[
  {
    "phase": 62,
    "mode": "original",
    "label": "op Some(21) LayoutRearrangement: TensorShape([2, 577, 4096]) F8F143 { scale_exponent: -4 } Amp(Left) -> TensorShape([2, 577, 4096]) Amp(Left)",
    "transfers": 147712,
    "receive_fragments": 2363392,
    "maximum_endpoint_fragments": 2020,
    "transmitted_bytes": 4726784,
    "received_bytes": 75628544,
    "maximum_outgoing_bytes": 8192,
    "maximum_incoming_bytes": 56448,
    "source_staging_max_bytes": 0,
    "destination_staging_max_bytes": 0,
    "combined_staging_max_bytes": 0,
    "source_copy_max_bytes": 0,
    "destination_copy_max_bytes": 0,
    "local_copy_runs": 0,
    "affine_copy_loops": 0,
    "copy_cycles_floor": 0,
    "replay_address_ranges_valid": true
  },
  {
    "phase": 62,
    "mode": "source",
    "label": "op Some(21) LayoutRearrangement: TensorShape([2, 577, 4096]) F8F143 { scale_exponent: -4 } Amp(Left) -> TensorShape([2, 577, 4096]) Amp(Left)",
    "transfers": 147712,
    "receive_fragments": 2363392,
    "maximum_endpoint_fragments": 2020,
    "transmitted_bytes": 4726784,
    "received_bytes": 75628544,
    "maximum_outgoing_bytes": 8192,
    "maximum_incoming_bytes": 56448,
    "source_staging_max_bytes": 8192,
    "destination_staging_max_bytes": 0,
    "combined_staging_max_bytes": 8192,
    "source_copy_max_bytes": 8192,
    "destination_copy_max_bytes": 0,
    "local_copy_runs": 577,
    "affine_copy_loops": 577,
    "copy_cycles_floor": 1024,
    "replay_address_ranges_valid": true
  },
  {
    "phase": 62,
    "mode": "destination",
    "label": "op Some(21) LayoutRearrangement: TensorShape([2, 577, 4096]) F8F143 { scale_exponent: -4 } Amp(Left) -> TensorShape([2, 577, 4096]) Amp(Left)",
    "transfers": 8655,
    "receive_fragments": 138480,
    "maximum_endpoint_fragments": 113,
    "transmitted_bytes": 4726784,
    "received_bytes": 75628544,
    "maximum_outgoing_bytes": 8192,
    "maximum_incoming_bytes": 56448,
    "source_staging_max_bytes": 0,
    "destination_staging_max_bytes": 56448,
    "combined_staging_max_bytes": 56448,
    "source_copy_max_bytes": 0,
    "destination_copy_max_bytes": 56448,
    "local_copy_runs": 2363392,
    "affine_copy_loops": 138480,
    "copy_cycles_floor": 7056,
    "replay_address_ranges_valid": true
  },
  {
    "phase": 62,
    "mode": "both",
    "label": "op Some(21) LayoutRearrangement: TensorShape([2, 577, 4096]) F8F143 { scale_exponent: -4 } Amp(Left) -> TensorShape([2, 577, 4096]) Amp(Left)",
    "transfers": 8655,
    "receive_fragments": 138480,
    "maximum_endpoint_fragments": 113,
    "transmitted_bytes": 4726784,
    "received_bytes": 75628544,
    "maximum_outgoing_bytes": 8192,
    "maximum_incoming_bytes": 56448,
    "source_staging_max_bytes": 8192,
    "destination_staging_max_bytes": 56448,
    "combined_staging_max_bytes": 64640,
    "source_copy_max_bytes": 8192,
    "destination_copy_max_bytes": 56448,
    "local_copy_runs": 2363969,
    "affine_copy_loops": 139057,
    "copy_cycles_floor": 8080,
    "replay_address_ranges_valid": true
  },
  {
    "phase": 101,
    "mode": "original",
    "label": "op Some(41) LayoutRearrangement: TensorShape([2, 1, 4096]) F16 RowMajor -> TensorShape([2, 1, 4096]) RowMajor",
    "transfers": 1023,
    "receive_fragments": 1023,
    "maximum_endpoint_fragments": 1023,
    "transmitted_bytes": 16368,
    "received_bytes": 16368,
    "maximum_outgoing_bytes": 16,
    "maximum_incoming_bytes": 16368,
    "source_staging_max_bytes": 0,
    "destination_staging_max_bytes": 0,
    "combined_staging_max_bytes": 0,
    "source_copy_max_bytes": 0,
    "destination_copy_max_bytes": 0,
    "local_copy_runs": 0,
    "affine_copy_loops": 0,
    "copy_cycles_floor": 0,
    "replay_address_ranges_valid": true
  },
  {
    "phase": 101,
    "mode": "source",
    "label": "op Some(41) LayoutRearrangement: TensorShape([2, 1, 4096]) F16 RowMajor -> TensorShape([2, 1, 4096]) RowMajor",
    "transfers": 1023,
    "receive_fragments": 1023,
    "maximum_endpoint_fragments": 1023,
    "transmitted_bytes": 16368,
    "received_bytes": 16368,
    "maximum_outgoing_bytes": 16,
    "maximum_incoming_bytes": 16368,
    "source_staging_max_bytes": 16,
    "destination_staging_max_bytes": 0,
    "combined_staging_max_bytes": 16,
    "source_copy_max_bytes": 16,
    "destination_copy_max_bytes": 0,
    "local_copy_runs": 1023,
    "affine_copy_loops": 1023,
    "copy_cycles_floor": 2,
    "replay_address_ranges_valid": true
  },
  {
    "phase": 101,
    "mode": "destination",
    "label": "op Some(41) LayoutRearrangement: TensorShape([2, 1, 4096]) F16 RowMajor -> TensorShape([2, 1, 4096]) RowMajor",
    "transfers": 1023,
    "receive_fragments": 1023,
    "maximum_endpoint_fragments": 1023,
    "transmitted_bytes": 16368,
    "received_bytes": 16368,
    "maximum_outgoing_bytes": 16,
    "maximum_incoming_bytes": 16368,
    "source_staging_max_bytes": 0,
    "destination_staging_max_bytes": 16368,
    "combined_staging_max_bytes": 16368,
    "source_copy_max_bytes": 0,
    "destination_copy_max_bytes": 16368,
    "local_copy_runs": 1,
    "affine_copy_loops": 1,
    "copy_cycles_floor": 2046,
    "replay_address_ranges_valid": true
  },
  {
    "phase": 101,
    "mode": "both",
    "label": "op Some(41) LayoutRearrangement: TensorShape([2, 1, 4096]) F16 RowMajor -> TensorShape([2, 1, 4096]) RowMajor",
    "transfers": 1023,
    "receive_fragments": 1023,
    "maximum_endpoint_fragments": 1023,
    "transmitted_bytes": 16368,
    "received_bytes": 16368,
    "maximum_outgoing_bytes": 16,
    "maximum_incoming_bytes": 16368,
    "source_staging_max_bytes": 16,
    "destination_staging_max_bytes": 16368,
    "combined_staging_max_bytes": 16368,
    "source_copy_max_bytes": 16,
    "destination_copy_max_bytes": 16368,
    "local_copy_runs": 1024,
    "affine_copy_loops": 1024,
    "copy_cycles_floor": 2048,
    "replay_address_ranges_valid": true
  },
  {
    "phase": 58,
    "mode": "original",
    "label": "op Some(18) OperatorInputs: TensorShape([2, 577, 1024]) F8F143 { scale_exponent: -4 } Amp(Left) -> TensorShape([2, 577, 1024]) Amp(Left)",
    "transfers": 45375,
    "receive_fragments": 1229447,
    "maximum_endpoint_fragments": 1000,
    "transmitted_bytes": 9565696,
    "received_bytes": 85128704,
    "maximum_outgoing_bytes": 37888,
    "maximum_incoming_bytes": 60416,
    "source_staging_max_bytes": 0,
    "destination_staging_max_bytes": 0,
    "combined_staging_max_bytes": 0,
    "source_copy_max_bytes": 0,
    "destination_copy_max_bytes": 0,
    "local_copy_runs": 0,
    "affine_copy_loops": 0,
    "copy_cycles_floor": 0,
    "replay_address_ranges_valid": true
  },
  {
    "phase": 58,
    "mode": "source",
    "label": "op Some(18) OperatorInputs: TensorShape([2, 577, 1024]) F8F143 { scale_exponent: -4 } Amp(Left) -> TensorShape([2, 577, 1024]) Amp(Left)",
    "transfers": 41406,
    "receive_fragments": 1224290,
    "maximum_endpoint_fragments": 972,
    "transmitted_bytes": 9565696,
    "received_bytes": 85128704,
    "maximum_outgoing_bytes": 37888,
    "maximum_incoming_bytes": 60416,
    "source_staging_max_bytes": 37888,
    "destination_staging_max_bytes": 0,
    "combined_staging_max_bytes": 37888,
    "source_copy_max_bytes": 37888,
    "destination_copy_max_bytes": 0,
    "local_copy_runs": 8656,
    "affine_copy_loops": 4801,
    "copy_cycles_floor": 4736,
    "replay_address_ranges_valid": true
  },
  {
    "phase": 58,
    "mode": "destination",
    "label": "op Some(18) OperatorInputs: TensorShape([2, 577, 1024]) F8F143 { scale_exponent: -4 } Amp(Left) -> TensorShape([2, 577, 1024]) Amp(Left)",
    "transfers": 4886,
    "receive_fragments": 79930,
    "maximum_endpoint_fragments": 87,
    "transmitted_bytes": 9565696,
    "received_bytes": 85128704,
    "maximum_outgoing_bytes": 37888,
    "maximum_incoming_bytes": 60416,
    "source_staging_max_bytes": 0,
    "destination_staging_max_bytes": 60416,
    "combined_staging_max_bytes": 60416,
    "source_copy_max_bytes": 0,
    "destination_copy_max_bytes": 60416,
    "local_copy_runs": 1226927,
    "affine_copy_loops": 80530,
    "copy_cycles_floor": 7552,
    "replay_address_ranges_valid": true
  },
  {
    "phase": 58,
    "mode": "both",
    "label": "op Some(18) OperatorInputs: TensorShape([2, 577, 1024]) F8F143 { scale_exponent: -4 } Amp(Left) -> TensorShape([2, 577, 1024]) Amp(Left)",
    "transfers": 4256,
    "receive_fragments": 79300,
    "maximum_endpoint_fragments": 80,
    "transmitted_bytes": 9565696,
    "received_bytes": 85128704,
    "maximum_outgoing_bytes": 37888,
    "maximum_incoming_bytes": 60416,
    "source_staging_max_bytes": 37888,
    "destination_staging_max_bytes": 60416,
    "combined_staging_max_bytes": 98304,
    "source_copy_max_bytes": 37888,
    "destination_copy_max_bytes": 60416,
    "local_copy_runs": 1230143,
    "affine_copy_loops": 83102,
    "copy_cycles_floor": 12288,
    "replay_address_ranges_valid": true
  },
  {
    "phase": 84,
    "mode": "original",
    "label": "op Some(35) OperatorInputs: TensorShape([2, 577, 1024]) F16 RowMajor -> TensorShape([16, 577, 128]) RowMajor",
    "transfers": 82504,
    "receive_fragments": 83079,
    "maximum_endpoint_fragments": 749,
    "transmitted_bytes": 4579072,
    "received_bytes": 4726272,
    "maximum_outgoing_bytes": 12000,
    "maximum_incoming_bytes": 22528,
    "source_staging_max_bytes": 0,
    "destination_staging_max_bytes": 0,
    "combined_staging_max_bytes": 0,
    "source_copy_max_bytes": 0,
    "destination_copy_max_bytes": 0,
    "local_copy_runs": 0,
    "affine_copy_loops": 0,
    "copy_cycles_floor": 0,
    "replay_address_ranges_valid": true
  },
  {
    "phase": 84,
    "mode": "source",
    "label": "op Some(35) OperatorInputs: TensorShape([2, 577, 1024]) F16 RowMajor -> TensorShape([16, 577, 128]) RowMajor",
    "transfers": 82504,
    "receive_fragments": 83079,
    "maximum_endpoint_fragments": 749,
    "transmitted_bytes": 4579072,
    "received_bytes": 4726272,
    "maximum_outgoing_bytes": 12000,
    "maximum_incoming_bytes": 22528,
    "source_staging_max_bytes": 12000,
    "destination_staging_max_bytes": 0,
    "combined_staging_max_bytes": 12000,
    "source_copy_max_bytes": 12000,
    "destination_copy_max_bytes": 0,
    "local_copy_runs": 76505,
    "affine_copy_loops": 9172,
    "copy_cycles_floor": 1500,
    "replay_address_ranges_valid": true
  },
  {
    "phase": 84,
    "mode": "destination",
    "label": "op Some(35) OperatorInputs: TensorShape([2, 577, 1024]) F16 RowMajor -> TensorShape([16, 577, 128]) RowMajor",
    "transfers": 82399,
    "receive_fragments": 82974,
    "maximum_endpoint_fragments": 749,
    "transmitted_bytes": 4579072,
    "received_bytes": 4726272,
    "maximum_outgoing_bytes": 12000,
    "maximum_incoming_bytes": 22528,
    "source_staging_max_bytes": 0,
    "destination_staging_max_bytes": 22528,
    "combined_staging_max_bytes": 22528,
    "source_copy_max_bytes": 0,
    "destination_copy_max_bytes": 22528,
    "local_copy_runs": 5512,
    "affine_copy_loops": 5364,
    "copy_cycles_floor": 2816,
    "replay_address_ranges_valid": true
  },
  {
    "phase": 84,
    "mode": "both",
    "label": "op Some(35) OperatorInputs: TensorShape([2, 577, 1024]) F16 RowMajor -> TensorShape([16, 577, 128]) RowMajor",
    "transfers": 80479,
    "receive_fragments": 81054,
    "maximum_endpoint_fragments": 719,
    "transmitted_bytes": 4579072,
    "received_bytes": 4726272,
    "maximum_outgoing_bytes": 12000,
    "maximum_incoming_bytes": 22528,
    "source_staging_max_bytes": 12000,
    "destination_staging_max_bytes": 22528,
    "combined_staging_max_bytes": 34272,
    "source_copy_max_bytes": 12000,
    "destination_copy_max_bytes": 22528,
    "local_copy_runs": 82017,
    "affine_copy_loops": 14536,
    "copy_cycles_floor": 4316,
    "replay_address_ranges_valid": true
  },
  {
    "phase": 17,
    "mode": "original",
    "label": "op Some(4) LayoutRearrangement: TensorShape([2, 577, 1024]) F8F143 { scale_exponent: -4 } Amp(Left) -> TensorShape([2, 577, 1024]) Amp(Left)",
    "transfers": 36928,
    "receive_fragments": 590848,
    "maximum_endpoint_fragments": 544,
    "transmitted_bytes": 1181696,
    "received_bytes": 18907136,
    "maximum_outgoing_bytes": 2048,
    "maximum_incoming_bytes": 15360,
    "source_staging_max_bytes": 0,
    "destination_staging_max_bytes": 0,
    "combined_staging_max_bytes": 0,
    "source_copy_max_bytes": 0,
    "destination_copy_max_bytes": 0,
    "local_copy_runs": 0,
    "affine_copy_loops": 0,
    "copy_cycles_floor": 0,
    "replay_address_ranges_valid": true
  },
  {
    "phase": 17,
    "mode": "source",
    "label": "op Some(4) LayoutRearrangement: TensorShape([2, 577, 1024]) F8F143 { scale_exponent: -4 } Amp(Left) -> TensorShape([2, 577, 1024]) Amp(Left)",
    "transfers": 36928,
    "receive_fragments": 590848,
    "maximum_endpoint_fragments": 544,
    "transmitted_bytes": 1181696,
    "received_bytes": 18907136,
    "maximum_outgoing_bytes": 2048,
    "maximum_incoming_bytes": 15360,
    "source_staging_max_bytes": 2048,
    "destination_staging_max_bytes": 0,
    "combined_staging_max_bytes": 2048,
    "source_copy_max_bytes": 2048,
    "destination_copy_max_bytes": 0,
    "local_copy_runs": 577,
    "affine_copy_loops": 577,
    "copy_cycles_floor": 256,
    "replay_address_ranges_valid": true
  },
  {
    "phase": 17,
    "mode": "destination",
    "label": "op Some(4) LayoutRearrangement: TensorShape([2, 577, 1024]) F8F143 { scale_exponent: -4 } Amp(Left) -> TensorShape([2, 577, 1024]) Amp(Left)",
    "transfers": 3462,
    "receive_fragments": 55392,
    "maximum_endpoint_fragments": 46,
    "transmitted_bytes": 1181696,
    "received_bytes": 18907136,
    "maximum_outgoing_bytes": 2048,
    "maximum_incoming_bytes": 15360,
    "source_staging_max_bytes": 0,
    "destination_staging_max_bytes": 15360,
    "combined_staging_max_bytes": 15360,
    "source_copy_max_bytes": 0,
    "destination_copy_max_bytes": 15360,
    "local_copy_runs": 590848,
    "affine_copy_loops": 55392,
    "copy_cycles_floor": 1920,
    "replay_address_ranges_valid": true
  },
  {
    "phase": 17,
    "mode": "both",
    "label": "op Some(4) LayoutRearrangement: TensorShape([2, 577, 1024]) F8F143 { scale_exponent: -4 } Amp(Left) -> TensorShape([2, 577, 1024]) Amp(Left)",
    "transfers": 3462,
    "receive_fragments": 55392,
    "maximum_endpoint_fragments": 46,
    "transmitted_bytes": 1181696,
    "received_bytes": 18907136,
    "maximum_outgoing_bytes": 2048,
    "maximum_incoming_bytes": 15360,
    "source_staging_max_bytes": 2048,
    "destination_staging_max_bytes": 15360,
    "combined_staging_max_bytes": 17408,
    "source_copy_max_bytes": 2048,
    "destination_copy_max_bytes": 15360,
    "local_copy_runs": 591425,
    "affine_copy_loops": 55969,
    "copy_cycles_floor": 2176,
    "replay_address_ranges_valid": true
  },
  {
    "phase": 69,
    "mode": "original",
    "label": "op Some(26) OperatorInputs: TensorShape([2, 577, 1024]) F8F143 { scale_exponent: -4 } Amp(Left) -> TensorShape([2, 577, 1024]) Amp(Left)",
    "transfers": 39101,
    "receive_fragments": 594173,
    "maximum_endpoint_fragments": 515,
    "transmitted_bytes": 5372416,
    "received_bytes": 41972224,
    "maximum_outgoing_bytes": 19968,
    "maximum_incoming_bytes": 30208,
    "source_staging_max_bytes": 0,
    "destination_staging_max_bytes": 0,
    "combined_staging_max_bytes": 0,
    "source_copy_max_bytes": 0,
    "destination_copy_max_bytes": 0,
    "local_copy_runs": 0,
    "affine_copy_loops": 0,
    "copy_cycles_floor": 0,
    "replay_address_ranges_valid": true
  },
  {
    "phase": 69,
    "mode": "source",
    "label": "op Some(26) OperatorInputs: TensorShape([2, 577, 1024]) F8F143 { scale_exponent: -4 } Amp(Left) -> TensorShape([2, 577, 1024]) Amp(Left)",
    "transfers": 39101,
    "receive_fragments": 594173,
    "maximum_endpoint_fragments": 515,
    "transmitted_bytes": 5372416,
    "received_bytes": 41972224,
    "maximum_outgoing_bytes": 19968,
    "maximum_incoming_bytes": 30208,
    "source_staging_max_bytes": 19968,
    "destination_staging_max_bytes": 0,
    "combined_staging_max_bytes": 19968,
    "source_copy_max_bytes": 19968,
    "destination_copy_max_bytes": 0,
    "local_copy_runs": 2258,
    "affine_copy_loops": 2222,
    "copy_cycles_floor": 2496,
    "replay_address_ranges_valid": true
  },
  {
    "phase": 69,
    "mode": "destination",
    "label": "op Some(26) OperatorInputs: TensorShape([2, 577, 1024]) F8F143 { scale_exponent: -4 } Amp(Left) -> TensorShape([2, 577, 1024]) Amp(Left)",
    "transfers": 6789,
    "receive_fragments": 77181,
    "maximum_endpoint_fragments": 81,
    "transmitted_bytes": 5372416,
    "received_bytes": 41972224,
    "maximum_outgoing_bytes": 19968,
    "maximum_incoming_bytes": 30208,
    "source_staging_max_bytes": 0,
    "destination_staging_max_bytes": 30208,
    "combined_staging_max_bytes": 30208,
    "source_copy_max_bytes": 0,
    "destination_copy_max_bytes": 30208,
    "local_copy_runs": 592334,
    "affine_copy_loops": 75342,
    "copy_cycles_floor": 3776,
    "replay_address_ranges_valid": true
  },
  {
    "phase": 69,
    "mode": "both",
    "label": "op Some(26) OperatorInputs: TensorShape([2, 577, 1024]) F8F143 { scale_exponent: -4 } Amp(Left) -> TensorShape([2, 577, 1024]) Amp(Left)",
    "transfers": 6789,
    "receive_fragments": 77181,
    "maximum_endpoint_fragments": 81,
    "transmitted_bytes": 5372416,
    "received_bytes": 41972224,
    "maximum_outgoing_bytes": 19968,
    "maximum_incoming_bytes": 30208,
    "source_staging_max_bytes": 19968,
    "destination_staging_max_bytes": 30208,
    "combined_staging_max_bytes": 50176,
    "source_copy_max_bytes": 19968,
    "destination_copy_max_bytes": 30208,
    "local_copy_runs": 594592,
    "affine_copy_loops": 77564,
    "copy_cycles_floor": 6272,
    "replay_address_ranges_valid": true
  },
  {
    "phase": 83,
    "mode": "original",
    "label": "op None OperatorInputs: TensorShape([2, 1, 1024]) F16 RowMajor -> TensorShape([16, 1, 128]) RowMajor",
    "transfers": 511,
    "receive_fragments": 511,
    "maximum_endpoint_fragments": 511,
    "transmitted_bytes": 4088,
    "received_bytes": 4088,
    "maximum_outgoing_bytes": 8,
    "maximum_incoming_bytes": 4088,
    "source_staging_max_bytes": 0,
    "destination_staging_max_bytes": 0,
    "combined_staging_max_bytes": 0,
    "source_copy_max_bytes": 0,
    "destination_copy_max_bytes": 0,
    "local_copy_runs": 0,
    "affine_copy_loops": 0,
    "copy_cycles_floor": 0,
    "replay_address_ranges_valid": true
  },
  {
    "phase": 83,
    "mode": "source",
    "label": "op None OperatorInputs: TensorShape([2, 1, 1024]) F16 RowMajor -> TensorShape([16, 1, 128]) RowMajor",
    "transfers": 511,
    "receive_fragments": 511,
    "maximum_endpoint_fragments": 511,
    "transmitted_bytes": 4088,
    "received_bytes": 4088,
    "maximum_outgoing_bytes": 8,
    "maximum_incoming_bytes": 4088,
    "source_staging_max_bytes": 8,
    "destination_staging_max_bytes": 0,
    "combined_staging_max_bytes": 8,
    "source_copy_max_bytes": 8,
    "destination_copy_max_bytes": 0,
    "local_copy_runs": 511,
    "affine_copy_loops": 511,
    "copy_cycles_floor": 1,
    "replay_address_ranges_valid": true
  },
  {
    "phase": 83,
    "mode": "destination",
    "label": "op None OperatorInputs: TensorShape([2, 1, 1024]) F16 RowMajor -> TensorShape([16, 1, 128]) RowMajor",
    "transfers": 511,
    "receive_fragments": 511,
    "maximum_endpoint_fragments": 511,
    "transmitted_bytes": 4088,
    "received_bytes": 4088,
    "maximum_outgoing_bytes": 8,
    "maximum_incoming_bytes": 4088,
    "source_staging_max_bytes": 0,
    "destination_staging_max_bytes": 4088,
    "combined_staging_max_bytes": 4088,
    "source_copy_max_bytes": 0,
    "destination_copy_max_bytes": 4088,
    "local_copy_runs": 1,
    "affine_copy_loops": 1,
    "copy_cycles_floor": 511,
    "replay_address_ranges_valid": true
  },
  {
    "phase": 83,
    "mode": "both",
    "label": "op None OperatorInputs: TensorShape([2, 1, 1024]) F16 RowMajor -> TensorShape([16, 1, 128]) RowMajor",
    "transfers": 511,
    "receive_fragments": 511,
    "maximum_endpoint_fragments": 511,
    "transmitted_bytes": 4088,
    "received_bytes": 4088,
    "maximum_outgoing_bytes": 8,
    "maximum_incoming_bytes": 4088,
    "source_staging_max_bytes": 8,
    "destination_staging_max_bytes": 4088,
    "combined_staging_max_bytes": 4088,
    "source_copy_max_bytes": 8,
    "destination_copy_max_bytes": 4088,
    "local_copy_runs": 512,
    "affine_copy_loops": 512,
    "copy_cycles_floor": 512,
    "replay_address_ranges_valid": true
  },
  {
    "phase": 93,
    "mode": "original",
    "label": "op Some(38) LayoutRearrangement: TensorShape([2, 1, 1024]) F16 RowMajor -> TensorShape([2, 1, 1024]) RowMajor",
    "transfers": 511,
    "receive_fragments": 511,
    "maximum_endpoint_fragments": 511,
    "transmitted_bytes": 4088,
    "received_bytes": 4088,
    "maximum_outgoing_bytes": 8,
    "maximum_incoming_bytes": 4088,
    "source_staging_max_bytes": 0,
    "destination_staging_max_bytes": 0,
    "combined_staging_max_bytes": 0,
    "source_copy_max_bytes": 0,
    "destination_copy_max_bytes": 0,
    "local_copy_runs": 0,
    "affine_copy_loops": 0,
    "copy_cycles_floor": 0,
    "replay_address_ranges_valid": true
  },
  {
    "phase": 93,
    "mode": "source",
    "label": "op Some(38) LayoutRearrangement: TensorShape([2, 1, 1024]) F16 RowMajor -> TensorShape([2, 1, 1024]) RowMajor",
    "transfers": 511,
    "receive_fragments": 511,
    "maximum_endpoint_fragments": 511,
    "transmitted_bytes": 4088,
    "received_bytes": 4088,
    "maximum_outgoing_bytes": 8,
    "maximum_incoming_bytes": 4088,
    "source_staging_max_bytes": 8,
    "destination_staging_max_bytes": 0,
    "combined_staging_max_bytes": 8,
    "source_copy_max_bytes": 8,
    "destination_copy_max_bytes": 0,
    "local_copy_runs": 511,
    "affine_copy_loops": 511,
    "copy_cycles_floor": 1,
    "replay_address_ranges_valid": true
  },
  {
    "phase": 93,
    "mode": "destination",
    "label": "op Some(38) LayoutRearrangement: TensorShape([2, 1, 1024]) F16 RowMajor -> TensorShape([2, 1, 1024]) RowMajor",
    "transfers": 511,
    "receive_fragments": 511,
    "maximum_endpoint_fragments": 511,
    "transmitted_bytes": 4088,
    "received_bytes": 4088,
    "maximum_outgoing_bytes": 8,
    "maximum_incoming_bytes": 4088,
    "source_staging_max_bytes": 0,
    "destination_staging_max_bytes": 4088,
    "combined_staging_max_bytes": 4088,
    "source_copy_max_bytes": 0,
    "destination_copy_max_bytes": 4088,
    "local_copy_runs": 1,
    "affine_copy_loops": 1,
    "copy_cycles_floor": 511,
    "replay_address_ranges_valid": true
  },
  {
    "phase": 93,
    "mode": "both",
    "label": "op Some(38) LayoutRearrangement: TensorShape([2, 1, 1024]) F16 RowMajor -> TensorShape([2, 1, 1024]) RowMajor",
    "transfers": 511,
    "receive_fragments": 511,
    "maximum_endpoint_fragments": 511,
    "transmitted_bytes": 4088,
    "received_bytes": 4088,
    "maximum_outgoing_bytes": 8,
    "maximum_incoming_bytes": 4088,
    "source_staging_max_bytes": 8,
    "destination_staging_max_bytes": 4088,
    "combined_staging_max_bytes": 4088,
    "source_copy_max_bytes": 8,
    "destination_copy_max_bytes": 4088,
    "local_copy_runs": 512,
    "affine_copy_loops": 512,
    "copy_cycles_floor": 512,
    "replay_address_ranges_valid": true
  },
  {
    "phase": 106,
    "mode": "original",
    "label": "op Some(44) LayoutRearrangement: TensorShape([2, 1, 1024]) F16 RowMajor -> TensorShape([2, 1, 1024]) RowMajor",
    "transfers": 511,
    "receive_fragments": 511,
    "maximum_endpoint_fragments": 511,
    "transmitted_bytes": 4088,
    "received_bytes": 4088,
    "maximum_outgoing_bytes": 8,
    "maximum_incoming_bytes": 4088,
    "source_staging_max_bytes": 0,
    "destination_staging_max_bytes": 0,
    "combined_staging_max_bytes": 0,
    "source_copy_max_bytes": 0,
    "destination_copy_max_bytes": 0,
    "local_copy_runs": 0,
    "affine_copy_loops": 0,
    "copy_cycles_floor": 0,
    "replay_address_ranges_valid": true
  },
  {
    "phase": 106,
    "mode": "source",
    "label": "op Some(44) LayoutRearrangement: TensorShape([2, 1, 1024]) F16 RowMajor -> TensorShape([2, 1, 1024]) RowMajor",
    "transfers": 511,
    "receive_fragments": 511,
    "maximum_endpoint_fragments": 511,
    "transmitted_bytes": 4088,
    "received_bytes": 4088,
    "maximum_outgoing_bytes": 8,
    "maximum_incoming_bytes": 4088,
    "source_staging_max_bytes": 8,
    "destination_staging_max_bytes": 0,
    "combined_staging_max_bytes": 8,
    "source_copy_max_bytes": 8,
    "destination_copy_max_bytes": 0,
    "local_copy_runs": 511,
    "affine_copy_loops": 511,
    "copy_cycles_floor": 1,
    "replay_address_ranges_valid": true
  },
  {
    "phase": 106,
    "mode": "destination",
    "label": "op Some(44) LayoutRearrangement: TensorShape([2, 1, 1024]) F16 RowMajor -> TensorShape([2, 1, 1024]) RowMajor",
    "transfers": 511,
    "receive_fragments": 511,
    "maximum_endpoint_fragments": 511,
    "transmitted_bytes": 4088,
    "received_bytes": 4088,
    "maximum_outgoing_bytes": 8,
    "maximum_incoming_bytes": 4088,
    "source_staging_max_bytes": 0,
    "destination_staging_max_bytes": 4088,
    "combined_staging_max_bytes": 4088,
    "source_copy_max_bytes": 0,
    "destination_copy_max_bytes": 4088,
    "local_copy_runs": 1,
    "affine_copy_loops": 1,
    "copy_cycles_floor": 511,
    "replay_address_ranges_valid": true
  },
  {
    "phase": 106,
    "mode": "both",
    "label": "op Some(44) LayoutRearrangement: TensorShape([2, 1, 1024]) F16 RowMajor -> TensorShape([2, 1, 1024]) RowMajor",
    "transfers": 511,
    "receive_fragments": 511,
    "maximum_endpoint_fragments": 511,
    "transmitted_bytes": 4088,
    "received_bytes": 4088,
    "maximum_outgoing_bytes": 8,
    "maximum_incoming_bytes": 4088,
    "source_staging_max_bytes": 8,
    "destination_staging_max_bytes": 4088,
    "combined_staging_max_bytes": 4088,
    "source_copy_max_bytes": 8,
    "destination_copy_max_bytes": 4088,
    "local_copy_runs": 512,
    "affine_copy_loops": 512,
    "copy_cycles_floor": 512,
    "replay_address_ranges_valid": true
  },
  {
    "phase": 60,
    "mode": "original",
    "label": "op Some(19) OperatorInputs: TensorShape([2, 577, 4096]) F16 Amp(Left) -> TensorShape([2, 577, 4096]) RowMajor",
    "transfers": 295194,
    "receive_fragments": 332090,
    "maximum_endpoint_fragments": 504,
    "transmitted_bytes": 9453376,
    "received_bytes": 18898752,
    "maximum_outgoing_bytes": 6912,
    "maximum_incoming_bytes": 16384,
    "source_staging_max_bytes": 0,
    "destination_staging_max_bytes": 0,
    "combined_staging_max_bytes": 0,
    "source_copy_max_bytes": 0,
    "destination_copy_max_bytes": 0,
    "local_copy_runs": 0,
    "affine_copy_loops": 0,
    "copy_cycles_floor": 0,
    "replay_address_ranges_valid": true
  },
  {
    "phase": 60,
    "mode": "source",
    "label": "op Some(19) OperatorInputs: TensorShape([2, 577, 4096]) F16 Amp(Left) -> TensorShape([2, 577, 4096]) RowMajor",
    "transfers": 152225,
    "receive_fragments": 189121,
    "maximum_endpoint_fragments": 272,
    "transmitted_bytes": 9453376,
    "received_bytes": 18898752,
    "maximum_outgoing_bytes": 6912,
    "maximum_incoming_bytes": 16384,
    "source_staging_max_bytes": 6912,
    "destination_staging_max_bytes": 0,
    "combined_staging_max_bytes": 6912,
    "source_copy_max_bytes": 6912,
    "destination_copy_max_bytes": 0,
    "local_copy_runs": 286074,
    "affine_copy_loops": 143097,
    "copy_cycles_floor": 864,
    "replay_address_ranges_valid": true
  },
  {
    "phase": 60,
    "mode": "destination",
    "label": "op Some(19) OperatorInputs: TensorShape([2, 577, 4096]) F16 Amp(Left) -> TensorShape([2, 577, 4096]) RowMajor",
    "transfers": 295194,
    "receive_fragments": 332090,
    "maximum_endpoint_fragments": 504,
    "transmitted_bytes": 9453376,
    "received_bytes": 18898752,
    "maximum_outgoing_bytes": 6912,
    "maximum_incoming_bytes": 16384,
    "source_staging_max_bytes": 0,
    "destination_staging_max_bytes": 16384,
    "combined_staging_max_bytes": 16384,
    "source_copy_max_bytes": 0,
    "destination_copy_max_bytes": 16384,
    "local_copy_runs": 8150,
    "affine_copy_loops": 7944,
    "copy_cycles_floor": 2048,
    "replay_address_ranges_valid": true
  },
  {
    "phase": 60,
    "mode": "both",
    "label": "op Some(19) OperatorInputs: TensorShape([2, 577, 4096]) F16 Amp(Left) -> TensorShape([2, 577, 4096]) RowMajor",
    "transfers": 152225,
    "receive_fragments": 189121,
    "maximum_endpoint_fragments": 272,
    "transmitted_bytes": 9453376,
    "received_bytes": 18898752,
    "maximum_outgoing_bytes": 6912,
    "maximum_incoming_bytes": 16384,
    "source_staging_max_bytes": 6912,
    "destination_staging_max_bytes": 16384,
    "combined_staging_max_bytes": 23296,
    "source_copy_max_bytes": 6912,
    "destination_copy_max_bytes": 16384,
    "local_copy_runs": 294224,
    "affine_copy_loops": 151041,
    "copy_cycles_floor": 2912,
    "replay_address_ranges_valid": true
  },
  {
    "phase": 0,
    "mode": "original",
    "label": "op Some(0) LayoutRearrangement: TensorShape([2, 577, 588]) F8F143 { scale_exponent: -4 } Amp(Left) -> TensorShape([2, 577, 588]) Amp(Left)",
    "transfers": 21926,
    "receive_fragments": 635854,
    "maximum_endpoint_fragments": 502,
    "transmitted_bytes": 678552,
    "received_bytes": 19678008,
    "maximum_outgoing_bytes": 1176,
    "maximum_incoming_bytes": 14848,
    "source_staging_max_bytes": 0,
    "destination_staging_max_bytes": 0,
    "combined_staging_max_bytes": 0,
    "source_copy_max_bytes": 0,
    "destination_copy_max_bytes": 0,
    "local_copy_runs": 0,
    "affine_copy_loops": 0,
    "copy_cycles_floor": 0,
    "replay_address_ranges_valid": true
  },
  {
    "phase": 0,
    "mode": "source",
    "label": "op Some(0) LayoutRearrangement: TensorShape([2, 577, 588]) F8F143 { scale_exponent: -4 } Amp(Left) -> TensorShape([2, 577, 588]) Amp(Left)",
    "transfers": 21926,
    "receive_fragments": 635854,
    "maximum_endpoint_fragments": 502,
    "transmitted_bytes": 678552,
    "received_bytes": 19678008,
    "maximum_outgoing_bytes": 1176,
    "maximum_incoming_bytes": 14848,
    "source_staging_max_bytes": 1176,
    "destination_staging_max_bytes": 0,
    "combined_staging_max_bytes": 1176,
    "source_copy_max_bytes": 1176,
    "destination_copy_max_bytes": 0,
    "local_copy_runs": 1154,
    "affine_copy_loops": 1154,
    "copy_cycles_floor": 147,
    "replay_address_ranges_valid": true
  },
  {
    "phase": 0,
    "mode": "destination",
    "label": "op Some(0) LayoutRearrangement: TensorShape([2, 577, 588]) F8F143 { scale_exponent: -4 } Amp(Left) -> TensorShape([2, 577, 588]) Amp(Left)",
    "transfers": 6347,
    "receive_fragments": 184063,
    "maximum_endpoint_fragments": 243,
    "transmitted_bytes": 678552,
    "received_bytes": 19678008,
    "maximum_outgoing_bytes": 1176,
    "maximum_incoming_bytes": 14848,
    "source_staging_max_bytes": 0,
    "destination_staging_max_bytes": 14848,
    "combined_staging_max_bytes": 14848,
    "source_copy_max_bytes": 0,
    "destination_copy_max_bytes": 14848,
    "local_copy_runs": 635854,
    "affine_copy_loops": 167330,
    "copy_cycles_floor": 1856,
    "replay_address_ranges_valid": true
  },
  {
    "phase": 0,
    "mode": "both",
    "label": "op Some(0) LayoutRearrangement: TensorShape([2, 577, 588]) F8F143 { scale_exponent: -4 } Amp(Left) -> TensorShape([2, 577, 588]) Amp(Left)",
    "transfers": 5770,
    "receive_fragments": 167330,
    "maximum_endpoint_fragments": 126,
    "transmitted_bytes": 678552,
    "received_bytes": 19678008,
    "maximum_outgoing_bytes": 1176,
    "maximum_incoming_bytes": 14848,
    "source_staging_max_bytes": 1176,
    "destination_staging_max_bytes": 14848,
    "combined_staging_max_bytes": 16024,
    "source_copy_max_bytes": 1176,
    "destination_copy_max_bytes": 14848,
    "local_copy_runs": 637008,
    "affine_copy_loops": 168484,
    "copy_cycles_floor": 2003,
    "replay_address_ranges_valid": true
  },
  {
    "phase": 26,
    "mode": "original",
    "label": "op Some(12) OperatorInputs: TensorShape([2, 577, 1024]) F16 RowMajor -> TensorShape([32, 577, 64]) RowMajor",
    "transfers": 166035,
    "receive_fragments": 166035,
    "maximum_endpoint_fragments": 429,
    "transmitted_bytes": 7084416,
    "received_bytes": 7084416,
    "maximum_outgoing_bytes": 12288,
    "maximum_incoming_bytes": 5760,
    "source_staging_max_bytes": 0,
    "destination_staging_max_bytes": 0,
    "combined_staging_max_bytes": 0,
    "source_copy_max_bytes": 0,
    "destination_copy_max_bytes": 0,
    "local_copy_runs": 0,
    "affine_copy_loops": 0,
    "copy_cycles_floor": 0,
    "replay_address_ranges_valid": true
  },
  {
    "phase": 26,
    "mode": "source",
    "label": "op Some(12) OperatorInputs: TensorShape([2, 577, 1024]) F16 RowMajor -> TensorShape([32, 577, 64]) RowMajor",
    "transfers": 166035,
    "receive_fragments": 166035,
    "maximum_endpoint_fragments": 429,
    "transmitted_bytes": 7084416,
    "received_bytes": 7084416,
    "maximum_outgoing_bytes": 12288,
    "maximum_incoming_bytes": 5760,
    "source_staging_max_bytes": 12288,
    "destination_staging_max_bytes": 0,
    "combined_staging_max_bytes": 12288,
    "source_copy_max_bytes": 12288,
    "destination_copy_max_bytes": 0,
    "local_copy_runs": 152488,
    "affine_copy_loops": 78696,
    "copy_cycles_floor": 1536,
    "replay_address_ranges_valid": true
  },
  {
    "phase": 26,
    "mode": "destination",
    "label": "op Some(12) OperatorInputs: TensorShape([2, 577, 1024]) F16 RowMajor -> TensorShape([32, 577, 64]) RowMajor",
    "transfers": 166035,
    "receive_fragments": 166035,
    "maximum_endpoint_fragments": 429,
    "transmitted_bytes": 7084416,
    "received_bytes": 7084416,
    "maximum_outgoing_bytes": 12288,
    "maximum_incoming_bytes": 5760,
    "source_staging_max_bytes": 0,
    "destination_staging_max_bytes": 5760,
    "combined_staging_max_bytes": 5760,
    "source_copy_max_bytes": 0,
    "destination_copy_max_bytes": 5760,
    "local_copy_runs": 153075,
    "affine_copy_loops": 79283,
    "copy_cycles_floor": 720,
    "replay_address_ranges_valid": true
  },
  {
    "phase": 26,
    "mode": "both",
    "label": "op Some(12) OperatorInputs: TensorShape([2, 577, 1024]) F16 RowMajor -> TensorShape([32, 577, 64]) RowMajor",
    "transfers": 87776,
    "receive_fragments": 87776,
    "maximum_endpoint_fragments": 237,
    "transmitted_bytes": 7084416,
    "received_bytes": 7084416,
    "maximum_outgoing_bytes": 12288,
    "maximum_incoming_bytes": 5760,
    "source_staging_max_bytes": 12288,
    "destination_staging_max_bytes": 5760,
    "combined_staging_max_bytes": 18048,
    "source_copy_max_bytes": 12288,
    "destination_copy_max_bytes": 5760,
    "local_copy_runs": 305563,
    "affine_copy_loops": 157979,
    "copy_cycles_floor": 2256,
    "replay_address_ranges_valid": true
  },
  {
    "phase": 38,
    "mode": "original",
    "label": "op Some(14) OperatorInputs: TensorShape([2, 577, 1024]) F8F143 { scale_exponent: -4 } Amp(Left) -> TensorShape([2, 577, 1024]) Amp(Left)",
    "transfers": 38158,
    "receive_fragments": 482334,
    "maximum_endpoint_fragments": 407,
    "transmitted_bytes": 3276800,
    "received_bytes": 22700032,
    "maximum_outgoing_bytes": 8192,
    "maximum_incoming_bytes": 15872,
    "source_staging_max_bytes": 0,
    "destination_staging_max_bytes": 0,
    "combined_staging_max_bytes": 0,
    "source_copy_max_bytes": 0,
    "destination_copy_max_bytes": 0,
    "local_copy_runs": 0,
    "affine_copy_loops": 0,
    "copy_cycles_floor": 0,
    "replay_address_ranges_valid": true
  },
  {
    "phase": 38,
    "mode": "source",
    "label": "op Some(14) OperatorInputs: TensorShape([2, 577, 1024]) F8F143 { scale_exponent: -4 } Amp(Left) -> TensorShape([2, 577, 1024]) Amp(Left)",
    "transfers": 38158,
    "receive_fragments": 482334,
    "maximum_endpoint_fragments": 407,
    "transmitted_bytes": 3276800,
    "received_bytes": 22700032,
    "maximum_outgoing_bytes": 8192,
    "maximum_incoming_bytes": 15872,
    "source_staging_max_bytes": 8192,
    "destination_staging_max_bytes": 0,
    "combined_staging_max_bytes": 8192,
    "source_copy_max_bytes": 8192,
    "destination_copy_max_bytes": 0,
    "local_copy_runs": 1969,
    "affine_copy_loops": 1951,
    "copy_cycles_floor": 1024,
    "replay_address_ranges_valid": true
  },
  {
    "phase": 38,
    "mode": "destination",
    "label": "op Some(14) OperatorInputs: TensorShape([2, 577, 1024]) F8F143 { scale_exponent: -4 } Amp(Left) -> TensorShape([2, 577, 1024]) Amp(Left)",
    "transfers": 10462,
    "receive_fragments": 122286,
    "maximum_endpoint_fragments": 107,
    "transmitted_bytes": 3276800,
    "received_bytes": 22700032,
    "maximum_outgoing_bytes": 8192,
    "maximum_incoming_bytes": 15872,
    "source_staging_max_bytes": 0,
    "destination_staging_max_bytes": 15872,
    "combined_staging_max_bytes": 15872,
    "source_copy_max_bytes": 0,
    "destination_copy_max_bytes": 15872,
    "local_copy_runs": 481585,
    "affine_copy_loops": 121537,
    "copy_cycles_floor": 1984,
    "replay_address_ranges_valid": true
  },
  {
    "phase": 38,
    "mode": "both",
    "label": "op Some(14) OperatorInputs: TensorShape([2, 577, 1024]) F8F143 { scale_exponent: -4 } Amp(Left) -> TensorShape([2, 577, 1024]) Amp(Left)",
    "transfers": 10462,
    "receive_fragments": 122286,
    "maximum_endpoint_fragments": 107,
    "transmitted_bytes": 3276800,
    "received_bytes": 22700032,
    "maximum_outgoing_bytes": 8192,
    "maximum_incoming_bytes": 15872,
    "source_staging_max_bytes": 8192,
    "destination_staging_max_bytes": 15872,
    "combined_staging_max_bytes": 24064,
    "source_copy_max_bytes": 8192,
    "destination_copy_max_bytes": 15872,
    "local_copy_runs": 483554,
    "affine_copy_loops": 123488,
    "copy_cycles_floor": 3008,
    "replay_address_ranges_valid": true
  },
  {
    "phase": 24,
    "mode": "original",
    "label": "op Some(5) OperatorInputs: TensorShape([2, 577, 3072]) F16 Amp(Left) -> TensorShape([2, 577, 3072]) RowMajor",
    "transfers": 221370,
    "receive_fragments": 249042,
    "maximum_endpoint_fragments": 376,
    "transmitted_bytes": 7089216,
    "received_bytes": 14173248,
    "maximum_outgoing_bytes": 5120,
    "maximum_incoming_bytes": 12288,
    "source_staging_max_bytes": 0,
    "destination_staging_max_bytes": 0,
    "combined_staging_max_bytes": 0,
    "source_copy_max_bytes": 0,
    "destination_copy_max_bytes": 0,
    "local_copy_runs": 0,
    "affine_copy_loops": 0,
    "copy_cycles_floor": 0,
    "replay_address_ranges_valid": true
  },
  {
    "phase": 24,
    "mode": "source",
    "label": "op Some(5) OperatorInputs: TensorShape([2, 577, 3072]) F16 Amp(Left) -> TensorShape([2, 577, 3072]) RowMajor",
    "transfers": 110697,
    "receive_fragments": 138369,
    "maximum_endpoint_fragments": 200,
    "transmitted_bytes": 7089216,
    "received_bytes": 14173248,
    "maximum_outgoing_bytes": 5120,
    "maximum_incoming_bytes": 12288,
    "source_staging_max_bytes": 5120,
    "destination_staging_max_bytes": 0,
    "combined_staging_max_bytes": 5120,
    "source_copy_max_bytes": 5120,
    "destination_copy_max_bytes": 0,
    "local_copy_runs": 221370,
    "affine_copy_loops": 110697,
    "copy_cycles_floor": 640,
    "replay_address_ranges_valid": true
  },
  {
    "phase": 24,
    "mode": "destination",
    "label": "op Some(5) OperatorInputs: TensorShape([2, 577, 3072]) F16 Amp(Left) -> TensorShape([2, 577, 3072]) RowMajor",
    "transfers": 221370,
    "receive_fragments": 249042,
    "maximum_endpoint_fragments": 376,
    "transmitted_bytes": 7089216,
    "received_bytes": 14173248,
    "maximum_outgoing_bytes": 5120,
    "maximum_incoming_bytes": 12288,
    "source_staging_max_bytes": 0,
    "destination_staging_max_bytes": 12288,
    "combined_staging_max_bytes": 12288,
    "source_copy_max_bytes": 0,
    "destination_copy_max_bytes": 12288,
    "local_copy_runs": 5214,
    "affine_copy_loops": 4912,
    "copy_cycles_floor": 1536,
    "replay_address_ranges_valid": true
  },
  {
    "phase": 24,
    "mode": "both",
    "label": "op Some(5) OperatorInputs: TensorShape([2, 577, 3072]) F16 Amp(Left) -> TensorShape([2, 577, 3072]) RowMajor",
    "transfers": 110697,
    "receive_fragments": 138369,
    "maximum_endpoint_fragments": 200,
    "transmitted_bytes": 7089216,
    "received_bytes": 14173248,
    "maximum_outgoing_bytes": 5120,
    "maximum_incoming_bytes": 12288,
    "source_staging_max_bytes": 5120,
    "destination_staging_max_bytes": 12288,
    "combined_staging_max_bytes": 17408,
    "source_copy_max_bytes": 5120,
    "destination_copy_max_bytes": 12288,
    "local_copy_runs": 226584,
    "affine_copy_loops": 115609,
    "copy_cycles_floor": 2176,
    "replay_address_ranges_valid": true
  },
  {
    "phase": 77,
    "mode": "original",
    "label": "op Some(27) OperatorInputs: TensorShape([2, 577, 2048]) F16 Amp(Left) -> TensorShape([2, 577, 2048]) RowMajor",
    "transfers": 147594,
    "receive_fragments": 166042,
    "maximum_endpoint_fragments": 252,
    "transmitted_bytes": 4726592,
    "received_bytes": 9449280,
    "maximum_outgoing_bytes": 3552,
    "maximum_incoming_bytes": 8192,
    "source_staging_max_bytes": 0,
    "destination_staging_max_bytes": 0,
    "combined_staging_max_bytes": 0,
    "source_copy_max_bytes": 0,
    "destination_copy_max_bytes": 0,
    "local_copy_runs": 0,
    "affine_copy_loops": 0,
    "copy_cycles_floor": 0,
    "replay_address_ranges_valid": true
  },
  {
    "phase": 77,
    "mode": "source",
    "label": "op Some(27) OperatorInputs: TensorShape([2, 577, 2048]) F16 Amp(Left) -> TensorShape([2, 577, 2048]) RowMajor",
    "transfers": 147594,
    "receive_fragments": 166042,
    "maximum_endpoint_fragments": 252,
    "transmitted_bytes": 4726592,
    "received_bytes": 9449280,
    "maximum_outgoing_bytes": 3552,
    "maximum_incoming_bytes": 8192,
    "source_staging_max_bytes": 3552,
    "destination_staging_max_bytes": 0,
    "combined_staging_max_bytes": 3552,
    "source_copy_max_bytes": 3552,
    "destination_copy_max_bytes": 0,
    "local_copy_runs": 1683,
    "affine_copy_loops": 1555,
    "copy_cycles_floor": 444,
    "replay_address_ranges_valid": true
  },
  {
    "phase": 77,
    "mode": "destination",
    "label": "op Some(27) OperatorInputs: TensorShape([2, 577, 2048]) F16 Amp(Left) -> TensorShape([2, 577, 2048]) RowMajor",
    "transfers": 147594,
    "receive_fragments": 166042,
    "maximum_endpoint_fragments": 252,
    "transmitted_bytes": 4726592,
    "received_bytes": 9449280,
    "maximum_outgoing_bytes": 3552,
    "maximum_incoming_bytes": 8192,
    "source_staging_max_bytes": 0,
    "destination_staging_max_bytes": 8192,
    "combined_staging_max_bytes": 8192,
    "source_copy_max_bytes": 0,
    "destination_copy_max_bytes": 8192,
    "local_copy_runs": 4773,
    "affine_copy_loops": 4472,
    "copy_cycles_floor": 1024,
    "replay_address_ranges_valid": true
  },
  {
    "phase": 77,
    "mode": "both",
    "label": "op Some(27) OperatorInputs: TensorShape([2, 577, 2048]) F16 Amp(Left) -> TensorShape([2, 577, 2048]) RowMajor",
    "transfers": 147594,
    "receive_fragments": 166042,
    "maximum_endpoint_fragments": 252,
    "transmitted_bytes": 4726592,
    "received_bytes": 9449280,
    "maximum_outgoing_bytes": 3552,
    "maximum_incoming_bytes": 8192,
    "source_staging_max_bytes": 3552,
    "destination_staging_max_bytes": 8192,
    "combined_staging_max_bytes": 11712,
    "source_copy_max_bytes": 3552,
    "destination_copy_max_bytes": 8192,
    "local_copy_runs": 6456,
    "affine_copy_loops": 6027,
    "copy_cycles_floor": 1468,
    "replay_address_ranges_valid": true
  },
  {
    "phase": 1,
    "mode": "original",
    "label": "op Some(0) OperatorInputs: TensorShape([588, 1024]) F8F143 { scale_exponent: -4 } Amp(TransposedLeft) -> TensorShape([588, 1024]) BlockMajor(Matrix { row_block: 64, column_block: 16 })",
    "transfers": 38130,
    "receive_fragments": 39000,
    "maximum_endpoint_fragments": 225,
    "transmitted_bytes": 1260800,
    "received_bytes": 3226880,
    "maximum_outgoing_bytes": 3584,
    "maximum_incoming_bytes": 3072,
    "source_staging_max_bytes": 0,
    "destination_staging_max_bytes": 0,
    "combined_staging_max_bytes": 0,
    "source_copy_max_bytes": 0,
    "destination_copy_max_bytes": 0,
    "local_copy_runs": 0,
    "affine_copy_loops": 0,
    "copy_cycles_floor": 0,
    "replay_address_ranges_valid": true
  },
  {
    "phase": 1,
    "mode": "source",
    "label": "op Some(0) OperatorInputs: TensorShape([588, 1024]) F8F143 { scale_exponent: -4 } Amp(TransposedLeft) -> TensorShape([588, 1024]) BlockMajor(Matrix { row_block: 64, column_block: 16 })",
    "transfers": 2449,
    "receive_fragments": 3319,
    "maximum_endpoint_fragments": 50,
    "transmitted_bytes": 1260800,
    "received_bytes": 3226880,
    "maximum_outgoing_bytes": 3584,
    "maximum_incoming_bytes": 3072,
    "source_staging_max_bytes": 3584,
    "destination_staging_max_bytes": 0,
    "combined_staging_max_bytes": 3584,
    "source_copy_max_bytes": 3584,
    "destination_copy_max_bytes": 0,
    "local_copy_runs": 2449,
    "affine_copy_loops": 1504,
    "copy_cycles_floor": 448,
    "replay_address_ranges_valid": true
  },
  {
    "phase": 1,
    "mode": "destination",
    "label": "op Some(0) OperatorInputs: TensorShape([588, 1024]) F8F143 { scale_exponent: -4 } Amp(TransposedLeft) -> TensorShape([588, 1024]) BlockMajor(Matrix { row_block: 64, column_block: 16 })",
    "transfers": 2449,
    "receive_fragments": 3319,
    "maximum_endpoint_fragments": 50,
    "transmitted_bytes": 1260800,
    "received_bytes": 3226880,
    "maximum_outgoing_bytes": 3584,
    "maximum_incoming_bytes": 3072,
    "source_staging_max_bytes": 0,
    "destination_staging_max_bytes": 3072,
    "combined_staging_max_bytes": 3072,
    "source_copy_max_bytes": 0,
    "destination_copy_max_bytes": 3072,
    "local_copy_runs": 2429,
    "affine_copy_loops": 1484,
    "copy_cycles_floor": 384,
    "replay_address_ranges_valid": true
  },
  {
    "phase": 1,
    "mode": "both",
    "label": "op Some(0) OperatorInputs: TensorShape([588, 1024]) F8F143 { scale_exponent: -4 } Amp(TransposedLeft) -> TensorShape([588, 1024]) BlockMajor(Matrix { row_block: 64, column_block: 16 })",
    "transfers": 1504,
    "receive_fragments": 2374,
    "maximum_endpoint_fragments": 8,
    "transmitted_bytes": 1260800,
    "received_bytes": 3226880,
    "maximum_outgoing_bytes": 3584,
    "maximum_incoming_bytes": 3072,
    "source_staging_max_bytes": 3584,
    "destination_staging_max_bytes": 3072,
    "combined_staging_max_bytes": 6656,
    "source_copy_max_bytes": 3584,
    "destination_copy_max_bytes": 3072,
    "local_copy_runs": 4878,
    "affine_copy_loops": 2988,
    "copy_cycles_floor": 832,
    "replay_address_ranges_valid": true
  },
  {
    "phase": 49,
    "mode": "original",
    "label": "op Some(14) OperatorInputs: TensorShape([2, 577, 1024]) F16 Amp(Left) -> TensorShape([960]) RowMajor",
    "transfers": 14720,
    "receive_fragments": 14720,
    "maximum_endpoint_fragments": 220,
    "transmitted_bytes": 2367488,
    "received_bytes": 2367488,
    "maximum_outgoing_bytes": 26240,
    "maximum_incoming_bytes": 1920,
    "source_staging_max_bytes": 0,
    "destination_staging_max_bytes": 0,
    "combined_staging_max_bytes": 0,
    "source_copy_max_bytes": 0,
    "destination_copy_max_bytes": 0,
    "local_copy_runs": 0,
    "affine_copy_loops": 0,
    "copy_cycles_floor": 0,
    "replay_address_ranges_valid": true
  },
  {
    "phase": 49,
    "mode": "source",
    "label": "op Some(14) OperatorInputs: TensorShape([2, 577, 1024]) F16 Amp(Left) -> TensorShape([960]) RowMajor",
    "transfers": 1495,
    "receive_fragments": 1495,
    "maximum_endpoint_fragments": 22,
    "transmitted_bytes": 2367488,
    "received_bytes": 2367488,
    "maximum_outgoing_bytes": 26240,
    "maximum_incoming_bytes": 1920,
    "source_staging_max_bytes": 26240,
    "destination_staging_max_bytes": 0,
    "combined_staging_max_bytes": 26240,
    "source_copy_max_bytes": 26240,
    "destination_copy_max_bytes": 0,
    "local_copy_runs": 14720,
    "affine_copy_loops": 1495,
    "copy_cycles_floor": 3280,
    "replay_address_ranges_valid": true
  },
  {
    "phase": 49,
    "mode": "destination",
    "label": "op Some(14) OperatorInputs: TensorShape([2, 577, 1024]) F16 Amp(Left) -> TensorShape([960]) RowMajor",
    "transfers": 14720,
    "receive_fragments": 14720,
    "maximum_endpoint_fragments": 220,
    "transmitted_bytes": 2367488,
    "received_bytes": 2367488,
    "maximum_outgoing_bytes": 26240,
    "maximum_incoming_bytes": 1920,
    "source_staging_max_bytes": 0,
    "destination_staging_max_bytes": 1920,
    "combined_staging_max_bytes": 1920,
    "source_copy_max_bytes": 0,
    "destination_copy_max_bytes": 1920,
    "local_copy_runs": 1495,
    "affine_copy_loops": 1482,
    "copy_cycles_floor": 240,
    "replay_address_ranges_valid": true
  },
  {
    "phase": 49,
    "mode": "both",
    "label": "op Some(14) OperatorInputs: TensorShape([2, 577, 1024]) F16 Amp(Left) -> TensorShape([960]) RowMajor",
    "transfers": 1495,
    "receive_fragments": 1495,
    "maximum_endpoint_fragments": 22,
    "transmitted_bytes": 2367488,
    "received_bytes": 2367488,
    "maximum_outgoing_bytes": 26240,
    "maximum_incoming_bytes": 1920,
    "source_staging_max_bytes": 26240,
    "destination_staging_max_bytes": 1920,
    "combined_staging_max_bytes": 28160,
    "source_copy_max_bytes": 26240,
    "destination_copy_max_bytes": 1920,
    "local_copy_runs": 16215,
    "affine_copy_loops": 2977,
    "copy_cycles_floor": 3520,
    "replay_address_ranges_valid": true
  },
  {
    "phase": 50,
    "mode": "original",
    "label": "op Some(14) OperatorInputs: TensorShape([2, 577, 1024]) F16 Amp(Left) -> TensorShape([960]) RowMajor",
    "transfers": 14720,
    "receive_fragments": 14720,
    "maximum_endpoint_fragments": 220,
    "transmitted_bytes": 2367488,
    "received_bytes": 2367488,
    "maximum_outgoing_bytes": 26880,
    "maximum_incoming_bytes": 1920,
    "source_staging_max_bytes": 0,
    "destination_staging_max_bytes": 0,
    "combined_staging_max_bytes": 0,
    "source_copy_max_bytes": 0,
    "destination_copy_max_bytes": 0,
    "local_copy_runs": 0,
    "affine_copy_loops": 0,
    "copy_cycles_floor": 0,
    "replay_address_ranges_valid": true
  },
  {
    "phase": 50,
    "mode": "source",
    "label": "op Some(14) OperatorInputs: TensorShape([2, 577, 1024]) F16 Amp(Left) -> TensorShape([960]) RowMajor",
    "transfers": 1495,
    "receive_fragments": 1495,
    "maximum_endpoint_fragments": 22,
    "transmitted_bytes": 2367488,
    "received_bytes": 2367488,
    "maximum_outgoing_bytes": 26880,
    "maximum_incoming_bytes": 1920,
    "source_staging_max_bytes": 26880,
    "destination_staging_max_bytes": 0,
    "combined_staging_max_bytes": 26880,
    "source_copy_max_bytes": 26880,
    "destination_copy_max_bytes": 0,
    "local_copy_runs": 14720,
    "affine_copy_loops": 1495,
    "copy_cycles_floor": 3360,
    "replay_address_ranges_valid": true
  },
  {
    "phase": 50,
    "mode": "destination",
    "label": "op Some(14) OperatorInputs: TensorShape([2, 577, 1024]) F16 Amp(Left) -> TensorShape([960]) RowMajor",
    "transfers": 14720,
    "receive_fragments": 14720,
    "maximum_endpoint_fragments": 220,
    "transmitted_bytes": 2367488,
    "received_bytes": 2367488,
    "maximum_outgoing_bytes": 26880,
    "maximum_incoming_bytes": 1920,
    "source_staging_max_bytes": 0,
    "destination_staging_max_bytes": 1920,
    "combined_staging_max_bytes": 1920,
    "source_copy_max_bytes": 0,
    "destination_copy_max_bytes": 1920,
    "local_copy_runs": 1495,
    "affine_copy_loops": 1482,
    "copy_cycles_floor": 240,
    "replay_address_ranges_valid": true
  },
  {
    "phase": 50,
    "mode": "both",
    "label": "op Some(14) OperatorInputs: TensorShape([2, 577, 1024]) F16 Amp(Left) -> TensorShape([960]) RowMajor",
    "transfers": 1495,
    "receive_fragments": 1495,
    "maximum_endpoint_fragments": 22,
    "transmitted_bytes": 2367488,
    "received_bytes": 2367488,
    "maximum_outgoing_bytes": 26880,
    "maximum_incoming_bytes": 1920,
    "source_staging_max_bytes": 26880,
    "destination_staging_max_bytes": 1920,
    "combined_staging_max_bytes": 28800,
    "source_copy_max_bytes": 26880,
    "destination_copy_max_bytes": 1920,
    "local_copy_runs": 16215,
    "affine_copy_loops": 2977,
    "copy_cycles_floor": 3600,
    "replay_address_ranges_valid": true
  },
  {
    "phase": 51,
    "mode": "original",
    "label": "op Some(14) OperatorInputs: TensorShape([2, 577, 1024]) F16 Amp(Left) -> TensorShape([960]) RowMajor",
    "transfers": 14720,
    "receive_fragments": 14720,
    "maximum_endpoint_fragments": 220,
    "transmitted_bytes": 2367488,
    "received_bytes": 2367488,
    "maximum_outgoing_bytes": 26880,
    "maximum_incoming_bytes": 1920,
    "source_staging_max_bytes": 0,
    "destination_staging_max_bytes": 0,
    "combined_staging_max_bytes": 0,
    "source_copy_max_bytes": 0,
    "destination_copy_max_bytes": 0,
    "local_copy_runs": 0,
    "affine_copy_loops": 0,
    "copy_cycles_floor": 0,
    "replay_address_ranges_valid": true
  },
  {
    "phase": 51,
    "mode": "source",
    "label": "op Some(14) OperatorInputs: TensorShape([2, 577, 1024]) F16 Amp(Left) -> TensorShape([960]) RowMajor",
    "transfers": 1495,
    "receive_fragments": 1495,
    "maximum_endpoint_fragments": 22,
    "transmitted_bytes": 2367488,
    "received_bytes": 2367488,
    "maximum_outgoing_bytes": 26880,
    "maximum_incoming_bytes": 1920,
    "source_staging_max_bytes": 26880,
    "destination_staging_max_bytes": 0,
    "combined_staging_max_bytes": 26880,
    "source_copy_max_bytes": 26880,
    "destination_copy_max_bytes": 0,
    "local_copy_runs": 14720,
    "affine_copy_loops": 1495,
    "copy_cycles_floor": 3360,
    "replay_address_ranges_valid": true
  },
  {
    "phase": 51,
    "mode": "destination",
    "label": "op Some(14) OperatorInputs: TensorShape([2, 577, 1024]) F16 Amp(Left) -> TensorShape([960]) RowMajor",
    "transfers": 14720,
    "receive_fragments": 14720,
    "maximum_endpoint_fragments": 220,
    "transmitted_bytes": 2367488,
    "received_bytes": 2367488,
    "maximum_outgoing_bytes": 26880,
    "maximum_incoming_bytes": 1920,
    "source_staging_max_bytes": 0,
    "destination_staging_max_bytes": 1920,
    "combined_staging_max_bytes": 1920,
    "source_copy_max_bytes": 0,
    "destination_copy_max_bytes": 1920,
    "local_copy_runs": 1495,
    "affine_copy_loops": 1482,
    "copy_cycles_floor": 240,
    "replay_address_ranges_valid": true
  },
  {
    "phase": 51,
    "mode": "both",
    "label": "op Some(14) OperatorInputs: TensorShape([2, 577, 1024]) F16 Amp(Left) -> TensorShape([960]) RowMajor",
    "transfers": 1495,
    "receive_fragments": 1495,
    "maximum_endpoint_fragments": 22,
    "transmitted_bytes": 2367488,
    "received_bytes": 2367488,
    "maximum_outgoing_bytes": 26880,
    "maximum_incoming_bytes": 1920,
    "source_staging_max_bytes": 26880,
    "destination_staging_max_bytes": 1920,
    "combined_staging_max_bytes": 28800,
    "source_copy_max_bytes": 26880,
    "destination_copy_max_bytes": 1920,
    "local_copy_runs": 16215,
    "affine_copy_loops": 2977,
    "copy_cycles_floor": 3600,
    "replay_address_ranges_valid": true
  },
  {
    "phase": 52,
    "mode": "original",
    "label": "op Some(14) OperatorInputs: TensorShape([2, 577, 1024]) F16 Amp(Left) -> TensorShape([960]) RowMajor",
    "transfers": 14720,
    "receive_fragments": 14720,
    "maximum_endpoint_fragments": 220,
    "transmitted_bytes": 2367488,
    "received_bytes": 2367488,
    "maximum_outgoing_bytes": 26880,
    "maximum_incoming_bytes": 1920,
    "source_staging_max_bytes": 0,
    "destination_staging_max_bytes": 0,
    "combined_staging_max_bytes": 0,
    "source_copy_max_bytes": 0,
    "destination_copy_max_bytes": 0,
    "local_copy_runs": 0,
    "affine_copy_loops": 0,
    "copy_cycles_floor": 0,
    "replay_address_ranges_valid": true
  },
  {
    "phase": 52,
    "mode": "source",
    "label": "op Some(14) OperatorInputs: TensorShape([2, 577, 1024]) F16 Amp(Left) -> TensorShape([960]) RowMajor",
    "transfers": 1495,
    "receive_fragments": 1495,
    "maximum_endpoint_fragments": 22,
    "transmitted_bytes": 2367488,
    "received_bytes": 2367488,
    "maximum_outgoing_bytes": 26880,
    "maximum_incoming_bytes": 1920,
    "source_staging_max_bytes": 26880,
    "destination_staging_max_bytes": 0,
    "combined_staging_max_bytes": 26880,
    "source_copy_max_bytes": 26880,
    "destination_copy_max_bytes": 0,
    "local_copy_runs": 14720,
    "affine_copy_loops": 1495,
    "copy_cycles_floor": 3360,
    "replay_address_ranges_valid": true
  },
  {
    "phase": 52,
    "mode": "destination",
    "label": "op Some(14) OperatorInputs: TensorShape([2, 577, 1024]) F16 Amp(Left) -> TensorShape([960]) RowMajor",
    "transfers": 14720,
    "receive_fragments": 14720,
    "maximum_endpoint_fragments": 220,
    "transmitted_bytes": 2367488,
    "received_bytes": 2367488,
    "maximum_outgoing_bytes": 26880,
    "maximum_incoming_bytes": 1920,
    "source_staging_max_bytes": 0,
    "destination_staging_max_bytes": 1920,
    "combined_staging_max_bytes": 1920,
    "source_copy_max_bytes": 0,
    "destination_copy_max_bytes": 1920,
    "local_copy_runs": 1495,
    "affine_copy_loops": 1482,
    "copy_cycles_floor": 240,
    "replay_address_ranges_valid": true
  },
  {
    "phase": 52,
    "mode": "both",
    "label": "op Some(14) OperatorInputs: TensorShape([2, 577, 1024]) F16 Amp(Left) -> TensorShape([960]) RowMajor",
    "transfers": 1495,
    "receive_fragments": 1495,
    "maximum_endpoint_fragments": 22,
    "transmitted_bytes": 2367488,
    "received_bytes": 2367488,
    "maximum_outgoing_bytes": 26880,
    "maximum_incoming_bytes": 1920,
    "source_staging_max_bytes": 26880,
    "destination_staging_max_bytes": 1920,
    "combined_staging_max_bytes": 28800,
    "source_copy_max_bytes": 26880,
    "destination_copy_max_bytes": 1920,
    "local_copy_runs": 16215,
    "affine_copy_loops": 2977,
    "copy_cycles_floor": 3600,
    "replay_address_ranges_valid": true
  },
  {
    "phase": 53,
    "mode": "original",
    "label": "op Some(14) OperatorInputs: TensorShape([2, 577, 1024]) F16 Amp(Left) -> TensorShape([960]) RowMajor",
    "transfers": 14720,
    "receive_fragments": 14720,
    "maximum_endpoint_fragments": 220,
    "transmitted_bytes": 2367488,
    "received_bytes": 2367488,
    "maximum_outgoing_bytes": 26880,
    "maximum_incoming_bytes": 1920,
    "source_staging_max_bytes": 0,
    "destination_staging_max_bytes": 0,
    "combined_staging_max_bytes": 0,
    "source_copy_max_bytes": 0,
    "destination_copy_max_bytes": 0,
    "local_copy_runs": 0,
    "affine_copy_loops": 0,
    "copy_cycles_floor": 0,
    "replay_address_ranges_valid": true
  },
  {
    "phase": 53,
    "mode": "source",
    "label": "op Some(14) OperatorInputs: TensorShape([2, 577, 1024]) F16 Amp(Left) -> TensorShape([960]) RowMajor",
    "transfers": 1495,
    "receive_fragments": 1495,
    "maximum_endpoint_fragments": 22,
    "transmitted_bytes": 2367488,
    "received_bytes": 2367488,
    "maximum_outgoing_bytes": 26880,
    "maximum_incoming_bytes": 1920,
    "source_staging_max_bytes": 26880,
    "destination_staging_max_bytes": 0,
    "combined_staging_max_bytes": 26880,
    "source_copy_max_bytes": 26880,
    "destination_copy_max_bytes": 0,
    "local_copy_runs": 14720,
    "affine_copy_loops": 1495,
    "copy_cycles_floor": 3360,
    "replay_address_ranges_valid": true
  },
  {
    "phase": 53,
    "mode": "destination",
    "label": "op Some(14) OperatorInputs: TensorShape([2, 577, 1024]) F16 Amp(Left) -> TensorShape([960]) RowMajor",
    "transfers": 14720,
    "receive_fragments": 14720,
    "maximum_endpoint_fragments": 220,
    "transmitted_bytes": 2367488,
    "received_bytes": 2367488,
    "maximum_outgoing_bytes": 26880,
    "maximum_incoming_bytes": 1920,
    "source_staging_max_bytes": 0,
    "destination_staging_max_bytes": 1920,
    "combined_staging_max_bytes": 1920,
    "source_copy_max_bytes": 0,
    "destination_copy_max_bytes": 1920,
    "local_copy_runs": 1495,
    "affine_copy_loops": 1482,
    "copy_cycles_floor": 240,
    "replay_address_ranges_valid": true
  },
  {
    "phase": 53,
    "mode": "both",
    "label": "op Some(14) OperatorInputs: TensorShape([2, 577, 1024]) F16 Amp(Left) -> TensorShape([960]) RowMajor",
    "transfers": 1495,
    "receive_fragments": 1495,
    "maximum_endpoint_fragments": 22,
    "transmitted_bytes": 2367488,
    "received_bytes": 2367488,
    "maximum_outgoing_bytes": 26880,
    "maximum_incoming_bytes": 1920,
    "source_staging_max_bytes": 26880,
    "destination_staging_max_bytes": 1920,
    "combined_staging_max_bytes": 28800,
    "source_copy_max_bytes": 26880,
    "destination_copy_max_bytes": 1920,
    "local_copy_runs": 16215,
    "affine_copy_loops": 2977,
    "copy_cycles_floor": 3600,
    "replay_address_ranges_valid": true
  },
  {
    "phase": 39,
    "mode": "original",
    "label": "op Some(14) OperatorInputs: TensorShape([2, 577, 1024]) F16 Amp(Left) -> TensorShape([960]) RowMajor",
    "transfers": 19328,
    "receive_fragments": 19328,
    "maximum_endpoint_fragments": 210,
    "transmitted_bytes": 3104768,
    "received_bytes": 3104768,
    "maximum_outgoing_bytes": 24960,
    "maximum_incoming_bytes": 3840,
    "source_staging_max_bytes": 0,
    "destination_staging_max_bytes": 0,
    "combined_staging_max_bytes": 0,
    "source_copy_max_bytes": 0,
    "destination_copy_max_bytes": 0,
    "local_copy_runs": 0,
    "affine_copy_loops": 0,
    "copy_cycles_floor": 0,
    "replay_address_ranges_valid": true
  },
  {
    "phase": 39,
    "mode": "source",
    "label": "op Some(14) OperatorInputs: TensorShape([2, 577, 1024]) F16 Amp(Left) -> TensorShape([960]) RowMajor",
    "transfers": 1963,
    "receive_fragments": 1963,
    "maximum_endpoint_fragments": 21,
    "transmitted_bytes": 3104768,
    "received_bytes": 3104768,
    "maximum_outgoing_bytes": 24960,
    "maximum_incoming_bytes": 3840,
    "source_staging_max_bytes": 24960,
    "destination_staging_max_bytes": 0,
    "combined_staging_max_bytes": 24960,
    "source_copy_max_bytes": 24960,
    "destination_copy_max_bytes": 0,
    "local_copy_runs": 19328,
    "affine_copy_loops": 1963,
    "copy_cycles_floor": 3120,
    "replay_address_ranges_valid": true
  },
  {
    "phase": 39,
    "mode": "destination",
    "label": "op Some(14) OperatorInputs: TensorShape([2, 577, 1024]) F16 Amp(Left) -> TensorShape([960]) RowMajor",
    "transfers": 19328,
    "receive_fragments": 19328,
    "maximum_endpoint_fragments": 210,
    "transmitted_bytes": 3104768,
    "received_bytes": 3104768,
    "maximum_outgoing_bytes": 24960,
    "maximum_incoming_bytes": 3840,
    "source_staging_max_bytes": 0,
    "destination_staging_max_bytes": 3840,
    "combined_staging_max_bytes": 3840,
    "source_copy_max_bytes": 0,
    "destination_copy_max_bytes": 3840,
    "local_copy_runs": 1469,
    "affine_copy_loops": 1469,
    "copy_cycles_floor": 480,
    "replay_address_ranges_valid": true
  },
  {
    "phase": 39,
    "mode": "both",
    "label": "op Some(14) OperatorInputs: TensorShape([2, 577, 1024]) F16 Amp(Left) -> TensorShape([960]) RowMajor",
    "transfers": 1963,
    "receive_fragments": 1963,
    "maximum_endpoint_fragments": 21,
    "transmitted_bytes": 3104768,
    "received_bytes": 3104768,
    "maximum_outgoing_bytes": 24960,
    "maximum_incoming_bytes": 3840,
    "source_staging_max_bytes": 24960,
    "destination_staging_max_bytes": 3840,
    "combined_staging_max_bytes": 26880,
    "source_copy_max_bytes": 24960,
    "destination_copy_max_bytes": 3840,
    "local_copy_runs": 20797,
    "affine_copy_loops": 3432,
    "copy_cycles_floor": 3600,
    "replay_address_ranges_valid": true
  },
  {
    "phase": 11,
    "mode": "original",
    "label": "op Some(0) LayoutRearrangement: TensorShape([2, 577, 1024]) F16 Amp(Left) -> TensorShape([2, 577, 1024]) RowMajor",
    "transfers": 73796,
    "receive_fragments": 73796,
    "maximum_endpoint_fragments": 200,
    "transmitted_bytes": 2361472,
    "received_bytes": 2361472,
    "maximum_outgoing_bytes": 2304,
    "maximum_incoming_bytes": 4096,
    "source_staging_max_bytes": 0,
    "destination_staging_max_bytes": 0,
    "combined_staging_max_bytes": 0,
    "source_copy_max_bytes": 0,
    "destination_copy_max_bytes": 0,
    "local_copy_runs": 0,
    "affine_copy_loops": 0,
    "copy_cycles_floor": 0,
    "replay_address_ranges_valid": true
  },
  {
    "phase": 11,
    "mode": "source",
    "label": "op Some(0) LayoutRearrangement: TensorShape([2, 577, 1024]) F16 Amp(Left) -> TensorShape([2, 577, 1024]) RowMajor",
    "transfers": 33442,
    "receive_fragments": 33442,
    "maximum_endpoint_fragments": 82,
    "transmitted_bytes": 2361472,
    "received_bytes": 2361472,
    "maximum_outgoing_bytes": 2304,
    "maximum_incoming_bytes": 4096,
    "source_staging_max_bytes": 2304,
    "destination_staging_max_bytes": 0,
    "combined_staging_max_bytes": 2304,
    "source_copy_max_bytes": 2304,
    "destination_copy_max_bytes": 0,
    "local_copy_runs": 73796,
    "affine_copy_loops": 33442,
    "copy_cycles_floor": 288,
    "replay_address_ranges_valid": true
  },
  {
    "phase": 11,
    "mode": "destination",
    "label": "op Some(0) LayoutRearrangement: TensorShape([2, 577, 1024]) F16 Amp(Left) -> TensorShape([2, 577, 1024]) RowMajor",
    "transfers": 73796,
    "receive_fragments": 73796,
    "maximum_endpoint_fragments": 200,
    "transmitted_bytes": 2361472,
    "received_bytes": 2361472,
    "maximum_outgoing_bytes": 2304,
    "maximum_incoming_bytes": 4096,
    "source_staging_max_bytes": 0,
    "destination_staging_max_bytes": 4096,
    "combined_staging_max_bytes": 4096,
    "source_copy_max_bytes": 0,
    "destination_copy_max_bytes": 4096,
    "local_copy_runs": 73796,
    "affine_copy_loops": 36898,
    "copy_cycles_floor": 512,
    "replay_address_ranges_valid": true
  },
  {
    "phase": 11,
    "mode": "both",
    "label": "op Some(0) LayoutRearrangement: TensorShape([2, 577, 1024]) F16 Amp(Left) -> TensorShape([2, 577, 1024]) RowMajor",
    "transfers": 16721,
    "receive_fragments": 16721,
    "maximum_endpoint_fragments": 41,
    "transmitted_bytes": 2361472,
    "received_bytes": 2361472,
    "maximum_outgoing_bytes": 2304,
    "maximum_incoming_bytes": 4096,
    "source_staging_max_bytes": 2304,
    "destination_staging_max_bytes": 4096,
    "combined_staging_max_bytes": 6400,
    "source_copy_max_bytes": 2304,
    "destination_copy_max_bytes": 4096,
    "local_copy_runs": 147592,
    "affine_copy_loops": 53619,
    "copy_cycles_floor": 800,
    "replay_address_ranges_valid": true
  },
  {
    "phase": 48,
    "mode": "original",
    "label": "op Some(14) OperatorInputs: TensorShape([2, 577, 1024]) F16 Amp(Left) -> TensorShape([960]) RowMajor",
    "transfers": 14720,
    "receive_fragments": 14720,
    "maximum_endpoint_fragments": 200,
    "transmitted_bytes": 2367488,
    "received_bytes": 2367488,
    "maximum_outgoing_bytes": 26240,
    "maximum_incoming_bytes": 1920,
    "source_staging_max_bytes": 0,
    "destination_staging_max_bytes": 0,
    "combined_staging_max_bytes": 0,
    "source_copy_max_bytes": 0,
    "destination_copy_max_bytes": 0,
    "local_copy_runs": 0,
    "affine_copy_loops": 0,
    "copy_cycles_floor": 0,
    "replay_address_ranges_valid": true
  },
  {
    "phase": 48,
    "mode": "source",
    "label": "op Some(14) OperatorInputs: TensorShape([2, 577, 1024]) F16 Amp(Left) -> TensorShape([960]) RowMajor",
    "transfers": 1495,
    "receive_fragments": 1495,
    "maximum_endpoint_fragments": 20,
    "transmitted_bytes": 2367488,
    "received_bytes": 2367488,
    "maximum_outgoing_bytes": 26240,
    "maximum_incoming_bytes": 1920,
    "source_staging_max_bytes": 26240,
    "destination_staging_max_bytes": 0,
    "combined_staging_max_bytes": 26240,
    "source_copy_max_bytes": 26240,
    "destination_copy_max_bytes": 0,
    "local_copy_runs": 14720,
    "affine_copy_loops": 1495,
    "copy_cycles_floor": 3280,
    "replay_address_ranges_valid": true
  },
  {
    "phase": 48,
    "mode": "destination",
    "label": "op Some(14) OperatorInputs: TensorShape([2, 577, 1024]) F16 Amp(Left) -> TensorShape([960]) RowMajor",
    "transfers": 14720,
    "receive_fragments": 14720,
    "maximum_endpoint_fragments": 200,
    "transmitted_bytes": 2367488,
    "received_bytes": 2367488,
    "maximum_outgoing_bytes": 26240,
    "maximum_incoming_bytes": 1920,
    "source_staging_max_bytes": 0,
    "destination_staging_max_bytes": 1920,
    "combined_staging_max_bytes": 1920,
    "source_copy_max_bytes": 0,
    "destination_copy_max_bytes": 1920,
    "local_copy_runs": 1495,
    "affine_copy_loops": 1482,
    "copy_cycles_floor": 240,
    "replay_address_ranges_valid": true
  },
  {
    "phase": 48,
    "mode": "both",
    "label": "op Some(14) OperatorInputs: TensorShape([2, 577, 1024]) F16 Amp(Left) -> TensorShape([960]) RowMajor",
    "transfers": 1495,
    "receive_fragments": 1495,
    "maximum_endpoint_fragments": 20,
    "transmitted_bytes": 2367488,
    "received_bytes": 2367488,
    "maximum_outgoing_bytes": 26240,
    "maximum_incoming_bytes": 1920,
    "source_staging_max_bytes": 26240,
    "destination_staging_max_bytes": 1920,
    "combined_staging_max_bytes": 28160,
    "source_copy_max_bytes": 26240,
    "destination_copy_max_bytes": 1920,
    "local_copy_runs": 16215,
    "affine_copy_loops": 2977,
    "copy_cycles_floor": 3520,
    "replay_address_ranges_valid": true
  },
  {
    "phase": 47,
    "mode": "original",
    "label": "op Some(14) OperatorInputs: TensorShape([2, 577, 1024]) F16 Amp(Left) -> TensorShape([960]) RowMajor",
    "transfers": 14720,
    "receive_fragments": 14720,
    "maximum_endpoint_fragments": 190,
    "transmitted_bytes": 2367488,
    "received_bytes": 2367488,
    "maximum_outgoing_bytes": 26240,
    "maximum_incoming_bytes": 1920,
    "source_staging_max_bytes": 0,
    "destination_staging_max_bytes": 0,
    "combined_staging_max_bytes": 0,
    "source_copy_max_bytes": 0,
    "destination_copy_max_bytes": 0,
    "local_copy_runs": 0,
    "affine_copy_loops": 0,
    "copy_cycles_floor": 0,
    "replay_address_ranges_valid": true
  },
  {
    "phase": 47,
    "mode": "source",
    "label": "op Some(14) OperatorInputs: TensorShape([2, 577, 1024]) F16 Amp(Left) -> TensorShape([960]) RowMajor",
    "transfers": 1495,
    "receive_fragments": 1495,
    "maximum_endpoint_fragments": 19,
    "transmitted_bytes": 2367488,
    "received_bytes": 2367488,
    "maximum_outgoing_bytes": 26240,
    "maximum_incoming_bytes": 1920,
    "source_staging_max_bytes": 26240,
    "destination_staging_max_bytes": 0,
    "combined_staging_max_bytes": 26240,
    "source_copy_max_bytes": 26240,
    "destination_copy_max_bytes": 0,
    "local_copy_runs": 14720,
    "affine_copy_loops": 1495,
    "copy_cycles_floor": 3280,
    "replay_address_ranges_valid": true
  },
  {
    "phase": 47,
    "mode": "destination",
    "label": "op Some(14) OperatorInputs: TensorShape([2, 577, 1024]) F16 Amp(Left) -> TensorShape([960]) RowMajor",
    "transfers": 14720,
    "receive_fragments": 14720,
    "maximum_endpoint_fragments": 190,
    "transmitted_bytes": 2367488,
    "received_bytes": 2367488,
    "maximum_outgoing_bytes": 26240,
    "maximum_incoming_bytes": 1920,
    "source_staging_max_bytes": 0,
    "destination_staging_max_bytes": 1920,
    "combined_staging_max_bytes": 1920,
    "source_copy_max_bytes": 0,
    "destination_copy_max_bytes": 1920,
    "local_copy_runs": 1495,
    "affine_copy_loops": 1482,
    "copy_cycles_floor": 240,
    "replay_address_ranges_valid": true
  },
  {
    "phase": 47,
    "mode": "both",
    "label": "op Some(14) OperatorInputs: TensorShape([2, 577, 1024]) F16 Amp(Left) -> TensorShape([960]) RowMajor",
    "transfers": 1495,
    "receive_fragments": 1495,
    "maximum_endpoint_fragments": 19,
    "transmitted_bytes": 2367488,
    "received_bytes": 2367488,
    "maximum_outgoing_bytes": 26240,
    "maximum_incoming_bytes": 1920,
    "source_staging_max_bytes": 26240,
    "destination_staging_max_bytes": 1920,
    "combined_staging_max_bytes": 28160,
    "source_copy_max_bytes": 26240,
    "destination_copy_max_bytes": 1920,
    "local_copy_runs": 16215,
    "affine_copy_loops": 2977,
    "copy_cycles_floor": 3520,
    "replay_address_ranges_valid": true
  },
  {
    "phase": 46,
    "mode": "original",
    "label": "op Some(14) OperatorInputs: TensorShape([2, 577, 1024]) F16 Amp(Left) -> TensorShape([960]) RowMajor",
    "transfers": 14720,
    "receive_fragments": 14720,
    "maximum_endpoint_fragments": 180,
    "transmitted_bytes": 2367488,
    "received_bytes": 2367488,
    "maximum_outgoing_bytes": 26240,
    "maximum_incoming_bytes": 1920,
    "source_staging_max_bytes": 0,
    "destination_staging_max_bytes": 0,
    "combined_staging_max_bytes": 0,
    "source_copy_max_bytes": 0,
    "destination_copy_max_bytes": 0,
    "local_copy_runs": 0,
    "affine_copy_loops": 0,
    "copy_cycles_floor": 0,
    "replay_address_ranges_valid": true
  },
  {
    "phase": 46,
    "mode": "source",
    "label": "op Some(14) OperatorInputs: TensorShape([2, 577, 1024]) F16 Amp(Left) -> TensorShape([960]) RowMajor",
    "transfers": 1495,
    "receive_fragments": 1495,
    "maximum_endpoint_fragments": 18,
    "transmitted_bytes": 2367488,
    "received_bytes": 2367488,
    "maximum_outgoing_bytes": 26240,
    "maximum_incoming_bytes": 1920,
    "source_staging_max_bytes": 26240,
    "destination_staging_max_bytes": 0,
    "combined_staging_max_bytes": 26240,
    "source_copy_max_bytes": 26240,
    "destination_copy_max_bytes": 0,
    "local_copy_runs": 14720,
    "affine_copy_loops": 1495,
    "copy_cycles_floor": 3280,
    "replay_address_ranges_valid": true
  },
  {
    "phase": 46,
    "mode": "destination",
    "label": "op Some(14) OperatorInputs: TensorShape([2, 577, 1024]) F16 Amp(Left) -> TensorShape([960]) RowMajor",
    "transfers": 14720,
    "receive_fragments": 14720,
    "maximum_endpoint_fragments": 180,
    "transmitted_bytes": 2367488,
    "received_bytes": 2367488,
    "maximum_outgoing_bytes": 26240,
    "maximum_incoming_bytes": 1920,
    "source_staging_max_bytes": 0,
    "destination_staging_max_bytes": 1920,
    "combined_staging_max_bytes": 1920,
    "source_copy_max_bytes": 0,
    "destination_copy_max_bytes": 1920,
    "local_copy_runs": 1495,
    "affine_copy_loops": 1482,
    "copy_cycles_floor": 240,
    "replay_address_ranges_valid": true
  },
  {
    "phase": 46,
    "mode": "both",
    "label": "op Some(14) OperatorInputs: TensorShape([2, 577, 1024]) F16 Amp(Left) -> TensorShape([960]) RowMajor",
    "transfers": 1495,
    "receive_fragments": 1495,
    "maximum_endpoint_fragments": 18,
    "transmitted_bytes": 2367488,
    "received_bytes": 2367488,
    "maximum_outgoing_bytes": 26240,
    "maximum_incoming_bytes": 1920,
    "source_staging_max_bytes": 26240,
    "destination_staging_max_bytes": 1920,
    "combined_staging_max_bytes": 28160,
    "source_copy_max_bytes": 26240,
    "destination_copy_max_bytes": 1920,
    "local_copy_runs": 16215,
    "affine_copy_loops": 2977,
    "copy_cycles_floor": 3520,
    "replay_address_ranges_valid": true
  },
  {
    "phase": 45,
    "mode": "original",
    "label": "op Some(14) OperatorInputs: TensorShape([2, 577, 1024]) F16 Amp(Left) -> TensorShape([960]) RowMajor",
    "transfers": 14720,
    "receive_fragments": 14720,
    "maximum_endpoint_fragments": 170,
    "transmitted_bytes": 2367488,
    "received_bytes": 2367488,
    "maximum_outgoing_bytes": 26240,
    "maximum_incoming_bytes": 1920,
    "source_staging_max_bytes": 0,
    "destination_staging_max_bytes": 0,
    "combined_staging_max_bytes": 0,
    "source_copy_max_bytes": 0,
    "destination_copy_max_bytes": 0,
    "local_copy_runs": 0,
    "affine_copy_loops": 0,
    "copy_cycles_floor": 0,
    "replay_address_ranges_valid": true
  },
  {
    "phase": 45,
    "mode": "source",
    "label": "op Some(14) OperatorInputs: TensorShape([2, 577, 1024]) F16 Amp(Left) -> TensorShape([960]) RowMajor",
    "transfers": 1495,
    "receive_fragments": 1495,
    "maximum_endpoint_fragments": 17,
    "transmitted_bytes": 2367488,
    "received_bytes": 2367488,
    "maximum_outgoing_bytes": 26240,
    "maximum_incoming_bytes": 1920,
    "source_staging_max_bytes": 26240,
    "destination_staging_max_bytes": 0,
    "combined_staging_max_bytes": 26240,
    "source_copy_max_bytes": 26240,
    "destination_copy_max_bytes": 0,
    "local_copy_runs": 14720,
    "affine_copy_loops": 1495,
    "copy_cycles_floor": 3280,
    "replay_address_ranges_valid": true
  },
  {
    "phase": 45,
    "mode": "destination",
    "label": "op Some(14) OperatorInputs: TensorShape([2, 577, 1024]) F16 Amp(Left) -> TensorShape([960]) RowMajor",
    "transfers": 14720,
    "receive_fragments": 14720,
    "maximum_endpoint_fragments": 170,
    "transmitted_bytes": 2367488,
    "received_bytes": 2367488,
    "maximum_outgoing_bytes": 26240,
    "maximum_incoming_bytes": 1920,
    "source_staging_max_bytes": 0,
    "destination_staging_max_bytes": 1920,
    "combined_staging_max_bytes": 1920,
    "source_copy_max_bytes": 0,
    "destination_copy_max_bytes": 1920,
    "local_copy_runs": 1495,
    "affine_copy_loops": 1482,
    "copy_cycles_floor": 240,
    "replay_address_ranges_valid": true
  },
  {
    "phase": 45,
    "mode": "both",
    "label": "op Some(14) OperatorInputs: TensorShape([2, 577, 1024]) F16 Amp(Left) -> TensorShape([960]) RowMajor",
    "transfers": 1495,
    "receive_fragments": 1495,
    "maximum_endpoint_fragments": 17,
    "transmitted_bytes": 2367488,
    "received_bytes": 2367488,
    "maximum_outgoing_bytes": 26240,
    "maximum_incoming_bytes": 1920,
    "source_staging_max_bytes": 26240,
    "destination_staging_max_bytes": 1920,
    "combined_staging_max_bytes": 27520,
    "source_copy_max_bytes": 26240,
    "destination_copy_max_bytes": 1920,
    "local_copy_runs": 16215,
    "affine_copy_loops": 2977,
    "copy_cycles_floor": 3520,
    "replay_address_ranges_valid": true
  },
  {
    "phase": 64,
    "mode": "original",
    "label": "op Some(21) OperatorInputs: TensorShape([2, 577, 1024]) F16 Amp(Left) -> TensorShape([1280]) RowMajor",
    "transfers": 66960,
    "receive_fragments": 66960,
    "maximum_endpoint_fragments": 168,
    "transmitted_bytes": 35373568,
    "received_bytes": 35373568,
    "maximum_outgoing_bytes": 25088,
    "maximum_incoming_bytes": 38400,
    "source_staging_max_bytes": 0,
    "destination_staging_max_bytes": 0,
    "combined_staging_max_bytes": 0,
    "source_copy_max_bytes": 0,
    "destination_copy_max_bytes": 0,
    "local_copy_runs": 0,
    "affine_copy_loops": 0,
    "copy_cycles_floor": 0,
    "replay_address_ranges_valid": true
  },
  {
    "phase": 64,
    "mode": "source",
    "label": "op Some(21) OperatorInputs: TensorShape([2, 577, 1024]) F16 Amp(Left) -> TensorShape([1280]) RowMajor",
    "transfers": 25110,
    "receive_fragments": 25110,
    "maximum_endpoint_fragments": 54,
    "transmitted_bytes": 35373568,
    "received_bytes": 35373568,
    "maximum_outgoing_bytes": 25088,
    "maximum_incoming_bytes": 38400,
    "source_staging_max_bytes": 25088,
    "destination_staging_max_bytes": 0,
    "combined_staging_max_bytes": 25088,
    "source_copy_max_bytes": 25088,
    "destination_copy_max_bytes": 0,
    "local_copy_runs": 66960,
    "affine_copy_loops": 8430,
    "copy_cycles_floor": 3136,
    "replay_address_ranges_valid": true
  },
  {
    "phase": 64,
    "mode": "destination",
    "label": "op Some(21) OperatorInputs: TensorShape([2, 577, 1024]) F16 Amp(Left) -> TensorShape([1280]) RowMajor",
    "transfers": 66960,
    "receive_fragments": 66960,
    "maximum_endpoint_fragments": 168,
    "transmitted_bytes": 35373568,
    "received_bytes": 35373568,
    "maximum_outgoing_bytes": 25088,
    "maximum_incoming_bytes": 38400,
    "source_staging_max_bytes": 0,
    "destination_staging_max_bytes": 38400,
    "combined_staging_max_bytes": 38400,
    "source_copy_max_bytes": 0,
    "destination_copy_max_bytes": 38400,
    "local_copy_runs": 3237,
    "affine_copy_loops": 3237,
    "copy_cycles_floor": 4800,
    "replay_address_ranges_valid": true
  },
  {
    "phase": 64,
    "mode": "both",
    "label": "op Some(21) OperatorInputs: TensorShape([2, 577, 1024]) F16 Amp(Left) -> TensorShape([1280]) RowMajor",
    "transfers": 25110,
    "receive_fragments": 25110,
    "maximum_endpoint_fragments": 54,
    "transmitted_bytes": 35373568,
    "received_bytes": 35373568,
    "maximum_outgoing_bytes": 25088,
    "maximum_incoming_bytes": 38400,
    "source_staging_max_bytes": 25088,
    "destination_staging_max_bytes": 38400,
    "combined_staging_max_bytes": 63488,
    "source_copy_max_bytes": 25088,
    "destination_copy_max_bytes": 38400,
    "local_copy_runs": 70197,
    "affine_copy_loops": 11667,
    "copy_cycles_floor": 7936,
    "replay_address_ranges_valid": true
  },
  {
    "phase": 44,
    "mode": "original",
    "label": "op Some(14) OperatorInputs: TensorShape([2, 577, 1024]) F16 Amp(Left) -> TensorShape([960]) RowMajor",
    "transfers": 14720,
    "receive_fragments": 14720,
    "maximum_endpoint_fragments": 160,
    "transmitted_bytes": 2367488,
    "received_bytes": 2367488,
    "maximum_outgoing_bytes": 24960,
    "maximum_incoming_bytes": 1920,
    "source_staging_max_bytes": 0,
    "destination_staging_max_bytes": 0,
    "combined_staging_max_bytes": 0,
    "source_copy_max_bytes": 0,
    "destination_copy_max_bytes": 0,
    "local_copy_runs": 0,
    "affine_copy_loops": 0,
    "copy_cycles_floor": 0,
    "replay_address_ranges_valid": true
  },
  {
    "phase": 44,
    "mode": "source",
    "label": "op Some(14) OperatorInputs: TensorShape([2, 577, 1024]) F16 Amp(Left) -> TensorShape([960]) RowMajor",
    "transfers": 1495,
    "receive_fragments": 1495,
    "maximum_endpoint_fragments": 16,
    "transmitted_bytes": 2367488,
    "received_bytes": 2367488,
    "maximum_outgoing_bytes": 24960,
    "maximum_incoming_bytes": 1920,
    "source_staging_max_bytes": 24960,
    "destination_staging_max_bytes": 0,
    "combined_staging_max_bytes": 24960,
    "source_copy_max_bytes": 24960,
    "destination_copy_max_bytes": 0,
    "local_copy_runs": 14720,
    "affine_copy_loops": 1495,
    "copy_cycles_floor": 3120,
    "replay_address_ranges_valid": true
  },
  {
    "phase": 44,
    "mode": "destination",
    "label": "op Some(14) OperatorInputs: TensorShape([2, 577, 1024]) F16 Amp(Left) -> TensorShape([960]) RowMajor",
    "transfers": 14720,
    "receive_fragments": 14720,
    "maximum_endpoint_fragments": 160,
    "transmitted_bytes": 2367488,
    "received_bytes": 2367488,
    "maximum_outgoing_bytes": 24960,
    "maximum_incoming_bytes": 1920,
    "source_staging_max_bytes": 0,
    "destination_staging_max_bytes": 1920,
    "combined_staging_max_bytes": 1920,
    "source_copy_max_bytes": 0,
    "destination_copy_max_bytes": 1920,
    "local_copy_runs": 1495,
    "affine_copy_loops": 1482,
    "copy_cycles_floor": 240,
    "replay_address_ranges_valid": true
  },
  {
    "phase": 44,
    "mode": "both",
    "label": "op Some(14) OperatorInputs: TensorShape([2, 577, 1024]) F16 Amp(Left) -> TensorShape([960]) RowMajor",
    "transfers": 1495,
    "receive_fragments": 1495,
    "maximum_endpoint_fragments": 16,
    "transmitted_bytes": 2367488,
    "received_bytes": 2367488,
    "maximum_outgoing_bytes": 24960,
    "maximum_incoming_bytes": 1920,
    "source_staging_max_bytes": 24960,
    "destination_staging_max_bytes": 1920,
    "combined_staging_max_bytes": 26240,
    "source_copy_max_bytes": 24960,
    "destination_copy_max_bytes": 1920,
    "local_copy_runs": 16215,
    "affine_copy_loops": 2977,
    "copy_cycles_floor": 3360,
    "replay_address_ranges_valid": true
  },
  {
    "phase": 65,
    "mode": "original",
    "label": "op Some(22) OperatorInputs: TensorShape([2, 577, 1024]) F16 Amp(Left) -> TensorShape([2, 577, 1024]) RowMajor",
    "transfers": 73800,
    "receive_fragments": 83024,
    "maximum_endpoint_fragments": 152,
    "transmitted_bytes": 2363392,
    "received_bytes": 4724736,
    "maximum_outgoing_bytes": 2560,
    "maximum_incoming_bytes": 4096,
    "source_staging_max_bytes": 0,
    "destination_staging_max_bytes": 0,
    "combined_staging_max_bytes": 0,
    "source_copy_max_bytes": 0,
    "destination_copy_max_bytes": 0,
    "local_copy_runs": 0,
    "affine_copy_loops": 0,
    "copy_cycles_floor": 0,
    "replay_address_ranges_valid": true
  },
  {
    "phase": 65,
    "mode": "source",
    "label": "op Some(22) OperatorInputs: TensorShape([2, 577, 1024]) F16 Amp(Left) -> TensorShape([2, 577, 1024]) RowMajor",
    "transfers": 55356,
    "receive_fragments": 64580,
    "maximum_endpoint_fragments": 97,
    "transmitted_bytes": 2363392,
    "received_bytes": 4724736,
    "maximum_outgoing_bytes": 2560,
    "maximum_incoming_bytes": 4096,
    "source_staging_max_bytes": 2560,
    "destination_staging_max_bytes": 0,
    "combined_staging_max_bytes": 2560,
    "source_copy_max_bytes": 2560,
    "destination_copy_max_bytes": 0,
    "local_copy_runs": 37911,
    "affine_copy_loops": 19435,
    "copy_cycles_floor": 320,
    "replay_address_ranges_valid": true
  },
  {
    "phase": 65,
    "mode": "destination",
    "label": "op Some(22) OperatorInputs: TensorShape([2, 577, 1024]) F16 Amp(Left) -> TensorShape([2, 577, 1024]) RowMajor",
    "transfers": 73800,
    "receive_fragments": 83024,
    "maximum_endpoint_fragments": 152,
    "transmitted_bytes": 2363392,
    "received_bytes": 4724736,
    "maximum_outgoing_bytes": 2560,
    "maximum_incoming_bytes": 4096,
    "source_staging_max_bytes": 0,
    "destination_staging_max_bytes": 4096,
    "combined_staging_max_bytes": 4096,
    "source_copy_max_bytes": 0,
    "destination_copy_max_bytes": 4096,
    "local_copy_runs": 2389,
    "affine_copy_loops": 1859,
    "copy_cycles_floor": 512,
    "replay_address_ranges_valid": true
  },
  {
    "phase": 65,
    "mode": "both",
    "label": "op Some(22) OperatorInputs: TensorShape([2, 577, 1024]) F16 Amp(Left) -> TensorShape([2, 577, 1024]) RowMajor",
    "transfers": 55356,
    "receive_fragments": 64580,
    "maximum_endpoint_fragments": 97,
    "transmitted_bytes": 2363392,
    "received_bytes": 4724736,
    "maximum_outgoing_bytes": 2560,
    "maximum_incoming_bytes": 4096,
    "source_staging_max_bytes": 2560,
    "destination_staging_max_bytes": 4096,
    "combined_staging_max_bytes": 6656,
    "source_copy_max_bytes": 2560,
    "destination_copy_max_bytes": 4096,
    "local_copy_runs": 40300,
    "affine_copy_loops": 21294,
    "copy_cycles_floor": 832,
    "replay_address_ranges_valid": true
  },
  {
    "phase": 40,
    "mode": "original",
    "label": "op Some(14) OperatorInputs: TensorShape([2, 577, 1024]) F16 Amp(Left) -> TensorShape([960]) RowMajor",
    "transfers": 14720,
    "receive_fragments": 14720,
    "maximum_endpoint_fragments": 150,
    "transmitted_bytes": 2367488,
    "received_bytes": 2367488,
    "maximum_outgoing_bytes": 23040,
    "maximum_incoming_bytes": 1920,
    "source_staging_max_bytes": 0,
    "destination_staging_max_bytes": 0,
    "combined_staging_max_bytes": 0,
    "source_copy_max_bytes": 0,
    "destination_copy_max_bytes": 0,
    "local_copy_runs": 0,
    "affine_copy_loops": 0,
    "copy_cycles_floor": 0,
    "replay_address_ranges_valid": true
  },
  {
    "phase": 40,
    "mode": "source",
    "label": "op Some(14) OperatorInputs: TensorShape([2, 577, 1024]) F16 Amp(Left) -> TensorShape([960]) RowMajor",
    "transfers": 1495,
    "receive_fragments": 1495,
    "maximum_endpoint_fragments": 15,
    "transmitted_bytes": 2367488,
    "received_bytes": 2367488,
    "maximum_outgoing_bytes": 23040,
    "maximum_incoming_bytes": 1920,
    "source_staging_max_bytes": 23040,
    "destination_staging_max_bytes": 0,
    "combined_staging_max_bytes": 23040,
    "source_copy_max_bytes": 23040,
    "destination_copy_max_bytes": 0,
    "local_copy_runs": 14720,
    "affine_copy_loops": 1495,
    "copy_cycles_floor": 2880,
    "replay_address_ranges_valid": true
  },
  {
    "phase": 40,
    "mode": "destination",
    "label": "op Some(14) OperatorInputs: TensorShape([2, 577, 1024]) F16 Amp(Left) -> TensorShape([960]) RowMajor",
    "transfers": 14720,
    "receive_fragments": 14720,
    "maximum_endpoint_fragments": 150,
    "transmitted_bytes": 2367488,
    "received_bytes": 2367488,
    "maximum_outgoing_bytes": 23040,
    "maximum_incoming_bytes": 1920,
    "source_staging_max_bytes": 0,
    "destination_staging_max_bytes": 1920,
    "combined_staging_max_bytes": 1920,
    "source_copy_max_bytes": 0,
    "destination_copy_max_bytes": 1920,
    "local_copy_runs": 1495,
    "affine_copy_loops": 1482,
    "copy_cycles_floor": 240,
    "replay_address_ranges_valid": true
  },
  {
    "phase": 40,
    "mode": "both",
    "label": "op Some(14) OperatorInputs: TensorShape([2, 577, 1024]) F16 Amp(Left) -> TensorShape([960]) RowMajor",
    "transfers": 1495,
    "receive_fragments": 1495,
    "maximum_endpoint_fragments": 15,
    "transmitted_bytes": 2367488,
    "received_bytes": 2367488,
    "maximum_outgoing_bytes": 23040,
    "maximum_incoming_bytes": 1920,
    "source_staging_max_bytes": 23040,
    "destination_staging_max_bytes": 1920,
    "combined_staging_max_bytes": 24960,
    "source_copy_max_bytes": 23040,
    "destination_copy_max_bytes": 1920,
    "local_copy_runs": 16215,
    "affine_copy_loops": 2977,
    "copy_cycles_floor": 3120,
    "replay_address_ranges_valid": true
  },
  {
    "phase": 43,
    "mode": "original",
    "label": "op Some(14) OperatorInputs: TensorShape([2, 577, 1024]) F16 Amp(Left) -> TensorShape([960]) RowMajor",
    "transfers": 14720,
    "receive_fragments": 14720,
    "maximum_endpoint_fragments": 150,
    "transmitted_bytes": 2367488,
    "received_bytes": 2367488,
    "maximum_outgoing_bytes": 23680,
    "maximum_incoming_bytes": 1920,
    "source_staging_max_bytes": 0,
    "destination_staging_max_bytes": 0,
    "combined_staging_max_bytes": 0,
    "source_copy_max_bytes": 0,
    "destination_copy_max_bytes": 0,
    "local_copy_runs": 0,
    "affine_copy_loops": 0,
    "copy_cycles_floor": 0,
    "replay_address_ranges_valid": true
  },
  {
    "phase": 43,
    "mode": "source",
    "label": "op Some(14) OperatorInputs: TensorShape([2, 577, 1024]) F16 Amp(Left) -> TensorShape([960]) RowMajor",
    "transfers": 1495,
    "receive_fragments": 1495,
    "maximum_endpoint_fragments": 15,
    "transmitted_bytes": 2367488,
    "received_bytes": 2367488,
    "maximum_outgoing_bytes": 23680,
    "maximum_incoming_bytes": 1920,
    "source_staging_max_bytes": 23680,
    "destination_staging_max_bytes": 0,
    "combined_staging_max_bytes": 23680,
    "source_copy_max_bytes": 23680,
    "destination_copy_max_bytes": 0,
    "local_copy_runs": 14720,
    "affine_copy_loops": 1495,
    "copy_cycles_floor": 2960,
    "replay_address_ranges_valid": true
  },
  {
    "phase": 43,
    "mode": "destination",
    "label": "op Some(14) OperatorInputs: TensorShape([2, 577, 1024]) F16 Amp(Left) -> TensorShape([960]) RowMajor",
    "transfers": 14720,
    "receive_fragments": 14720,
    "maximum_endpoint_fragments": 150,
    "transmitted_bytes": 2367488,
    "received_bytes": 2367488,
    "maximum_outgoing_bytes": 23680,
    "maximum_incoming_bytes": 1920,
    "source_staging_max_bytes": 0,
    "destination_staging_max_bytes": 1920,
    "combined_staging_max_bytes": 1920,
    "source_copy_max_bytes": 0,
    "destination_copy_max_bytes": 1920,
    "local_copy_runs": 1495,
    "affine_copy_loops": 1482,
    "copy_cycles_floor": 240,
    "replay_address_ranges_valid": true
  },
  {
    "phase": 43,
    "mode": "both",
    "label": "op Some(14) OperatorInputs: TensorShape([2, 577, 1024]) F16 Amp(Left) -> TensorShape([960]) RowMajor",
    "transfers": 1495,
    "receive_fragments": 1495,
    "maximum_endpoint_fragments": 15,
    "transmitted_bytes": 2367488,
    "received_bytes": 2367488,
    "maximum_outgoing_bytes": 23680,
    "maximum_incoming_bytes": 1920,
    "source_staging_max_bytes": 23680,
    "destination_staging_max_bytes": 1920,
    "combined_staging_max_bytes": 24960,
    "source_copy_max_bytes": 23680,
    "destination_copy_max_bytes": 1920,
    "local_copy_runs": 16215,
    "affine_copy_loops": 2977,
    "copy_cycles_floor": 3200,
    "replay_address_ranges_valid": true
  },
  {
    "phase": 41,
    "mode": "original",
    "label": "op Some(14) OperatorInputs: TensorShape([2, 577, 1024]) F16 Amp(Left) -> TensorShape([960]) RowMajor",
    "transfers": 14720,
    "receive_fragments": 14720,
    "maximum_endpoint_fragments": 140,
    "transmitted_bytes": 2367488,
    "received_bytes": 2367488,
    "maximum_outgoing_bytes": 21120,
    "maximum_incoming_bytes": 1920,
    "source_staging_max_bytes": 0,
    "destination_staging_max_bytes": 0,
    "combined_staging_max_bytes": 0,
    "source_copy_max_bytes": 0,
    "destination_copy_max_bytes": 0,
    "local_copy_runs": 0,
    "affine_copy_loops": 0,
    "copy_cycles_floor": 0,
    "replay_address_ranges_valid": true
  },
  {
    "phase": 41,
    "mode": "source",
    "label": "op Some(14) OperatorInputs: TensorShape([2, 577, 1024]) F16 Amp(Left) -> TensorShape([960]) RowMajor",
    "transfers": 1495,
    "receive_fragments": 1495,
    "maximum_endpoint_fragments": 14,
    "transmitted_bytes": 2367488,
    "received_bytes": 2367488,
    "maximum_outgoing_bytes": 21120,
    "maximum_incoming_bytes": 1920,
    "source_staging_max_bytes": 21120,
    "destination_staging_max_bytes": 0,
    "combined_staging_max_bytes": 21120,
    "source_copy_max_bytes": 21120,
    "destination_copy_max_bytes": 0,
    "local_copy_runs": 14720,
    "affine_copy_loops": 1495,
    "copy_cycles_floor": 2640,
    "replay_address_ranges_valid": true
  },
  {
    "phase": 41,
    "mode": "destination",
    "label": "op Some(14) OperatorInputs: TensorShape([2, 577, 1024]) F16 Amp(Left) -> TensorShape([960]) RowMajor",
    "transfers": 14720,
    "receive_fragments": 14720,
    "maximum_endpoint_fragments": 140,
    "transmitted_bytes": 2367488,
    "received_bytes": 2367488,
    "maximum_outgoing_bytes": 21120,
    "maximum_incoming_bytes": 1920,
    "source_staging_max_bytes": 0,
    "destination_staging_max_bytes": 1920,
    "combined_staging_max_bytes": 1920,
    "source_copy_max_bytes": 0,
    "destination_copy_max_bytes": 1920,
    "local_copy_runs": 1495,
    "affine_copy_loops": 1482,
    "copy_cycles_floor": 240,
    "replay_address_ranges_valid": true
  },
  {
    "phase": 41,
    "mode": "both",
    "label": "op Some(14) OperatorInputs: TensorShape([2, 577, 1024]) F16 Amp(Left) -> TensorShape([960]) RowMajor",
    "transfers": 1495,
    "receive_fragments": 1495,
    "maximum_endpoint_fragments": 14,
    "transmitted_bytes": 2367488,
    "received_bytes": 2367488,
    "maximum_outgoing_bytes": 21120,
    "maximum_incoming_bytes": 1920,
    "source_staging_max_bytes": 21120,
    "destination_staging_max_bytes": 1920,
    "combined_staging_max_bytes": 23040,
    "source_copy_max_bytes": 21120,
    "destination_copy_max_bytes": 1920,
    "local_copy_runs": 16215,
    "affine_copy_loops": 2977,
    "copy_cycles_floor": 2880,
    "replay_address_ranges_valid": true
  },
  {
    "phase": 42,
    "mode": "original",
    "label": "op Some(14) OperatorInputs: TensorShape([2, 577, 1024]) F16 Amp(Left) -> TensorShape([960]) RowMajor",
    "transfers": 14720,
    "receive_fragments": 14720,
    "maximum_endpoint_fragments": 140,
    "transmitted_bytes": 2367488,
    "received_bytes": 2367488,
    "maximum_outgoing_bytes": 22400,
    "maximum_incoming_bytes": 1920,
    "source_staging_max_bytes": 0,
    "destination_staging_max_bytes": 0,
    "combined_staging_max_bytes": 0,
    "source_copy_max_bytes": 0,
    "destination_copy_max_bytes": 0,
    "local_copy_runs": 0,
    "affine_copy_loops": 0,
    "copy_cycles_floor": 0,
    "replay_address_ranges_valid": true
  },
  {
    "phase": 42,
    "mode": "source",
    "label": "op Some(14) OperatorInputs: TensorShape([2, 577, 1024]) F16 Amp(Left) -> TensorShape([960]) RowMajor",
    "transfers": 1495,
    "receive_fragments": 1495,
    "maximum_endpoint_fragments": 14,
    "transmitted_bytes": 2367488,
    "received_bytes": 2367488,
    "maximum_outgoing_bytes": 22400,
    "maximum_incoming_bytes": 1920,
    "source_staging_max_bytes": 22400,
    "destination_staging_max_bytes": 0,
    "combined_staging_max_bytes": 22400,
    "source_copy_max_bytes": 22400,
    "destination_copy_max_bytes": 0,
    "local_copy_runs": 14720,
    "affine_copy_loops": 1495,
    "copy_cycles_floor": 2800,
    "replay_address_ranges_valid": true
  },
  {
    "phase": 42,
    "mode": "destination",
    "label": "op Some(14) OperatorInputs: TensorShape([2, 577, 1024]) F16 Amp(Left) -> TensorShape([960]) RowMajor",
    "transfers": 14720,
    "receive_fragments": 14720,
    "maximum_endpoint_fragments": 140,
    "transmitted_bytes": 2367488,
    "received_bytes": 2367488,
    "maximum_outgoing_bytes": 22400,
    "maximum_incoming_bytes": 1920,
    "source_staging_max_bytes": 0,
    "destination_staging_max_bytes": 1920,
    "combined_staging_max_bytes": 1920,
    "source_copy_max_bytes": 0,
    "destination_copy_max_bytes": 1920,
    "local_copy_runs": 1495,
    "affine_copy_loops": 1482,
    "copy_cycles_floor": 240,
    "replay_address_ranges_valid": true
  },
  {
    "phase": 42,
    "mode": "both",
    "label": "op Some(14) OperatorInputs: TensorShape([2, 577, 1024]) F16 Amp(Left) -> TensorShape([960]) RowMajor",
    "transfers": 1495,
    "receive_fragments": 1495,
    "maximum_endpoint_fragments": 14,
    "transmitted_bytes": 2367488,
    "received_bytes": 2367488,
    "maximum_outgoing_bytes": 22400,
    "maximum_incoming_bytes": 1920,
    "source_staging_max_bytes": 22400,
    "destination_staging_max_bytes": 1920,
    "combined_staging_max_bytes": 23680,
    "source_copy_max_bytes": 22400,
    "destination_copy_max_bytes": 1920,
    "local_copy_runs": 16215,
    "affine_copy_loops": 2977,
    "copy_cycles_floor": 3040,
    "replay_address_ranges_valid": true
  },
  {
    "phase": 54,
    "mode": "original",
    "label": "op Some(15) OperatorInputs: TensorShape([2, 577, 1024]) F16 Amp(Left) -> TensorShape([2, 577, 1024]) RowMajor",
    "transfers": 73814,
    "receive_fragments": 83038,
    "maximum_endpoint_fragments": 133,
    "transmitted_bytes": 2363840,
    "received_bytes": 4725184,
    "maximum_outgoing_bytes": 2176,
    "maximum_incoming_bytes": 4096,
    "source_staging_max_bytes": 0,
    "destination_staging_max_bytes": 0,
    "combined_staging_max_bytes": 0,
    "source_copy_max_bytes": 0,
    "destination_copy_max_bytes": 0,
    "local_copy_runs": 0,
    "affine_copy_loops": 0,
    "copy_cycles_floor": 0,
    "replay_address_ranges_valid": true
  },
  {
    "phase": 54,
    "mode": "source",
    "label": "op Some(15) OperatorInputs: TensorShape([2, 577, 1024]) F16 Amp(Left) -> TensorShape([2, 577, 1024]) RowMajor",
    "transfers": 15000,
    "receive_fragments": 24224,
    "maximum_endpoint_fragments": 34,
    "transmitted_bytes": 2363840,
    "received_bytes": 4725184,
    "maximum_outgoing_bytes": 2176,
    "maximum_incoming_bytes": 4096,
    "source_staging_max_bytes": 2176,
    "destination_staging_max_bytes": 0,
    "combined_staging_max_bytes": 2176,
    "source_copy_max_bytes": 2176,
    "destination_copy_max_bytes": 0,
    "local_copy_runs": 73814,
    "affine_copy_loops": 15000,
    "copy_cycles_floor": 272,
    "replay_address_ranges_valid": true
  },
  {
    "phase": 54,
    "mode": "destination",
    "label": "op Some(15) OperatorInputs: TensorShape([2, 577, 1024]) F16 Amp(Left) -> TensorShape([2, 577, 1024]) RowMajor",
    "transfers": 73814,
    "receive_fragments": 83038,
    "maximum_endpoint_fragments": 133,
    "transmitted_bytes": 2363840,
    "received_bytes": 4725184,
    "maximum_outgoing_bytes": 2176,
    "maximum_incoming_bytes": 4096,
    "source_staging_max_bytes": 0,
    "destination_staging_max_bytes": 4096,
    "combined_staging_max_bytes": 4096,
    "source_copy_max_bytes": 0,
    "destination_copy_max_bytes": 4096,
    "local_copy_runs": 2293,
    "affine_copy_loops": 1359,
    "copy_cycles_floor": 512,
    "replay_address_ranges_valid": true
  },
  {
    "phase": 54,
    "mode": "both",
    "label": "op Some(15) OperatorInputs: TensorShape([2, 577, 1024]) F16 Amp(Left) -> TensorShape([2, 577, 1024]) RowMajor",
    "transfers": 15000,
    "receive_fragments": 24224,
    "maximum_endpoint_fragments": 34,
    "transmitted_bytes": 2363840,
    "received_bytes": 4725184,
    "maximum_outgoing_bytes": 2176,
    "maximum_incoming_bytes": 4096,
    "source_staging_max_bytes": 2176,
    "destination_staging_max_bytes": 4096,
    "combined_staging_max_bytes": 6272,
    "source_copy_max_bytes": 2176,
    "destination_copy_max_bytes": 4096,
    "local_copy_runs": 76107,
    "affine_copy_loops": 16359,
    "copy_cycles_floor": 784,
    "replay_address_ranges_valid": true
  },
  {
    "phase": 110,
    "mode": "original",
    "label": "op Some(46) LayoutRearrangement: TensorShape([2, 1, 1024]) F16 Amp(Left) -> TensorShape([2, 1, 1024]) RowMajor",
    "transfers": 126,
    "receive_fragments": 126,
    "maximum_endpoint_fragments": 126,
    "transmitted_bytes": 4032,
    "received_bytes": 4032,
    "maximum_outgoing_bytes": 64,
    "maximum_incoming_bytes": 4032,
    "source_staging_max_bytes": 0,
    "destination_staging_max_bytes": 0,
    "combined_staging_max_bytes": 0,
    "source_copy_max_bytes": 0,
    "destination_copy_max_bytes": 0,
    "local_copy_runs": 0,
    "affine_copy_loops": 0,
    "copy_cycles_floor": 0,
    "replay_address_ranges_valid": true
  },
  {
    "phase": 110,
    "mode": "source",
    "label": "op Some(46) LayoutRearrangement: TensorShape([2, 1, 1024]) F16 Amp(Left) -> TensorShape([2, 1, 1024]) RowMajor",
    "transfers": 126,
    "receive_fragments": 126,
    "maximum_endpoint_fragments": 126,
    "transmitted_bytes": 4032,
    "received_bytes": 4032,
    "maximum_outgoing_bytes": 64,
    "maximum_incoming_bytes": 4032,
    "source_staging_max_bytes": 64,
    "destination_staging_max_bytes": 0,
    "combined_staging_max_bytes": 64,
    "source_copy_max_bytes": 64,
    "destination_copy_max_bytes": 0,
    "local_copy_runs": 126,
    "affine_copy_loops": 63,
    "copy_cycles_floor": 8,
    "replay_address_ranges_valid": true
  },
  {
    "phase": 110,
    "mode": "destination",
    "label": "op Some(46) LayoutRearrangement: TensorShape([2, 1, 1024]) F16 Amp(Left) -> TensorShape([2, 1, 1024]) RowMajor",
    "transfers": 126,
    "receive_fragments": 126,
    "maximum_endpoint_fragments": 126,
    "transmitted_bytes": 4032,
    "received_bytes": 4032,
    "maximum_outgoing_bytes": 64,
    "maximum_incoming_bytes": 4032,
    "source_staging_max_bytes": 0,
    "destination_staging_max_bytes": 4032,
    "combined_staging_max_bytes": 4032,
    "source_copy_max_bytes": 0,
    "destination_copy_max_bytes": 4032,
    "local_copy_runs": 126,
    "affine_copy_loops": 63,
    "copy_cycles_floor": 504,
    "replay_address_ranges_valid": true
  },
  {
    "phase": 110,
    "mode": "both",
    "label": "op Some(46) LayoutRearrangement: TensorShape([2, 1, 1024]) F16 Amp(Left) -> TensorShape([2, 1, 1024]) RowMajor",
    "transfers": 63,
    "receive_fragments": 63,
    "maximum_endpoint_fragments": 63,
    "transmitted_bytes": 4032,
    "received_bytes": 4032,
    "maximum_outgoing_bytes": 64,
    "maximum_incoming_bytes": 4032,
    "source_staging_max_bytes": 64,
    "destination_staging_max_bytes": 4032,
    "combined_staging_max_bytes": 4032,
    "source_copy_max_bytes": 64,
    "destination_copy_max_bytes": 4032,
    "local_copy_runs": 252,
    "affine_copy_loops": 126,
    "copy_cycles_floor": 512,
    "replay_address_ranges_valid": true
  },
  {
    "phase": 2,
    "mode": "original",
    "label": "op Some(0) OperatorInputs: TensorShape([2, 577, 1024]) F16 Amp(Left) -> TensorShape([1152]) RowMajor",
    "transfers": 7424,
    "receive_fragments": 7424,
    "maximum_endpoint_fragments": 72,
    "transmitted_bytes": 2523136,
    "received_bytes": 2523136,
    "maximum_outgoing_bytes": 19968,
    "maximum_incoming_bytes": 4608,
    "source_staging_max_bytes": 0,
    "destination_staging_max_bytes": 0,
    "combined_staging_max_bytes": 0,
    "source_copy_max_bytes": 0,
    "destination_copy_max_bytes": 0,
    "local_copy_runs": 0,
    "affine_copy_loops": 0,
    "copy_cycles_floor": 0,
    "replay_address_ranges_valid": true
  },
  {
    "phase": 2,
    "mode": "source",
    "label": "op Some(0) OperatorInputs: TensorShape([2, 577, 1024]) F16 Amp(Left) -> TensorShape([1152]) RowMajor",
    "transfers": 1682,
    "receive_fragments": 1682,
    "maximum_endpoint_fragments": 12,
    "transmitted_bytes": 2523136,
    "received_bytes": 2523136,
    "maximum_outgoing_bytes": 19968,
    "maximum_incoming_bytes": 4608,
    "source_staging_max_bytes": 19968,
    "destination_staging_max_bytes": 0,
    "combined_staging_max_bytes": 19968,
    "source_copy_max_bytes": 19968,
    "destination_copy_max_bytes": 0,
    "local_copy_runs": 7424,
    "affine_copy_loops": 1682,
    "copy_cycles_floor": 2496,
    "replay_address_ranges_valid": true
  },
  {
    "phase": 2,
    "mode": "destination",
    "label": "op Some(0) OperatorInputs: TensorShape([2, 577, 1024]) F16 Amp(Left) -> TensorShape([1152]) RowMajor",
    "transfers": 7424,
    "receive_fragments": 7424,
    "maximum_endpoint_fragments": 72,
    "transmitted_bytes": 2523136,
    "received_bytes": 2523136,
    "maximum_outgoing_bytes": 19968,
    "maximum_incoming_bytes": 4608,
    "source_staging_max_bytes": 0,
    "destination_staging_max_bytes": 4608,
    "combined_staging_max_bytes": 4608,
    "source_copy_max_bytes": 0,
    "destination_copy_max_bytes": 4608,
    "local_copy_runs": 1479,
    "affine_copy_loops": 1479,
    "copy_cycles_floor": 576,
    "replay_address_ranges_valid": true
  },
  {
    "phase": 2,
    "mode": "both",
    "label": "op Some(0) OperatorInputs: TensorShape([2, 577, 1024]) F16 Amp(Left) -> TensorShape([1152]) RowMajor",
    "transfers": 1682,
    "receive_fragments": 1682,
    "maximum_endpoint_fragments": 12,
    "transmitted_bytes": 2523136,
    "received_bytes": 2523136,
    "maximum_outgoing_bytes": 19968,
    "maximum_incoming_bytes": 4608,
    "source_staging_max_bytes": 19968,
    "destination_staging_max_bytes": 4608,
    "combined_staging_max_bytes": 22272,
    "source_copy_max_bytes": 19968,
    "destination_copy_max_bytes": 4608,
    "local_copy_runs": 8903,
    "affine_copy_loops": 3161,
    "copy_cycles_floor": 3072,
    "replay_address_ranges_valid": true
  },
  {
    "phase": 10,
    "mode": "original",
    "label": "op Some(0) OperatorInputs: TensorShape([2, 577, 1024]) F16 Amp(Left) -> TensorShape([768]) RowMajor",
    "transfers": 6784,
    "receive_fragments": 6784,
    "maximum_endpoint_fragments": 72,
    "transmitted_bytes": 2367488,
    "received_bytes": 2367488,
    "maximum_outgoing_bytes": 22272,
    "maximum_incoming_bytes": 2304,
    "source_staging_max_bytes": 0,
    "destination_staging_max_bytes": 0,
    "combined_staging_max_bytes": 0,
    "source_copy_max_bytes": 0,
    "destination_copy_max_bytes": 0,
    "local_copy_runs": 0,
    "affine_copy_loops": 0,
    "copy_cycles_floor": 0,
    "replay_address_ranges_valid": true
  },
  {
    "phase": 10,
    "mode": "source",
    "label": "op Some(0) OperatorInputs: TensorShape([2, 577, 1024]) F16 Amp(Left) -> TensorShape([768]) RowMajor",
    "transfers": 1537,
    "receive_fragments": 1537,
    "maximum_endpoint_fragments": 12,
    "transmitted_bytes": 2367488,
    "received_bytes": 2367488,
    "maximum_outgoing_bytes": 22272,
    "maximum_incoming_bytes": 2304,
    "source_staging_max_bytes": 22272,
    "destination_staging_max_bytes": 0,
    "combined_staging_max_bytes": 22272,
    "source_copy_max_bytes": 22272,
    "destination_copy_max_bytes": 0,
    "local_copy_runs": 6784,
    "affine_copy_loops": 1537,
    "copy_cycles_floor": 2784,
    "replay_address_ranges_valid": true
  },
  {
    "phase": 10,
    "mode": "destination",
    "label": "op Some(0) OperatorInputs: TensorShape([2, 577, 1024]) F16 Amp(Left) -> TensorShape([768]) RowMajor",
    "transfers": 6784,
    "receive_fragments": 6784,
    "maximum_endpoint_fragments": 72,
    "transmitted_bytes": 2367488,
    "received_bytes": 2367488,
    "maximum_outgoing_bytes": 22272,
    "maximum_incoming_bytes": 2304,
    "source_staging_max_bytes": 0,
    "destination_staging_max_bytes": 2304,
    "combined_staging_max_bytes": 2304,
    "source_copy_max_bytes": 0,
    "destination_copy_max_bytes": 2304,
    "local_copy_runs": 1537,
    "affine_copy_loops": 1537,
    "copy_cycles_floor": 288,
    "replay_address_ranges_valid": true
  },
  {
    "phase": 10,
    "mode": "both",
    "label": "op Some(0) OperatorInputs: TensorShape([2, 577, 1024]) F16 Amp(Left) -> TensorShape([768]) RowMajor",
    "transfers": 1537,
    "receive_fragments": 1537,
    "maximum_endpoint_fragments": 12,
    "transmitted_bytes": 2367488,
    "received_bytes": 2367488,
    "maximum_outgoing_bytes": 22272,
    "maximum_incoming_bytes": 2304,
    "source_staging_max_bytes": 22272,
    "destination_staging_max_bytes": 2304,
    "combined_staging_max_bytes": 24576,
    "source_copy_max_bytes": 22272,
    "destination_copy_max_bytes": 2304,
    "local_copy_runs": 8321,
    "affine_copy_loops": 3074,
    "copy_cycles_floor": 3072,
    "replay_address_ranges_valid": true
  },
  {
    "phase": 9,
    "mode": "original",
    "label": "op Some(0) OperatorInputs: TensorShape([2, 577, 1024]) F16 Amp(Left) -> TensorShape([1152]) RowMajor",
    "transfers": 6784,
    "receive_fragments": 6784,
    "maximum_endpoint_fragments": 66,
    "transmitted_bytes": 2367488,
    "received_bytes": 2367488,
    "maximum_outgoing_bytes": 20736,
    "maximum_incoming_bytes": 2304,
    "source_staging_max_bytes": 0,
    "destination_staging_max_bytes": 0,
    "combined_staging_max_bytes": 0,
    "source_copy_max_bytes": 0,
    "destination_copy_max_bytes": 0,
    "local_copy_runs": 0,
    "affine_copy_loops": 0,
    "copy_cycles_floor": 0,
    "replay_address_ranges_valid": true
  },
  {
    "phase": 9,
    "mode": "source",
    "label": "op Some(0) OperatorInputs: TensorShape([2, 577, 1024]) F16 Amp(Left) -> TensorShape([1152]) RowMajor",
    "transfers": 1537,
    "receive_fragments": 1537,
    "maximum_endpoint_fragments": 11,
    "transmitted_bytes": 2367488,
    "received_bytes": 2367488,
    "maximum_outgoing_bytes": 20736,
    "maximum_incoming_bytes": 2304,
    "source_staging_max_bytes": 20736,
    "destination_staging_max_bytes": 0,
    "combined_staging_max_bytes": 20736,
    "source_copy_max_bytes": 20736,
    "destination_copy_max_bytes": 0,
    "local_copy_runs": 6784,
    "affine_copy_loops": 1537,
    "copy_cycles_floor": 2592,
    "replay_address_ranges_valid": true
  },
  {
    "phase": 9,
    "mode": "destination",
    "label": "op Some(0) OperatorInputs: TensorShape([2, 577, 1024]) F16 Amp(Left) -> TensorShape([1152]) RowMajor",
    "transfers": 6784,
    "receive_fragments": 6784,
    "maximum_endpoint_fragments": 66,
    "transmitted_bytes": 2367488,
    "received_bytes": 2367488,
    "maximum_outgoing_bytes": 20736,
    "maximum_incoming_bytes": 2304,
    "source_staging_max_bytes": 0,
    "destination_staging_max_bytes": 2304,
    "combined_staging_max_bytes": 2304,
    "source_copy_max_bytes": 0,
    "destination_copy_max_bytes": 2304,
    "local_copy_runs": 1537,
    "affine_copy_loops": 1537,
    "copy_cycles_floor": 288,
    "replay_address_ranges_valid": true
  },
  {
    "phase": 9,
    "mode": "both",
    "label": "op Some(0) OperatorInputs: TensorShape([2, 577, 1024]) F16 Amp(Left) -> TensorShape([1152]) RowMajor",
    "transfers": 1537,
    "receive_fragments": 1537,
    "maximum_endpoint_fragments": 11,
    "transmitted_bytes": 2367488,
    "received_bytes": 2367488,
    "maximum_outgoing_bytes": 20736,
    "maximum_incoming_bytes": 2304,
    "source_staging_max_bytes": 20736,
    "destination_staging_max_bytes": 2304,
    "combined_staging_max_bytes": 23040,
    "source_copy_max_bytes": 20736,
    "destination_copy_max_bytes": 2304,
    "local_copy_runs": 8321,
    "affine_copy_loops": 3074,
    "copy_cycles_floor": 2880,
    "replay_address_ranges_valid": true
  },
  {
    "phase": 94,
    "mode": "original",
    "label": "op Some(39) OperatorInputs: TensorShape([2, 1, 1024]) F16 RowMajor -> TensorShape([2, 1, 1024]) RowMajor",
    "transfers": 127,
    "receive_fragments": 191,
    "maximum_endpoint_fragments": 65,
    "transmitted_bytes": 8128,
    "received_bytes": 12224,
    "maximum_outgoing_bytes": 4032,
    "maximum_incoming_bytes": 192,
    "source_staging_max_bytes": 0,
    "destination_staging_max_bytes": 0,
    "combined_staging_max_bytes": 0,
    "source_copy_max_bytes": 0,
    "destination_copy_max_bytes": 0,
    "local_copy_runs": 0,
    "affine_copy_loops": 0,
    "copy_cycles_floor": 0,
    "replay_address_ranges_valid": true
  },
  {
    "phase": 94,
    "mode": "source",
    "label": "op Some(39) OperatorInputs: TensorShape([2, 1, 1024]) F16 RowMajor -> TensorShape([2, 1, 1024]) RowMajor",
    "transfers": 95,
    "receive_fragments": 127,
    "maximum_endpoint_fragments": 64,
    "transmitted_bytes": 8128,
    "received_bytes": 12224,
    "maximum_outgoing_bytes": 4032,
    "maximum_incoming_bytes": 192,
    "source_staging_max_bytes": 4032,
    "destination_staging_max_bytes": 0,
    "combined_staging_max_bytes": 4032,
    "source_copy_max_bytes": 4032,
    "destination_copy_max_bytes": 0,
    "local_copy_runs": 65,
    "affine_copy_loops": 33,
    "copy_cycles_floor": 504,
    "replay_address_ranges_valid": true
  },
  {
    "phase": 94,
    "mode": "destination",
    "label": "op Some(39) OperatorInputs: TensorShape([2, 1, 1024]) F16 RowMajor -> TensorShape([2, 1, 1024]) RowMajor",
    "transfers": 127,
    "receive_fragments": 191,
    "maximum_endpoint_fragments": 65,
    "transmitted_bytes": 8128,
    "received_bytes": 12224,
    "maximum_outgoing_bytes": 4032,
    "maximum_incoming_bytes": 192,
    "source_staging_max_bytes": 0,
    "destination_staging_max_bytes": 192,
    "combined_staging_max_bytes": 192,
    "source_copy_max_bytes": 0,
    "destination_copy_max_bytes": 192,
    "local_copy_runs": 64,
    "affine_copy_loops": 64,
    "copy_cycles_floor": 24,
    "replay_address_ranges_valid": true
  },
  {
    "phase": 94,
    "mode": "both",
    "label": "op Some(39) OperatorInputs: TensorShape([2, 1, 1024]) F16 RowMajor -> TensorShape([2, 1, 1024]) RowMajor",
    "transfers": 95,
    "receive_fragments": 127,
    "maximum_endpoint_fragments": 64,
    "transmitted_bytes": 8128,
    "received_bytes": 12224,
    "maximum_outgoing_bytes": 4032,
    "maximum_incoming_bytes": 192,
    "source_staging_max_bytes": 4032,
    "destination_staging_max_bytes": 192,
    "combined_staging_max_bytes": 4160,
    "source_copy_max_bytes": 4032,
    "destination_copy_max_bytes": 192,
    "local_copy_runs": 129,
    "affine_copy_loops": 97,
    "copy_cycles_floor": 528,
    "replay_address_ranges_valid": true
  },
  {
    "phase": 96,
    "mode": "original",
    "label": "op Some(39) LayoutRearrangement: TensorShape([2, 1, 1024]) F16 RowMajor -> TensorShape([2, 1, 1024]) RowMajor",
    "transfers": 63,
    "receive_fragments": 63,
    "maximum_endpoint_fragments": 63,
    "transmitted_bytes": 4032,
    "received_bytes": 4032,
    "maximum_outgoing_bytes": 64,
    "maximum_incoming_bytes": 4032,
    "source_staging_max_bytes": 0,
    "destination_staging_max_bytes": 0,
    "combined_staging_max_bytes": 0,
    "source_copy_max_bytes": 0,
    "destination_copy_max_bytes": 0,
    "local_copy_runs": 0,
    "affine_copy_loops": 0,
    "copy_cycles_floor": 0,
    "replay_address_ranges_valid": true
  },
  {
    "phase": 96,
    "mode": "source",
    "label": "op Some(39) LayoutRearrangement: TensorShape([2, 1, 1024]) F16 RowMajor -> TensorShape([2, 1, 1024]) RowMajor",
    "transfers": 63,
    "receive_fragments": 63,
    "maximum_endpoint_fragments": 63,
    "transmitted_bytes": 4032,
    "received_bytes": 4032,
    "maximum_outgoing_bytes": 64,
    "maximum_incoming_bytes": 4032,
    "source_staging_max_bytes": 64,
    "destination_staging_max_bytes": 0,
    "combined_staging_max_bytes": 64,
    "source_copy_max_bytes": 64,
    "destination_copy_max_bytes": 0,
    "local_copy_runs": 63,
    "affine_copy_loops": 63,
    "copy_cycles_floor": 8,
    "replay_address_ranges_valid": true
  },
  {
    "phase": 96,
    "mode": "destination",
    "label": "op Some(39) LayoutRearrangement: TensorShape([2, 1, 1024]) F16 RowMajor -> TensorShape([2, 1, 1024]) RowMajor",
    "transfers": 63,
    "receive_fragments": 63,
    "maximum_endpoint_fragments": 63,
    "transmitted_bytes": 4032,
    "received_bytes": 4032,
    "maximum_outgoing_bytes": 64,
    "maximum_incoming_bytes": 4032,
    "source_staging_max_bytes": 0,
    "destination_staging_max_bytes": 4032,
    "combined_staging_max_bytes": 4032,
    "source_copy_max_bytes": 0,
    "destination_copy_max_bytes": 4032,
    "local_copy_runs": 1,
    "affine_copy_loops": 1,
    "copy_cycles_floor": 504,
    "replay_address_ranges_valid": true
  },
  {
    "phase": 96,
    "mode": "both",
    "label": "op Some(39) LayoutRearrangement: TensorShape([2, 1, 1024]) F16 RowMajor -> TensorShape([2, 1, 1024]) RowMajor",
    "transfers": 63,
    "receive_fragments": 63,
    "maximum_endpoint_fragments": 63,
    "transmitted_bytes": 4032,
    "received_bytes": 4032,
    "maximum_outgoing_bytes": 64,
    "maximum_incoming_bytes": 4032,
    "source_staging_max_bytes": 64,
    "destination_staging_max_bytes": 4032,
    "combined_staging_max_bytes": 4032,
    "source_copy_max_bytes": 64,
    "destination_copy_max_bytes": 4032,
    "local_copy_runs": 64,
    "affine_copy_loops": 64,
    "copy_cycles_floor": 512,
    "replay_address_ranges_valid": true
  },
  {
    "phase": 8,
    "mode": "original",
    "label": "op Some(0) OperatorInputs: TensorShape([2, 577, 1024]) F16 Amp(Left) -> TensorShape([1152]) RowMajor",
    "transfers": 6784,
    "receive_fragments": 6784,
    "maximum_endpoint_fragments": 60,
    "transmitted_bytes": 2367488,
    "received_bytes": 2367488,
    "maximum_outgoing_bytes": 18432,
    "maximum_incoming_bytes": 2304,
    "source_staging_max_bytes": 0,
    "destination_staging_max_bytes": 0,
    "combined_staging_max_bytes": 0,
    "source_copy_max_bytes": 0,
    "destination_copy_max_bytes": 0,
    "local_copy_runs": 0,
    "affine_copy_loops": 0,
    "copy_cycles_floor": 0,
    "replay_address_ranges_valid": true
  },
  {
    "phase": 8,
    "mode": "source",
    "label": "op Some(0) OperatorInputs: TensorShape([2, 577, 1024]) F16 Amp(Left) -> TensorShape([1152]) RowMajor",
    "transfers": 1537,
    "receive_fragments": 1537,
    "maximum_endpoint_fragments": 10,
    "transmitted_bytes": 2367488,
    "received_bytes": 2367488,
    "maximum_outgoing_bytes": 18432,
    "maximum_incoming_bytes": 2304,
    "source_staging_max_bytes": 18432,
    "destination_staging_max_bytes": 0,
    "combined_staging_max_bytes": 18432,
    "source_copy_max_bytes": 18432,
    "destination_copy_max_bytes": 0,
    "local_copy_runs": 6784,
    "affine_copy_loops": 1537,
    "copy_cycles_floor": 2304,
    "replay_address_ranges_valid": true
  },
  {
    "phase": 8,
    "mode": "destination",
    "label": "op Some(0) OperatorInputs: TensorShape([2, 577, 1024]) F16 Amp(Left) -> TensorShape([1152]) RowMajor",
    "transfers": 6784,
    "receive_fragments": 6784,
    "maximum_endpoint_fragments": 60,
    "transmitted_bytes": 2367488,
    "received_bytes": 2367488,
    "maximum_outgoing_bytes": 18432,
    "maximum_incoming_bytes": 2304,
    "source_staging_max_bytes": 0,
    "destination_staging_max_bytes": 2304,
    "combined_staging_max_bytes": 2304,
    "source_copy_max_bytes": 0,
    "destination_copy_max_bytes": 2304,
    "local_copy_runs": 1537,
    "affine_copy_loops": 1537,
    "copy_cycles_floor": 288,
    "replay_address_ranges_valid": true
  },
  {
    "phase": 8,
    "mode": "both",
    "label": "op Some(0) OperatorInputs: TensorShape([2, 577, 1024]) F16 Amp(Left) -> TensorShape([1152]) RowMajor",
    "transfers": 1537,
    "receive_fragments": 1537,
    "maximum_endpoint_fragments": 10,
    "transmitted_bytes": 2367488,
    "received_bytes": 2367488,
    "maximum_outgoing_bytes": 18432,
    "maximum_incoming_bytes": 2304,
    "source_staging_max_bytes": 18432,
    "destination_staging_max_bytes": 2304,
    "combined_staging_max_bytes": 20736,
    "source_copy_max_bytes": 18432,
    "destination_copy_max_bytes": 2304,
    "local_copy_runs": 8321,
    "affine_copy_loops": 3074,
    "copy_cycles_floor": 2592,
    "replay_address_ranges_valid": true
  },
  {
    "phase": 3,
    "mode": "original",
    "label": "op Some(0) OperatorInputs: TensorShape([2, 577, 1024]) F16 Amp(Left) -> TensorShape([1152]) RowMajor",
    "transfers": 6784,
    "receive_fragments": 6784,
    "maximum_endpoint_fragments": 54,
    "transmitted_bytes": 2367488,
    "received_bytes": 2367488,
    "maximum_outgoing_bytes": 17664,
    "maximum_incoming_bytes": 2304,
    "source_staging_max_bytes": 0,
    "destination_staging_max_bytes": 0,
    "combined_staging_max_bytes": 0,
    "source_copy_max_bytes": 0,
    "destination_copy_max_bytes": 0,
    "local_copy_runs": 0,
    "affine_copy_loops": 0,
    "copy_cycles_floor": 0,
    "replay_address_ranges_valid": true
  },
  {
    "phase": 3,
    "mode": "source",
    "label": "op Some(0) OperatorInputs: TensorShape([2, 577, 1024]) F16 Amp(Left) -> TensorShape([1152]) RowMajor",
    "transfers": 1537,
    "receive_fragments": 1537,
    "maximum_endpoint_fragments": 9,
    "transmitted_bytes": 2367488,
    "received_bytes": 2367488,
    "maximum_outgoing_bytes": 17664,
    "maximum_incoming_bytes": 2304,
    "source_staging_max_bytes": 17664,
    "destination_staging_max_bytes": 0,
    "combined_staging_max_bytes": 17664,
    "source_copy_max_bytes": 17664,
    "destination_copy_max_bytes": 0,
    "local_copy_runs": 6784,
    "affine_copy_loops": 1537,
    "copy_cycles_floor": 2208,
    "replay_address_ranges_valid": true
  },
  {
    "phase": 3,
    "mode": "destination",
    "label": "op Some(0) OperatorInputs: TensorShape([2, 577, 1024]) F16 Amp(Left) -> TensorShape([1152]) RowMajor",
    "transfers": 6784,
    "receive_fragments": 6784,
    "maximum_endpoint_fragments": 54,
    "transmitted_bytes": 2367488,
    "received_bytes": 2367488,
    "maximum_outgoing_bytes": 17664,
    "maximum_incoming_bytes": 2304,
    "source_staging_max_bytes": 0,
    "destination_staging_max_bytes": 2304,
    "combined_staging_max_bytes": 2304,
    "source_copy_max_bytes": 0,
    "destination_copy_max_bytes": 2304,
    "local_copy_runs": 1537,
    "affine_copy_loops": 1537,
    "copy_cycles_floor": 288,
    "replay_address_ranges_valid": true
  },
  {
    "phase": 3,
    "mode": "both",
    "label": "op Some(0) OperatorInputs: TensorShape([2, 577, 1024]) F16 Amp(Left) -> TensorShape([1152]) RowMajor",
    "transfers": 1537,
    "receive_fragments": 1537,
    "maximum_endpoint_fragments": 9,
    "transmitted_bytes": 2367488,
    "received_bytes": 2367488,
    "maximum_outgoing_bytes": 17664,
    "maximum_incoming_bytes": 2304,
    "source_staging_max_bytes": 17664,
    "destination_staging_max_bytes": 2304,
    "combined_staging_max_bytes": 19968,
    "source_copy_max_bytes": 17664,
    "destination_copy_max_bytes": 2304,
    "local_copy_runs": 8321,
    "affine_copy_loops": 3074,
    "copy_cycles_floor": 2496,
    "replay_address_ranges_valid": true
  },
  {
    "phase": 7,
    "mode": "original",
    "label": "op Some(0) OperatorInputs: TensorShape([2, 577, 1024]) F16 Amp(Left) -> TensorShape([1152]) RowMajor",
    "transfers": 6784,
    "receive_fragments": 6784,
    "maximum_endpoint_fragments": 54,
    "transmitted_bytes": 2367488,
    "received_bytes": 2367488,
    "maximum_outgoing_bytes": 16128,
    "maximum_incoming_bytes": 2304,
    "source_staging_max_bytes": 0,
    "destination_staging_max_bytes": 0,
    "combined_staging_max_bytes": 0,
    "source_copy_max_bytes": 0,
    "destination_copy_max_bytes": 0,
    "local_copy_runs": 0,
    "affine_copy_loops": 0,
    "copy_cycles_floor": 0,
    "replay_address_ranges_valid": true
  },
  {
    "phase": 7,
    "mode": "source",
    "label": "op Some(0) OperatorInputs: TensorShape([2, 577, 1024]) F16 Amp(Left) -> TensorShape([1152]) RowMajor",
    "transfers": 1537,
    "receive_fragments": 1537,
    "maximum_endpoint_fragments": 9,
    "transmitted_bytes": 2367488,
    "received_bytes": 2367488,
    "maximum_outgoing_bytes": 16128,
    "maximum_incoming_bytes": 2304,
    "source_staging_max_bytes": 16128,
    "destination_staging_max_bytes": 0,
    "combined_staging_max_bytes": 16128,
    "source_copy_max_bytes": 16128,
    "destination_copy_max_bytes": 0,
    "local_copy_runs": 6784,
    "affine_copy_loops": 1537,
    "copy_cycles_floor": 2016,
    "replay_address_ranges_valid": true
  },
  {
    "phase": 7,
    "mode": "destination",
    "label": "op Some(0) OperatorInputs: TensorShape([2, 577, 1024]) F16 Amp(Left) -> TensorShape([1152]) RowMajor",
    "transfers": 6784,
    "receive_fragments": 6784,
    "maximum_endpoint_fragments": 54,
    "transmitted_bytes": 2367488,
    "received_bytes": 2367488,
    "maximum_outgoing_bytes": 16128,
    "maximum_incoming_bytes": 2304,
    "source_staging_max_bytes": 0,
    "destination_staging_max_bytes": 2304,
    "combined_staging_max_bytes": 2304,
    "source_copy_max_bytes": 0,
    "destination_copy_max_bytes": 2304,
    "local_copy_runs": 1537,
    "affine_copy_loops": 1537,
    "copy_cycles_floor": 288,
    "replay_address_ranges_valid": true
  },
  {
    "phase": 7,
    "mode": "both",
    "label": "op Some(0) OperatorInputs: TensorShape([2, 577, 1024]) F16 Amp(Left) -> TensorShape([1152]) RowMajor",
    "transfers": 1537,
    "receive_fragments": 1537,
    "maximum_endpoint_fragments": 9,
    "transmitted_bytes": 2367488,
    "received_bytes": 2367488,
    "maximum_outgoing_bytes": 16128,
    "maximum_incoming_bytes": 2304,
    "source_staging_max_bytes": 16128,
    "destination_staging_max_bytes": 2304,
    "combined_staging_max_bytes": 18432,
    "source_copy_max_bytes": 16128,
    "destination_copy_max_bytes": 2304,
    "local_copy_runs": 8321,
    "affine_copy_loops": 3074,
    "copy_cycles_floor": 2304,
    "replay_address_ranges_valid": true
  },
  {
    "phase": 4,
    "mode": "original",
    "label": "op Some(0) OperatorInputs: TensorShape([2, 577, 1024]) F16 Amp(Left) -> TensorShape([1152]) RowMajor",
    "transfers": 6784,
    "receive_fragments": 6784,
    "maximum_endpoint_fragments": 48,
    "transmitted_bytes": 2367488,
    "received_bytes": 2367488,
    "maximum_outgoing_bytes": 15360,
    "maximum_incoming_bytes": 2304,
    "source_staging_max_bytes": 0,
    "destination_staging_max_bytes": 0,
    "combined_staging_max_bytes": 0,
    "source_copy_max_bytes": 0,
    "destination_copy_max_bytes": 0,
    "local_copy_runs": 0,
    "affine_copy_loops": 0,
    "copy_cycles_floor": 0,
    "replay_address_ranges_valid": true
  },
  {
    "phase": 4,
    "mode": "source",
    "label": "op Some(0) OperatorInputs: TensorShape([2, 577, 1024]) F16 Amp(Left) -> TensorShape([1152]) RowMajor",
    "transfers": 1537,
    "receive_fragments": 1537,
    "maximum_endpoint_fragments": 8,
    "transmitted_bytes": 2367488,
    "received_bytes": 2367488,
    "maximum_outgoing_bytes": 15360,
    "maximum_incoming_bytes": 2304,
    "source_staging_max_bytes": 15360,
    "destination_staging_max_bytes": 0,
    "combined_staging_max_bytes": 15360,
    "source_copy_max_bytes": 15360,
    "destination_copy_max_bytes": 0,
    "local_copy_runs": 6784,
    "affine_copy_loops": 1537,
    "copy_cycles_floor": 1920,
    "replay_address_ranges_valid": true
  },
  {
    "phase": 4,
    "mode": "destination",
    "label": "op Some(0) OperatorInputs: TensorShape([2, 577, 1024]) F16 Amp(Left) -> TensorShape([1152]) RowMajor",
    "transfers": 6784,
    "receive_fragments": 6784,
    "maximum_endpoint_fragments": 48,
    "transmitted_bytes": 2367488,
    "received_bytes": 2367488,
    "maximum_outgoing_bytes": 15360,
    "maximum_incoming_bytes": 2304,
    "source_staging_max_bytes": 0,
    "destination_staging_max_bytes": 2304,
    "combined_staging_max_bytes": 2304,
    "source_copy_max_bytes": 0,
    "destination_copy_max_bytes": 2304,
    "local_copy_runs": 1537,
    "affine_copy_loops": 1537,
    "copy_cycles_floor": 288,
    "replay_address_ranges_valid": true
  },
  {
    "phase": 4,
    "mode": "both",
    "label": "op Some(0) OperatorInputs: TensorShape([2, 577, 1024]) F16 Amp(Left) -> TensorShape([1152]) RowMajor",
    "transfers": 1537,
    "receive_fragments": 1537,
    "maximum_endpoint_fragments": 8,
    "transmitted_bytes": 2367488,
    "received_bytes": 2367488,
    "maximum_outgoing_bytes": 15360,
    "maximum_incoming_bytes": 2304,
    "source_staging_max_bytes": 15360,
    "destination_staging_max_bytes": 2304,
    "combined_staging_max_bytes": 17664,
    "source_copy_max_bytes": 15360,
    "destination_copy_max_bytes": 2304,
    "local_copy_runs": 8321,
    "affine_copy_loops": 3074,
    "copy_cycles_floor": 2208,
    "replay_address_ranges_valid": true
  },
  {
    "phase": 6,
    "mode": "original",
    "label": "op Some(0) OperatorInputs: TensorShape([2, 577, 1024]) F16 Amp(Left) -> TensorShape([1152]) RowMajor",
    "transfers": 6784,
    "receive_fragments": 6784,
    "maximum_endpoint_fragments": 48,
    "transmitted_bytes": 2367488,
    "received_bytes": 2367488,
    "maximum_outgoing_bytes": 13824,
    "maximum_incoming_bytes": 2304,
    "source_staging_max_bytes": 0,
    "destination_staging_max_bytes": 0,
    "combined_staging_max_bytes": 0,
    "source_copy_max_bytes": 0,
    "destination_copy_max_bytes": 0,
    "local_copy_runs": 0,
    "affine_copy_loops": 0,
    "copy_cycles_floor": 0,
    "replay_address_ranges_valid": true
  },
  {
    "phase": 6,
    "mode": "source",
    "label": "op Some(0) OperatorInputs: TensorShape([2, 577, 1024]) F16 Amp(Left) -> TensorShape([1152]) RowMajor",
    "transfers": 1537,
    "receive_fragments": 1537,
    "maximum_endpoint_fragments": 8,
    "transmitted_bytes": 2367488,
    "received_bytes": 2367488,
    "maximum_outgoing_bytes": 13824,
    "maximum_incoming_bytes": 2304,
    "source_staging_max_bytes": 13824,
    "destination_staging_max_bytes": 0,
    "combined_staging_max_bytes": 13824,
    "source_copy_max_bytes": 13824,
    "destination_copy_max_bytes": 0,
    "local_copy_runs": 6784,
    "affine_copy_loops": 1537,
    "copy_cycles_floor": 1728,
    "replay_address_ranges_valid": true
  },
  {
    "phase": 6,
    "mode": "destination",
    "label": "op Some(0) OperatorInputs: TensorShape([2, 577, 1024]) F16 Amp(Left) -> TensorShape([1152]) RowMajor",
    "transfers": 6784,
    "receive_fragments": 6784,
    "maximum_endpoint_fragments": 48,
    "transmitted_bytes": 2367488,
    "received_bytes": 2367488,
    "maximum_outgoing_bytes": 13824,
    "maximum_incoming_bytes": 2304,
    "source_staging_max_bytes": 0,
    "destination_staging_max_bytes": 2304,
    "combined_staging_max_bytes": 2304,
    "source_copy_max_bytes": 0,
    "destination_copy_max_bytes": 2304,
    "local_copy_runs": 1537,
    "affine_copy_loops": 1537,
    "copy_cycles_floor": 288,
    "replay_address_ranges_valid": true
  },
  {
    "phase": 6,
    "mode": "both",
    "label": "op Some(0) OperatorInputs: TensorShape([2, 577, 1024]) F16 Amp(Left) -> TensorShape([1152]) RowMajor",
    "transfers": 1537,
    "receive_fragments": 1537,
    "maximum_endpoint_fragments": 8,
    "transmitted_bytes": 2367488,
    "received_bytes": 2367488,
    "maximum_outgoing_bytes": 13824,
    "maximum_incoming_bytes": 2304,
    "source_staging_max_bytes": 13824,
    "destination_staging_max_bytes": 2304,
    "combined_staging_max_bytes": 16128,
    "source_copy_max_bytes": 13824,
    "destination_copy_max_bytes": 2304,
    "local_copy_runs": 8321,
    "affine_copy_loops": 3074,
    "copy_cycles_floor": 2016,
    "replay_address_ranges_valid": true
  },
  {
    "phase": 37,
    "mode": "original",
    "label": "op Some(12) LayoutRearrangement: TensorShape([32, 577, 64]) F16 RowMajor -> TensorShape([32, 577, 64]) RowMajor",
    "transfers": 18451,
    "receive_fragments": 18451,
    "maximum_endpoint_fragments": 45,
    "transmitted_bytes": 2361728,
    "received_bytes": 2361728,
    "maximum_outgoing_bytes": 1664,
    "maximum_incoming_bytes": 4096,
    "source_staging_max_bytes": 0,
    "destination_staging_max_bytes": 0,
    "combined_staging_max_bytes": 0,
    "source_copy_max_bytes": 0,
    "destination_copy_max_bytes": 0,
    "local_copy_runs": 0,
    "affine_copy_loops": 0,
    "copy_cycles_floor": 0,
    "replay_address_ranges_valid": true
  },
  {
    "phase": 37,
    "mode": "source",
    "label": "op Some(12) LayoutRearrangement: TensorShape([32, 577, 64]) F16 RowMajor -> TensorShape([32, 577, 64]) RowMajor",
    "transfers": 18451,
    "receive_fragments": 18451,
    "maximum_endpoint_fragments": 45,
    "transmitted_bytes": 2361728,
    "received_bytes": 2361728,
    "maximum_outgoing_bytes": 1664,
    "maximum_incoming_bytes": 4096,
    "source_staging_max_bytes": 1664,
    "destination_staging_max_bytes": 0,
    "combined_staging_max_bytes": 1664,
    "source_copy_max_bytes": 1664,
    "destination_copy_max_bytes": 0,
    "local_copy_runs": 1483,
    "affine_copy_loops": 1482,
    "copy_cycles_floor": 208,
    "replay_address_ranges_valid": true
  },
  {
    "phase": 37,
    "mode": "destination",
    "label": "op Some(12) LayoutRearrangement: TensorShape([32, 577, 64]) F16 RowMajor -> TensorShape([32, 577, 64]) RowMajor",
    "transfers": 18451,
    "receive_fragments": 18451,
    "maximum_endpoint_fragments": 45,
    "transmitted_bytes": 2361728,
    "received_bytes": 2361728,
    "maximum_outgoing_bytes": 1664,
    "maximum_incoming_bytes": 4096,
    "source_staging_max_bytes": 0,
    "destination_staging_max_bytes": 4096,
    "combined_staging_max_bytes": 4096,
    "source_copy_max_bytes": 0,
    "destination_copy_max_bytes": 4096,
    "local_copy_runs": 589,
    "affine_copy_loops": 589,
    "copy_cycles_floor": 512,
    "replay_address_ranges_valid": true
  },
  {
    "phase": 37,
    "mode": "both",
    "label": "op Some(12) LayoutRearrangement: TensorShape([32, 577, 64]) F16 RowMajor -> TensorShape([32, 577, 64]) RowMajor",
    "transfers": 18451,
    "receive_fragments": 18451,
    "maximum_endpoint_fragments": 45,
    "transmitted_bytes": 2361728,
    "received_bytes": 2361728,
    "maximum_outgoing_bytes": 1664,
    "maximum_incoming_bytes": 4096,
    "source_staging_max_bytes": 1664,
    "destination_staging_max_bytes": 4096,
    "combined_staging_max_bytes": 5760,
    "source_copy_max_bytes": 1664,
    "destination_copy_max_bytes": 4096,
    "local_copy_runs": 2072,
    "affine_copy_loops": 2071,
    "copy_cycles_floor": 720,
    "replay_address_ranges_valid": true
  },
  {
    "phase": 102,
    "mode": "original",
    "label": "op Some(43) LayoutRearrangement: TensorShape([2, 1, 4096]) F16 RowMajor -> TensorShape([2, 1, 4096]) RowMajor",
    "transfers": 44,
    "receive_fragments": 2814,
    "maximum_endpoint_fragments": 44,
    "transmitted_bytes": 16384,
    "received_bytes": 1047808,
    "maximum_outgoing_bytes": 16384,
    "maximum_incoming_bytes": 768,
    "source_staging_max_bytes": 0,
    "destination_staging_max_bytes": 0,
    "combined_staging_max_bytes": 0,
    "source_copy_max_bytes": 0,
    "destination_copy_max_bytes": 0,
    "local_copy_runs": 0,
    "affine_copy_loops": 0,
    "copy_cycles_floor": 0,
    "replay_address_ranges_valid": true
  },
  {
    "phase": 102,
    "mode": "source",
    "label": "op Some(43) LayoutRearrangement: TensorShape([2, 1, 4096]) F16 RowMajor -> TensorShape([2, 1, 4096]) RowMajor",
    "transfers": 22,
    "receive_fragments": 1407,
    "maximum_endpoint_fragments": 22,
    "transmitted_bytes": 16384,
    "received_bytes": 1047808,
    "maximum_outgoing_bytes": 16384,
    "maximum_incoming_bytes": 768,
    "source_staging_max_bytes": 16384,
    "destination_staging_max_bytes": 0,
    "combined_staging_max_bytes": 16384,
    "source_copy_max_bytes": 16384,
    "destination_copy_max_bytes": 0,
    "local_copy_runs": 44,
    "affine_copy_loops": 22,
    "copy_cycles_floor": 2048,
    "replay_address_ranges_valid": true
  },
  {
    "phase": 102,
    "mode": "destination",
    "label": "op Some(43) LayoutRearrangement: TensorShape([2, 1, 4096]) F16 RowMajor -> TensorShape([2, 1, 4096]) RowMajor",
    "transfers": 44,
    "receive_fragments": 2814,
    "maximum_endpoint_fragments": 44,
    "transmitted_bytes": 16384,
    "received_bytes": 1047808,
    "maximum_outgoing_bytes": 16384,
    "maximum_incoming_bytes": 768,
    "source_staging_max_bytes": 0,
    "destination_staging_max_bytes": 768,
    "combined_staging_max_bytes": 768,
    "source_copy_max_bytes": 0,
    "destination_copy_max_bytes": 768,
    "local_copy_runs": 1407,
    "affine_copy_loops": 1407,
    "copy_cycles_floor": 96,
    "replay_address_ranges_valid": true
  },
  {
    "phase": 102,
    "mode": "both",
    "label": "op Some(43) LayoutRearrangement: TensorShape([2, 1, 4096]) F16 RowMajor -> TensorShape([2, 1, 4096]) RowMajor",
    "transfers": 22,
    "receive_fragments": 1407,
    "maximum_endpoint_fragments": 22,
    "transmitted_bytes": 16384,
    "received_bytes": 1047808,
    "maximum_outgoing_bytes": 16384,
    "maximum_incoming_bytes": 768,
    "source_staging_max_bytes": 16384,
    "destination_staging_max_bytes": 768,
    "combined_staging_max_bytes": 16384,
    "source_copy_max_bytes": 16384,
    "destination_copy_max_bytes": 768,
    "local_copy_runs": 1451,
    "affine_copy_loops": 1429,
    "copy_cycles_floor": 2144,
    "replay_address_ranges_valid": true
  },
  {
    "phase": 5,
    "mode": "original",
    "label": "op Some(0) OperatorInputs: TensorShape([2, 577, 1024]) F16 Amp(Left) -> TensorShape([1152]) RowMajor",
    "transfers": 6784,
    "receive_fragments": 6784,
    "maximum_endpoint_fragments": 42,
    "transmitted_bytes": 2367488,
    "received_bytes": 2367488,
    "maximum_outgoing_bytes": 13056,
    "maximum_incoming_bytes": 2304,
    "source_staging_max_bytes": 0,
    "destination_staging_max_bytes": 0,
    "combined_staging_max_bytes": 0,
    "source_copy_max_bytes": 0,
    "destination_copy_max_bytes": 0,
    "local_copy_runs": 0,
    "affine_copy_loops": 0,
    "copy_cycles_floor": 0,
    "replay_address_ranges_valid": true
  },
  {
    "phase": 5,
    "mode": "source",
    "label": "op Some(0) OperatorInputs: TensorShape([2, 577, 1024]) F16 Amp(Left) -> TensorShape([1152]) RowMajor",
    "transfers": 1537,
    "receive_fragments": 1537,
    "maximum_endpoint_fragments": 7,
    "transmitted_bytes": 2367488,
    "received_bytes": 2367488,
    "maximum_outgoing_bytes": 13056,
    "maximum_incoming_bytes": 2304,
    "source_staging_max_bytes": 13056,
    "destination_staging_max_bytes": 0,
    "combined_staging_max_bytes": 13056,
    "source_copy_max_bytes": 13056,
    "destination_copy_max_bytes": 0,
    "local_copy_runs": 6784,
    "affine_copy_loops": 1537,
    "copy_cycles_floor": 1632,
    "replay_address_ranges_valid": true
  },
  {
    "phase": 5,
    "mode": "destination",
    "label": "op Some(0) OperatorInputs: TensorShape([2, 577, 1024]) F16 Amp(Left) -> TensorShape([1152]) RowMajor",
    "transfers": 6784,
    "receive_fragments": 6784,
    "maximum_endpoint_fragments": 42,
    "transmitted_bytes": 2367488,
    "received_bytes": 2367488,
    "maximum_outgoing_bytes": 13056,
    "maximum_incoming_bytes": 2304,
    "source_staging_max_bytes": 0,
    "destination_staging_max_bytes": 2304,
    "combined_staging_max_bytes": 2304,
    "source_copy_max_bytes": 0,
    "destination_copy_max_bytes": 2304,
    "local_copy_runs": 1537,
    "affine_copy_loops": 1537,
    "copy_cycles_floor": 288,
    "replay_address_ranges_valid": true
  },
  {
    "phase": 5,
    "mode": "both",
    "label": "op Some(0) OperatorInputs: TensorShape([2, 577, 1024]) F16 Amp(Left) -> TensorShape([1152]) RowMajor",
    "transfers": 1537,
    "receive_fragments": 1537,
    "maximum_endpoint_fragments": 7,
    "transmitted_bytes": 2367488,
    "received_bytes": 2367488,
    "maximum_outgoing_bytes": 13056,
    "maximum_incoming_bytes": 2304,
    "source_staging_max_bytes": 13056,
    "destination_staging_max_bytes": 2304,
    "combined_staging_max_bytes": 15360,
    "source_copy_max_bytes": 13056,
    "destination_copy_max_bytes": 2304,
    "local_copy_runs": 8321,
    "affine_copy_loops": 3074,
    "copy_cycles_floor": 1920,
    "replay_address_ranges_valid": true
  },
  {
    "phase": 85,
    "mode": "original",
    "label": "op Some(35) OperatorInputs: TensorShape([16, 1, 577]) F16 Amp(Left) -> TensorShape([16, 1, 577]) Amp(Left)",
    "transfers": 576,
    "receive_fragments": 576,
    "maximum_endpoint_fragments": 36,
    "transmitted_bytes": 18432,
    "received_bytes": 18432,
    "maximum_outgoing_bytes": 32,
    "maximum_incoming_bytes": 1152,
    "source_staging_max_bytes": 0,
    "destination_staging_max_bytes": 0,
    "combined_staging_max_bytes": 0,
    "source_copy_max_bytes": 0,
    "destination_copy_max_bytes": 0,
    "local_copy_runs": 0,
    "affine_copy_loops": 0,
    "copy_cycles_floor": 0,
    "replay_address_ranges_valid": true
  },
  {
    "phase": 85,
    "mode": "source",
    "label": "op Some(35) OperatorInputs: TensorShape([16, 1, 577]) F16 Amp(Left) -> TensorShape([16, 1, 577]) Amp(Left)",
    "transfers": 576,
    "receive_fragments": 576,
    "maximum_endpoint_fragments": 36,
    "transmitted_bytes": 18432,
    "received_bytes": 18432,
    "maximum_outgoing_bytes": 32,
    "maximum_incoming_bytes": 1152,
    "source_staging_max_bytes": 32,
    "destination_staging_max_bytes": 0,
    "combined_staging_max_bytes": 32,
    "source_copy_max_bytes": 32,
    "destination_copy_max_bytes": 0,
    "local_copy_runs": 576,
    "affine_copy_loops": 576,
    "copy_cycles_floor": 4,
    "replay_address_ranges_valid": true
  },
  {
    "phase": 85,
    "mode": "destination",
    "label": "op Some(35) OperatorInputs: TensorShape([16, 1, 577]) F16 Amp(Left) -> TensorShape([16, 1, 577]) Amp(Left)",
    "transfers": 576,
    "receive_fragments": 576,
    "maximum_endpoint_fragments": 36,
    "transmitted_bytes": 18432,
    "received_bytes": 18432,
    "maximum_outgoing_bytes": 32,
    "maximum_incoming_bytes": 1152,
    "source_staging_max_bytes": 0,
    "destination_staging_max_bytes": 1152,
    "combined_staging_max_bytes": 1152,
    "source_copy_max_bytes": 0,
    "destination_copy_max_bytes": 1152,
    "local_copy_runs": 16,
    "affine_copy_loops": 16,
    "copy_cycles_floor": 144,
    "replay_address_ranges_valid": true
  },
  {
    "phase": 85,
    "mode": "both",
    "label": "op Some(35) OperatorInputs: TensorShape([16, 1, 577]) F16 Amp(Left) -> TensorShape([16, 1, 577]) Amp(Left)",
    "transfers": 576,
    "receive_fragments": 576,
    "maximum_endpoint_fragments": 36,
    "transmitted_bytes": 18432,
    "received_bytes": 18432,
    "maximum_outgoing_bytes": 32,
    "maximum_incoming_bytes": 1152,
    "source_staging_max_bytes": 32,
    "destination_staging_max_bytes": 1152,
    "combined_staging_max_bytes": 1152,
    "source_copy_max_bytes": 32,
    "destination_copy_max_bytes": 1152,
    "local_copy_runs": 592,
    "affine_copy_loops": 592,
    "copy_cycles_floor": 148,
    "replay_address_ranges_valid": true
  },
  {
    "phase": 95,
    "mode": "original",
    "label": "op Some(39) OperatorInputs: TensorShape([2, 1, 32, 2]) F32 RowMajor -> TensorShape([2, 1, 32, 2]) RowMajor",
    "transfers": 64,
    "receive_fragments": 2048,
    "maximum_endpoint_fragments": 33,
    "transmitted_bytes": 512,
    "received_bytes": 16384,
    "maximum_outgoing_bytes": 8,
    "maximum_incoming_bytes": 256,
    "source_staging_max_bytes": 0,
    "destination_staging_max_bytes": 0,
    "combined_staging_max_bytes": 0,
    "source_copy_max_bytes": 0,
    "destination_copy_max_bytes": 0,
    "local_copy_runs": 0,
    "affine_copy_loops": 0,
    "copy_cycles_floor": 0,
    "replay_address_ranges_valid": true
  },
  {
    "phase": 95,
    "mode": "source",
    "label": "op Some(39) OperatorInputs: TensorShape([2, 1, 32, 2]) F32 RowMajor -> TensorShape([2, 1, 32, 2]) RowMajor",
    "transfers": 64,
    "receive_fragments": 2048,
    "maximum_endpoint_fragments": 33,
    "transmitted_bytes": 512,
    "received_bytes": 16384,
    "maximum_outgoing_bytes": 8,
    "maximum_incoming_bytes": 256,
    "source_staging_max_bytes": 8,
    "destination_staging_max_bytes": 0,
    "combined_staging_max_bytes": 8,
    "source_copy_max_bytes": 8,
    "destination_copy_max_bytes": 0,
    "local_copy_runs": 64,
    "affine_copy_loops": 64,
    "copy_cycles_floor": 1,
    "replay_address_ranges_valid": true
  },
  {
    "phase": 95,
    "mode": "destination",
    "label": "op Some(39) OperatorInputs: TensorShape([2, 1, 32, 2]) F32 RowMajor -> TensorShape([2, 1, 32, 2]) RowMajor",
    "transfers": 64,
    "receive_fragments": 2048,
    "maximum_endpoint_fragments": 33,
    "transmitted_bytes": 512,
    "received_bytes": 16384,
    "maximum_outgoing_bytes": 8,
    "maximum_incoming_bytes": 256,
    "source_staging_max_bytes": 0,
    "destination_staging_max_bytes": 256,
    "combined_staging_max_bytes": 256,
    "source_copy_max_bytes": 0,
    "destination_copy_max_bytes": 256,
    "local_copy_runs": 64,
    "affine_copy_loops": 64,
    "copy_cycles_floor": 32,
    "replay_address_ranges_valid": true
  },
  {
    "phase": 95,
    "mode": "both",
    "label": "op Some(39) OperatorInputs: TensorShape([2, 1, 32, 2]) F32 RowMajor -> TensorShape([2, 1, 32, 2]) RowMajor",
    "transfers": 64,
    "receive_fragments": 2048,
    "maximum_endpoint_fragments": 33,
    "transmitted_bytes": 512,
    "received_bytes": 16384,
    "maximum_outgoing_bytes": 8,
    "maximum_incoming_bytes": 256,
    "source_staging_max_bytes": 8,
    "destination_staging_max_bytes": 256,
    "combined_staging_max_bytes": 264,
    "source_copy_max_bytes": 8,
    "destination_copy_max_bytes": 256,
    "local_copy_runs": 128,
    "affine_copy_loops": 128,
    "copy_cycles_floor": 33,
    "replay_address_ranges_valid": true
  },
  {
    "phase": 82,
    "mode": "original",
    "label": "op Some(29) OperatorInputs: TensorShape([2, 1, 1024]) F16 Amp(Left) -> TensorShape([2, 1, 1024]) RowMajor",
    "transfers": 767,
    "receive_fragments": 1023,
    "maximum_endpoint_fragments": 32,
    "transmitted_bytes": 6136,
    "received_bytes": 8184,
    "maximum_outgoing_bytes": 256,
    "maximum_incoming_bytes": 16,
    "source_staging_max_bytes": 0,
    "destination_staging_max_bytes": 0,
    "combined_staging_max_bytes": 0,
    "source_copy_max_bytes": 0,
    "destination_copy_max_bytes": 0,
    "local_copy_runs": 0,
    "affine_copy_loops": 0,
    "copy_cycles_floor": 0,
    "replay_address_ranges_valid": true
  },
  {
    "phase": 82,
    "mode": "source",
    "label": "op Some(29) OperatorInputs: TensorShape([2, 1, 1024]) F16 Amp(Left) -> TensorShape([2, 1, 1024]) RowMajor",
    "transfers": 767,
    "receive_fragments": 1023,
    "maximum_endpoint_fragments": 32,
    "transmitted_bytes": 6136,
    "received_bytes": 8184,
    "maximum_outgoing_bytes": 256,
    "maximum_incoming_bytes": 16,
    "source_staging_max_bytes": 256,
    "destination_staging_max_bytes": 0,
    "combined_staging_max_bytes": 256,
    "source_copy_max_bytes": 256,
    "destination_copy_max_bytes": 0,
    "local_copy_runs": 136,
    "affine_copy_loops": 73,
    "copy_cycles_floor": 32,
    "replay_address_ranges_valid": true
  },
  {
    "phase": 82,
    "mode": "destination",
    "label": "op Some(29) OperatorInputs: TensorShape([2, 1, 1024]) F16 Amp(Left) -> TensorShape([2, 1, 1024]) RowMajor",
    "transfers": 767,
    "receive_fragments": 1023,
    "maximum_endpoint_fragments": 32,
    "transmitted_bytes": 6136,
    "received_bytes": 8184,
    "maximum_outgoing_bytes": 256,
    "maximum_incoming_bytes": 16,
    "source_staging_max_bytes": 0,
    "destination_staging_max_bytes": 16,
    "combined_staging_max_bytes": 16,
    "source_copy_max_bytes": 0,
    "destination_copy_max_bytes": 16,
    "local_copy_runs": 512,
    "affine_copy_loops": 512,
    "copy_cycles_floor": 2,
    "replay_address_ranges_valid": true
  },
  {
    "phase": 82,
    "mode": "both",
    "label": "op Some(29) OperatorInputs: TensorShape([2, 1, 1024]) F16 Amp(Left) -> TensorShape([2, 1, 1024]) RowMajor",
    "transfers": 767,
    "receive_fragments": 1023,
    "maximum_endpoint_fragments": 32,
    "transmitted_bytes": 6136,
    "received_bytes": 8184,
    "maximum_outgoing_bytes": 256,
    "maximum_incoming_bytes": 16,
    "source_staging_max_bytes": 256,
    "destination_staging_max_bytes": 16,
    "combined_staging_max_bytes": 256,
    "source_copy_max_bytes": 256,
    "destination_copy_max_bytes": 16,
    "local_copy_runs": 648,
    "affine_copy_loops": 585,
    "copy_cycles_floor": 34,
    "replay_address_ranges_valid": true
  },
  {
    "phase": 89,
    "mode": "original",
    "label": "op Some(37) LayoutRearrangement: TensorShape([2, 1, 1024]) F16 RowMajor -> TensorShape([2, 1, 1024]) RowMajor",
    "transfers": 32,
    "receive_fragments": 2046,
    "maximum_endpoint_fragments": 32,
    "transmitted_bytes": 4096,
    "received_bytes": 261888,
    "maximum_outgoing_bytes": 4096,
    "maximum_incoming_bytes": 256,
    "source_staging_max_bytes": 0,
    "destination_staging_max_bytes": 0,
    "combined_staging_max_bytes": 0,
    "source_copy_max_bytes": 0,
    "destination_copy_max_bytes": 0,
    "local_copy_runs": 0,
    "affine_copy_loops": 0,
    "copy_cycles_floor": 0,
    "replay_address_ranges_valid": true
  },
  {
    "phase": 89,
    "mode": "source",
    "label": "op Some(37) LayoutRearrangement: TensorShape([2, 1, 1024]) F16 RowMajor -> TensorShape([2, 1, 1024]) RowMajor",
    "transfers": 16,
    "receive_fragments": 1023,
    "maximum_endpoint_fragments": 16,
    "transmitted_bytes": 4096,
    "received_bytes": 261888,
    "maximum_outgoing_bytes": 4096,
    "maximum_incoming_bytes": 256,
    "source_staging_max_bytes": 4096,
    "destination_staging_max_bytes": 0,
    "combined_staging_max_bytes": 4096,
    "source_copy_max_bytes": 4096,
    "destination_copy_max_bytes": 0,
    "local_copy_runs": 32,
    "affine_copy_loops": 16,
    "copy_cycles_floor": 512,
    "replay_address_ranges_valid": true
  },
  {
    "phase": 89,
    "mode": "destination",
    "label": "op Some(37) LayoutRearrangement: TensorShape([2, 1, 1024]) F16 RowMajor -> TensorShape([2, 1, 1024]) RowMajor",
    "transfers": 32,
    "receive_fragments": 2046,
    "maximum_endpoint_fragments": 32,
    "transmitted_bytes": 4096,
    "received_bytes": 261888,
    "maximum_outgoing_bytes": 4096,
    "maximum_incoming_bytes": 256,
    "source_staging_max_bytes": 0,
    "destination_staging_max_bytes": 256,
    "combined_staging_max_bytes": 256,
    "source_copy_max_bytes": 0,
    "destination_copy_max_bytes": 256,
    "local_copy_runs": 1023,
    "affine_copy_loops": 1023,
    "copy_cycles_floor": 32,
    "replay_address_ranges_valid": true
  },
  {
    "phase": 89,
    "mode": "both",
    "label": "op Some(37) LayoutRearrangement: TensorShape([2, 1, 1024]) F16 RowMajor -> TensorShape([2, 1, 1024]) RowMajor",
    "transfers": 16,
    "receive_fragments": 1023,
    "maximum_endpoint_fragments": 16,
    "transmitted_bytes": 4096,
    "received_bytes": 261888,
    "maximum_outgoing_bytes": 4096,
    "maximum_incoming_bytes": 256,
    "source_staging_max_bytes": 4096,
    "destination_staging_max_bytes": 256,
    "combined_staging_max_bytes": 4096,
    "source_copy_max_bytes": 4096,
    "destination_copy_max_bytes": 256,
    "local_copy_runs": 1055,
    "affine_copy_loops": 1039,
    "copy_cycles_floor": 544,
    "replay_address_ranges_valid": true
  },
  {
    "phase": 92,
    "mode": "original",
    "label": "op Some(38) OperatorInputs: TensorShape([2, 1, 1024]) F16 Amp(Left) -> TensorShape([2, 1, 1024]) RowMajor",
    "transfers": 767,
    "receive_fragments": 1023,
    "maximum_endpoint_fragments": 32,
    "transmitted_bytes": 6136,
    "received_bytes": 8184,
    "maximum_outgoing_bytes": 256,
    "maximum_incoming_bytes": 16,
    "source_staging_max_bytes": 0,
    "destination_staging_max_bytes": 0,
    "combined_staging_max_bytes": 0,
    "source_copy_max_bytes": 0,
    "destination_copy_max_bytes": 0,
    "local_copy_runs": 0,
    "affine_copy_loops": 0,
    "copy_cycles_floor": 0,
    "replay_address_ranges_valid": true
  },
  {
    "phase": 92,
    "mode": "source",
    "label": "op Some(38) OperatorInputs: TensorShape([2, 1, 1024]) F16 Amp(Left) -> TensorShape([2, 1, 1024]) RowMajor",
    "transfers": 767,
    "receive_fragments": 1023,
    "maximum_endpoint_fragments": 32,
    "transmitted_bytes": 6136,
    "received_bytes": 8184,
    "maximum_outgoing_bytes": 256,
    "maximum_incoming_bytes": 16,
    "source_staging_max_bytes": 256,
    "destination_staging_max_bytes": 0,
    "combined_staging_max_bytes": 256,
    "source_copy_max_bytes": 256,
    "destination_copy_max_bytes": 0,
    "local_copy_runs": 136,
    "affine_copy_loops": 73,
    "copy_cycles_floor": 32,
    "replay_address_ranges_valid": true
  },
  {
    "phase": 92,
    "mode": "destination",
    "label": "op Some(38) OperatorInputs: TensorShape([2, 1, 1024]) F16 Amp(Left) -> TensorShape([2, 1, 1024]) RowMajor",
    "transfers": 767,
    "receive_fragments": 1023,
    "maximum_endpoint_fragments": 32,
    "transmitted_bytes": 6136,
    "received_bytes": 8184,
    "maximum_outgoing_bytes": 256,
    "maximum_incoming_bytes": 16,
    "source_staging_max_bytes": 0,
    "destination_staging_max_bytes": 16,
    "combined_staging_max_bytes": 16,
    "source_copy_max_bytes": 0,
    "destination_copy_max_bytes": 16,
    "local_copy_runs": 512,
    "affine_copy_loops": 512,
    "copy_cycles_floor": 2,
    "replay_address_ranges_valid": true
  },
  {
    "phase": 92,
    "mode": "both",
    "label": "op Some(38) OperatorInputs: TensorShape([2, 1, 1024]) F16 Amp(Left) -> TensorShape([2, 1, 1024]) RowMajor",
    "transfers": 767,
    "receive_fragments": 1023,
    "maximum_endpoint_fragments": 32,
    "transmitted_bytes": 6136,
    "received_bytes": 8184,
    "maximum_outgoing_bytes": 256,
    "maximum_incoming_bytes": 16,
    "source_staging_max_bytes": 256,
    "destination_staging_max_bytes": 16,
    "combined_staging_max_bytes": 256,
    "source_copy_max_bytes": 256,
    "destination_copy_max_bytes": 16,
    "local_copy_runs": 648,
    "affine_copy_loops": 585,
    "copy_cycles_floor": 34,
    "replay_address_ranges_valid": true
  },
  {
    "phase": 105,
    "mode": "original",
    "label": "op Some(44) OperatorInputs: TensorShape([2, 1, 1024]) F16 Amp(Left) -> TensorShape([2, 1, 1024]) RowMajor",
    "transfers": 767,
    "receive_fragments": 1023,
    "maximum_endpoint_fragments": 32,
    "transmitted_bytes": 6136,
    "received_bytes": 8184,
    "maximum_outgoing_bytes": 256,
    "maximum_incoming_bytes": 16,
    "source_staging_max_bytes": 0,
    "destination_staging_max_bytes": 0,
    "combined_staging_max_bytes": 0,
    "source_copy_max_bytes": 0,
    "destination_copy_max_bytes": 0,
    "local_copy_runs": 0,
    "affine_copy_loops": 0,
    "copy_cycles_floor": 0,
    "replay_address_ranges_valid": true
  },
  {
    "phase": 105,
    "mode": "source",
    "label": "op Some(44) OperatorInputs: TensorShape([2, 1, 1024]) F16 Amp(Left) -> TensorShape([2, 1, 1024]) RowMajor",
    "transfers": 767,
    "receive_fragments": 1023,
    "maximum_endpoint_fragments": 32,
    "transmitted_bytes": 6136,
    "received_bytes": 8184,
    "maximum_outgoing_bytes": 256,
    "maximum_incoming_bytes": 16,
    "source_staging_max_bytes": 256,
    "destination_staging_max_bytes": 0,
    "combined_staging_max_bytes": 256,
    "source_copy_max_bytes": 256,
    "destination_copy_max_bytes": 0,
    "local_copy_runs": 136,
    "affine_copy_loops": 73,
    "copy_cycles_floor": 32,
    "replay_address_ranges_valid": true
  },
  {
    "phase": 105,
    "mode": "destination",
    "label": "op Some(44) OperatorInputs: TensorShape([2, 1, 1024]) F16 Amp(Left) -> TensorShape([2, 1, 1024]) RowMajor",
    "transfers": 767,
    "receive_fragments": 1023,
    "maximum_endpoint_fragments": 32,
    "transmitted_bytes": 6136,
    "received_bytes": 8184,
    "maximum_outgoing_bytes": 256,
    "maximum_incoming_bytes": 16,
    "source_staging_max_bytes": 0,
    "destination_staging_max_bytes": 16,
    "combined_staging_max_bytes": 16,
    "source_copy_max_bytes": 0,
    "destination_copy_max_bytes": 16,
    "local_copy_runs": 512,
    "affine_copy_loops": 512,
    "copy_cycles_floor": 2,
    "replay_address_ranges_valid": true
  },
  {
    "phase": 105,
    "mode": "both",
    "label": "op Some(44) OperatorInputs: TensorShape([2, 1, 1024]) F16 Amp(Left) -> TensorShape([2, 1, 1024]) RowMajor",
    "transfers": 767,
    "receive_fragments": 1023,
    "maximum_endpoint_fragments": 32,
    "transmitted_bytes": 6136,
    "received_bytes": 8184,
    "maximum_outgoing_bytes": 256,
    "maximum_incoming_bytes": 16,
    "source_staging_max_bytes": 256,
    "destination_staging_max_bytes": 16,
    "combined_staging_max_bytes": 256,
    "source_copy_max_bytes": 256,
    "destination_copy_max_bytes": 16,
    "local_copy_runs": 648,
    "affine_copy_loops": 585,
    "copy_cycles_floor": 34,
    "replay_address_ranges_valid": true
  },
  {
    "phase": 107,
    "mode": "original",
    "label": "op Some(46) LayoutRearrangement: TensorShape([2, 1, 1024]) F16 RowMajor -> TensorShape([2, 1, 1024]) RowMajor",
    "transfers": 32,
    "receive_fragments": 2046,
    "maximum_endpoint_fragments": 32,
    "transmitted_bytes": 4096,
    "received_bytes": 261888,
    "maximum_outgoing_bytes": 4096,
    "maximum_incoming_bytes": 256,
    "source_staging_max_bytes": 0,
    "destination_staging_max_bytes": 0,
    "combined_staging_max_bytes": 0,
    "source_copy_max_bytes": 0,
    "destination_copy_max_bytes": 0,
    "local_copy_runs": 0,
    "affine_copy_loops": 0,
    "copy_cycles_floor": 0,
    "replay_address_ranges_valid": true
  },
  {
    "phase": 107,
    "mode": "source",
    "label": "op Some(46) LayoutRearrangement: TensorShape([2, 1, 1024]) F16 RowMajor -> TensorShape([2, 1, 1024]) RowMajor",
    "transfers": 16,
    "receive_fragments": 1023,
    "maximum_endpoint_fragments": 16,
    "transmitted_bytes": 4096,
    "received_bytes": 261888,
    "maximum_outgoing_bytes": 4096,
    "maximum_incoming_bytes": 256,
    "source_staging_max_bytes": 4096,
    "destination_staging_max_bytes": 0,
    "combined_staging_max_bytes": 4096,
    "source_copy_max_bytes": 4096,
    "destination_copy_max_bytes": 0,
    "local_copy_runs": 32,
    "affine_copy_loops": 16,
    "copy_cycles_floor": 512,
    "replay_address_ranges_valid": true
  },
  {
    "phase": 107,
    "mode": "destination",
    "label": "op Some(46) LayoutRearrangement: TensorShape([2, 1, 1024]) F16 RowMajor -> TensorShape([2, 1, 1024]) RowMajor",
    "transfers": 32,
    "receive_fragments": 2046,
    "maximum_endpoint_fragments": 32,
    "transmitted_bytes": 4096,
    "received_bytes": 261888,
    "maximum_outgoing_bytes": 4096,
    "maximum_incoming_bytes": 256,
    "source_staging_max_bytes": 0,
    "destination_staging_max_bytes": 256,
    "combined_staging_max_bytes": 256,
    "source_copy_max_bytes": 0,
    "destination_copy_max_bytes": 256,
    "local_copy_runs": 1023,
    "affine_copy_loops": 1023,
    "copy_cycles_floor": 32,
    "replay_address_ranges_valid": true
  },
  {
    "phase": 107,
    "mode": "both",
    "label": "op Some(46) LayoutRearrangement: TensorShape([2, 1, 1024]) F16 RowMajor -> TensorShape([2, 1, 1024]) RowMajor",
    "transfers": 16,
    "receive_fragments": 1023,
    "maximum_endpoint_fragments": 16,
    "transmitted_bytes": 4096,
    "received_bytes": 261888,
    "maximum_outgoing_bytes": 4096,
    "maximum_incoming_bytes": 256,
    "source_staging_max_bytes": 4096,
    "destination_staging_max_bytes": 256,
    "combined_staging_max_bytes": 4096,
    "source_copy_max_bytes": 4096,
    "destination_copy_max_bytes": 256,
    "local_copy_runs": 1055,
    "affine_copy_loops": 1039,
    "copy_cycles_floor": 544,
    "replay_address_ranges_valid": true
  },
  {
    "phase": 18,
    "mode": "original",
    "label": "op Some(4) OperatorInputs: TensorShape([1024, 3072]) F8F143 { scale_exponent: -4 } Amp(TransposedLeft) -> TensorShape([1024, 3072]) BlockMajor(Matrix { row_block: 192, column_block: 16 })",
    "transfers": 2398,
    "receive_fragments": 6142,
    "maximum_endpoint_fragments": 29,
    "transmitted_bytes": 6681088,
    "received_bytes": 52687360,
    "maximum_outgoing_bytes": 39424,
    "maximum_incoming_bytes": 36864,
    "source_staging_max_bytes": 0,
    "destination_staging_max_bytes": 0,
    "combined_staging_max_bytes": 0,
    "source_copy_max_bytes": 0,
    "destination_copy_max_bytes": 0,
    "local_copy_runs": 0,
    "affine_copy_loops": 0,
    "copy_cycles_floor": 0,
    "replay_address_ranges_valid": true
  },
  {
    "phase": 18,
    "mode": "source",
    "label": "op Some(4) OperatorInputs: TensorShape([1024, 3072]) F8F143 { scale_exponent: -4 } Amp(TransposedLeft) -> TensorShape([1024, 3072]) BlockMajor(Matrix { row_block: 192, column_block: 16 })",
    "transfers": 2206,
    "receive_fragments": 3454,
    "maximum_endpoint_fragments": 27,
    "transmitted_bytes": 6681088,
    "received_bytes": 52687360,
    "maximum_outgoing_bytes": 39424,
    "maximum_incoming_bytes": 36864,
    "source_staging_max_bytes": 39424,
    "destination_staging_max_bytes": 0,
    "combined_staging_max_bytes": 39424,
    "source_copy_max_bytes": 39424,
    "destination_copy_max_bytes": 0,
    "local_copy_runs": 1439,
    "affine_copy_loops": 1439,
    "copy_cycles_floor": 4928,
    "replay_address_ranges_valid": true
  },
  {
    "phase": 18,
    "mode": "destination",
    "label": "op Some(4) OperatorInputs: TensorShape([1024, 3072]) F8F143 { scale_exponent: -4 } Amp(TransposedLeft) -> TensorShape([1024, 3072]) BlockMajor(Matrix { row_block: 192, column_block: 16 })",
    "transfers": 2206,
    "receive_fragments": 3454,
    "maximum_endpoint_fragments": 27,
    "transmitted_bytes": 6681088,
    "received_bytes": 52687360,
    "maximum_outgoing_bytes": 39424,
    "maximum_incoming_bytes": 36864,
    "source_staging_max_bytes": 0,
    "destination_staging_max_bytes": 36864,
    "combined_staging_max_bytes": 36864,
    "source_copy_max_bytes": 0,
    "destination_copy_max_bytes": 36864,
    "local_copy_runs": 1617,
    "affine_copy_loops": 1441,
    "copy_cycles_floor": 4608,
    "replay_address_ranges_valid": true
  },
  {
    "phase": 18,
    "mode": "both",
    "label": "op Some(4) OperatorInputs: TensorShape([1024, 3072]) F8F143 { scale_exponent: -4 } Amp(TransposedLeft) -> TensorShape([1024, 3072]) BlockMajor(Matrix { row_block: 192, column_block: 16 })",
    "transfers": 2206,
    "receive_fragments": 3454,
    "maximum_endpoint_fragments": 27,
    "transmitted_bytes": 6681088,
    "received_bytes": 52687360,
    "maximum_outgoing_bytes": 39424,
    "maximum_incoming_bytes": 36864,
    "source_staging_max_bytes": 39424,
    "destination_staging_max_bytes": 36864,
    "combined_staging_max_bytes": 76288,
    "source_copy_max_bytes": 39424,
    "destination_copy_max_bytes": 36864,
    "local_copy_runs": 3056,
    "affine_copy_loops": 2880,
    "copy_cycles_floor": 9536,
    "replay_address_ranges_valid": true
  },
  {
    "phase": 27,
    "mode": "original",
    "label": "op Some(12) OperatorInputs: TensorShape([32, 577, 64]) F16 Amp(TransposedRight) -> TensorShape([32, 64, 64]) Amp(TransposedRight)",
    "transfers": 640,
    "receive_fragments": 29440,
    "maximum_endpoint_fragments": 25,
    "transmitted_bytes": 524288,
    "received_bytes": 24117248,
    "maximum_outgoing_bytes": 4096,
    "maximum_incoming_bytes": 16384,
    "source_staging_max_bytes": 0,
    "destination_staging_max_bytes": 0,
    "combined_staging_max_bytes": 0,
    "source_copy_max_bytes": 0,
    "destination_copy_max_bytes": 0,
    "local_copy_runs": 0,
    "affine_copy_loops": 0,
    "copy_cycles_floor": 0,
    "replay_address_ranges_valid": true
  },
  {
    "phase": 27,
    "mode": "source",
    "label": "op Some(12) OperatorInputs: TensorShape([32, 577, 64]) F16 Amp(TransposedRight) -> TensorShape([32, 64, 64]) Amp(TransposedRight)",
    "transfers": 640,
    "receive_fragments": 29440,
    "maximum_endpoint_fragments": 25,
    "transmitted_bytes": 524288,
    "received_bytes": 24117248,
    "maximum_outgoing_bytes": 4096,
    "maximum_incoming_bytes": 16384,
    "source_staging_max_bytes": 4096,
    "destination_staging_max_bytes": 0,
    "combined_staging_max_bytes": 4096,
    "source_copy_max_bytes": 4096,
    "destination_copy_max_bytes": 0,
    "local_copy_runs": 256,
    "affine_copy_loops": 128,
    "copy_cycles_floor": 512,
    "replay_address_ranges_valid": true
  },
  {
    "phase": 27,
    "mode": "destination",
    "label": "op Some(12) OperatorInputs: TensorShape([32, 577, 64]) F16 Amp(TransposedRight) -> TensorShape([32, 64, 64]) Amp(TransposedRight)",
    "transfers": 256,
    "receive_fragments": 11776,
    "maximum_endpoint_fragments": 10,
    "transmitted_bytes": 524288,
    "received_bytes": 24117248,
    "maximum_outgoing_bytes": 4096,
    "maximum_incoming_bytes": 16384,
    "source_staging_max_bytes": 0,
    "destination_staging_max_bytes": 16384,
    "combined_staging_max_bytes": 16384,
    "source_copy_max_bytes": 0,
    "destination_copy_max_bytes": 16384,
    "local_copy_runs": 29440,
    "affine_copy_loops": 11776,
    "copy_cycles_floor": 2048,
    "replay_address_ranges_valid": true
  },
  {
    "phase": 27,
    "mode": "both",
    "label": "op Some(12) OperatorInputs: TensorShape([32, 577, 64]) F16 Amp(TransposedRight) -> TensorShape([32, 64, 64]) Amp(TransposedRight)",
    "transfers": 128,
    "receive_fragments": 5888,
    "maximum_endpoint_fragments": 5,
    "transmitted_bytes": 524288,
    "received_bytes": 24117248,
    "maximum_outgoing_bytes": 4096,
    "maximum_incoming_bytes": 16384,
    "source_staging_max_bytes": 4096,
    "destination_staging_max_bytes": 16384,
    "combined_staging_max_bytes": 20480,
    "source_copy_max_bytes": 4096,
    "destination_copy_max_bytes": 16384,
    "local_copy_runs": 29696,
    "affine_copy_loops": 11904,
    "copy_cycles_floor": 2560,
    "replay_address_ranges_valid": true
  },
  {
    "phase": 28,
    "mode": "original",
    "label": "op Some(12) OperatorInputs: TensorShape([32, 577, 64]) F16 Amp(TransposedRight) -> TensorShape([32, 64, 64]) Amp(TransposedRight)",
    "transfers": 640,
    "receive_fragments": 29440,
    "maximum_endpoint_fragments": 25,
    "transmitted_bytes": 524288,
    "received_bytes": 24117248,
    "maximum_outgoing_bytes": 4096,
    "maximum_incoming_bytes": 16384,
    "source_staging_max_bytes": 0,
    "destination_staging_max_bytes": 0,
    "combined_staging_max_bytes": 0,
    "source_copy_max_bytes": 0,
    "destination_copy_max_bytes": 0,
    "local_copy_runs": 0,
    "affine_copy_loops": 0,
    "copy_cycles_floor": 0,
    "replay_address_ranges_valid": true
  },
  {
    "phase": 28,
    "mode": "source",
    "label": "op Some(12) OperatorInputs: TensorShape([32, 577, 64]) F16 Amp(TransposedRight) -> TensorShape([32, 64, 64]) Amp(TransposedRight)",
    "transfers": 640,
    "receive_fragments": 29440,
    "maximum_endpoint_fragments": 25,
    "transmitted_bytes": 524288,
    "received_bytes": 24117248,
    "maximum_outgoing_bytes": 4096,
    "maximum_incoming_bytes": 16384,
    "source_staging_max_bytes": 4096,
    "destination_staging_max_bytes": 0,
    "combined_staging_max_bytes": 4096,
    "source_copy_max_bytes": 4096,
    "destination_copy_max_bytes": 0,
    "local_copy_runs": 256,
    "affine_copy_loops": 128,
    "copy_cycles_floor": 512,
    "replay_address_ranges_valid": true
  },
  {
    "phase": 28,
    "mode": "destination",
    "label": "op Some(12) OperatorInputs: TensorShape([32, 577, 64]) F16 Amp(TransposedRight) -> TensorShape([32, 64, 64]) Amp(TransposedRight)",
    "transfers": 256,
    "receive_fragments": 11776,
    "maximum_endpoint_fragments": 10,
    "transmitted_bytes": 524288,
    "received_bytes": 24117248,
    "maximum_outgoing_bytes": 4096,
    "maximum_incoming_bytes": 16384,
    "source_staging_max_bytes": 0,
    "destination_staging_max_bytes": 16384,
    "combined_staging_max_bytes": 16384,
    "source_copy_max_bytes": 0,
    "destination_copy_max_bytes": 16384,
    "local_copy_runs": 29440,
    "affine_copy_loops": 11776,
    "copy_cycles_floor": 2048,
    "replay_address_ranges_valid": true
  },
  {
    "phase": 28,
    "mode": "both",
    "label": "op Some(12) OperatorInputs: TensorShape([32, 577, 64]) F16 Amp(TransposedRight) -> TensorShape([32, 64, 64]) Amp(TransposedRight)",
    "transfers": 128,
    "receive_fragments": 5888,
    "maximum_endpoint_fragments": 5,
    "transmitted_bytes": 524288,
    "received_bytes": 24117248,
    "maximum_outgoing_bytes": 4096,
    "maximum_incoming_bytes": 16384,
    "source_staging_max_bytes": 4096,
    "destination_staging_max_bytes": 16384,
    "combined_staging_max_bytes": 20480,
    "source_copy_max_bytes": 4096,
    "destination_copy_max_bytes": 16384,
    "local_copy_runs": 29696,
    "affine_copy_loops": 11904,
    "copy_cycles_floor": 2560,
    "replay_address_ranges_valid": true
  },
  {
    "phase": 29,
    "mode": "original",
    "label": "op Some(12) OperatorInputs: TensorShape([32, 577, 64]) F16 Amp(TransposedRight) -> TensorShape([32, 64, 64]) Amp(TransposedRight)",
    "transfers": 640,
    "receive_fragments": 29440,
    "maximum_endpoint_fragments": 25,
    "transmitted_bytes": 524288,
    "received_bytes": 24117248,
    "maximum_outgoing_bytes": 4096,
    "maximum_incoming_bytes": 16384,
    "source_staging_max_bytes": 0,
    "destination_staging_max_bytes": 0,
    "combined_staging_max_bytes": 0,
    "source_copy_max_bytes": 0,
    "destination_copy_max_bytes": 0,
    "local_copy_runs": 0,
    "affine_copy_loops": 0,
    "copy_cycles_floor": 0,
    "replay_address_ranges_valid": true
  },
  {
    "phase": 29,
    "mode": "source",
    "label": "op Some(12) OperatorInputs: TensorShape([32, 577, 64]) F16 Amp(TransposedRight) -> TensorShape([32, 64, 64]) Amp(TransposedRight)",
    "transfers": 640,
    "receive_fragments": 29440,
    "maximum_endpoint_fragments": 25,
    "transmitted_bytes": 524288,
    "received_bytes": 24117248,
    "maximum_outgoing_bytes": 4096,
    "maximum_incoming_bytes": 16384,
    "source_staging_max_bytes": 4096,
    "destination_staging_max_bytes": 0,
    "combined_staging_max_bytes": 4096,
    "source_copy_max_bytes": 4096,
    "destination_copy_max_bytes": 0,
    "local_copy_runs": 256,
    "affine_copy_loops": 128,
    "copy_cycles_floor": 512,
    "replay_address_ranges_valid": true
  },
  {
    "phase": 29,
    "mode": "destination",
    "label": "op Some(12) OperatorInputs: TensorShape([32, 577, 64]) F16 Amp(TransposedRight) -> TensorShape([32, 64, 64]) Amp(TransposedRight)",
    "transfers": 256,
    "receive_fragments": 11776,
    "maximum_endpoint_fragments": 10,
    "transmitted_bytes": 524288,
    "received_bytes": 24117248,
    "maximum_outgoing_bytes": 4096,
    "maximum_incoming_bytes": 16384,
    "source_staging_max_bytes": 0,
    "destination_staging_max_bytes": 16384,
    "combined_staging_max_bytes": 16384,
    "source_copy_max_bytes": 0,
    "destination_copy_max_bytes": 16384,
    "local_copy_runs": 29440,
    "affine_copy_loops": 11776,
    "copy_cycles_floor": 2048,
    "replay_address_ranges_valid": true
  },
  {
    "phase": 29,
    "mode": "both",
    "label": "op Some(12) OperatorInputs: TensorShape([32, 577, 64]) F16 Amp(TransposedRight) -> TensorShape([32, 64, 64]) Amp(TransposedRight)",
    "transfers": 128,
    "receive_fragments": 5888,
    "maximum_endpoint_fragments": 5,
    "transmitted_bytes": 524288,
    "received_bytes": 24117248,
    "maximum_outgoing_bytes": 4096,
    "maximum_incoming_bytes": 16384,
    "source_staging_max_bytes": 4096,
    "destination_staging_max_bytes": 16384,
    "combined_staging_max_bytes": 20480,
    "source_copy_max_bytes": 4096,
    "destination_copy_max_bytes": 16384,
    "local_copy_runs": 29696,
    "affine_copy_loops": 11904,
    "copy_cycles_floor": 2560,
    "replay_address_ranges_valid": true
  },
  {
    "phase": 30,
    "mode": "original",
    "label": "op Some(12) OperatorInputs: TensorShape([32, 577, 64]) F16 Amp(TransposedRight) -> TensorShape([32, 64, 64]) Amp(TransposedRight)",
    "transfers": 640,
    "receive_fragments": 29440,
    "maximum_endpoint_fragments": 25,
    "transmitted_bytes": 524288,
    "received_bytes": 24117248,
    "maximum_outgoing_bytes": 4096,
    "maximum_incoming_bytes": 16384,
    "source_staging_max_bytes": 0,
    "destination_staging_max_bytes": 0,
    "combined_staging_max_bytes": 0,
    "source_copy_max_bytes": 0,
    "destination_copy_max_bytes": 0,
    "local_copy_runs": 0,
    "affine_copy_loops": 0,
    "copy_cycles_floor": 0,
    "replay_address_ranges_valid": true
  },
  {
    "phase": 30,
    "mode": "source",
    "label": "op Some(12) OperatorInputs: TensorShape([32, 577, 64]) F16 Amp(TransposedRight) -> TensorShape([32, 64, 64]) Amp(TransposedRight)",
    "transfers": 640,
    "receive_fragments": 29440,
    "maximum_endpoint_fragments": 25,
    "transmitted_bytes": 524288,
    "received_bytes": 24117248,
    "maximum_outgoing_bytes": 4096,
    "maximum_incoming_bytes": 16384,
    "source_staging_max_bytes": 4096,
    "destination_staging_max_bytes": 0,
    "combined_staging_max_bytes": 4096,
    "source_copy_max_bytes": 4096,
    "destination_copy_max_bytes": 0,
    "local_copy_runs": 256,
    "affine_copy_loops": 128,
    "copy_cycles_floor": 512,
    "replay_address_ranges_valid": true
  },
  {
    "phase": 30,
    "mode": "destination",
    "label": "op Some(12) OperatorInputs: TensorShape([32, 577, 64]) F16 Amp(TransposedRight) -> TensorShape([32, 64, 64]) Amp(TransposedRight)",
    "transfers": 256,
    "receive_fragments": 11776,
    "maximum_endpoint_fragments": 10,
    "transmitted_bytes": 524288,
    "received_bytes": 24117248,
    "maximum_outgoing_bytes": 4096,
    "maximum_incoming_bytes": 16384,
    "source_staging_max_bytes": 0,
    "destination_staging_max_bytes": 16384,
    "combined_staging_max_bytes": 16384,
    "source_copy_max_bytes": 0,
    "destination_copy_max_bytes": 16384,
    "local_copy_runs": 29440,
    "affine_copy_loops": 11776,
    "copy_cycles_floor": 2048,
    "replay_address_ranges_valid": true
  },
  {
    "phase": 30,
    "mode": "both",
    "label": "op Some(12) OperatorInputs: TensorShape([32, 577, 64]) F16 Amp(TransposedRight) -> TensorShape([32, 64, 64]) Amp(TransposedRight)",
    "transfers": 128,
    "receive_fragments": 5888,
    "maximum_endpoint_fragments": 5,
    "transmitted_bytes": 524288,
    "received_bytes": 24117248,
    "maximum_outgoing_bytes": 4096,
    "maximum_incoming_bytes": 16384,
    "source_staging_max_bytes": 4096,
    "destination_staging_max_bytes": 16384,
    "combined_staging_max_bytes": 20480,
    "source_copy_max_bytes": 4096,
    "destination_copy_max_bytes": 16384,
    "local_copy_runs": 29696,
    "affine_copy_loops": 11904,
    "copy_cycles_floor": 2560,
    "replay_address_ranges_valid": true
  },
  {
    "phase": 31,
    "mode": "original",
    "label": "op Some(12) OperatorInputs: TensorShape([32, 577, 64]) F16 Amp(TransposedRight) -> TensorShape([32, 64, 64]) Amp(TransposedRight)",
    "transfers": 640,
    "receive_fragments": 29440,
    "maximum_endpoint_fragments": 25,
    "transmitted_bytes": 524288,
    "received_bytes": 24117248,
    "maximum_outgoing_bytes": 4096,
    "maximum_incoming_bytes": 16384,
    "source_staging_max_bytes": 0,
    "destination_staging_max_bytes": 0,
    "combined_staging_max_bytes": 0,
    "source_copy_max_bytes": 0,
    "destination_copy_max_bytes": 0,
    "local_copy_runs": 0,
    "affine_copy_loops": 0,
    "copy_cycles_floor": 0,
    "replay_address_ranges_valid": true
  },
  {
    "phase": 31,
    "mode": "source",
    "label": "op Some(12) OperatorInputs: TensorShape([32, 577, 64]) F16 Amp(TransposedRight) -> TensorShape([32, 64, 64]) Amp(TransposedRight)",
    "transfers": 640,
    "receive_fragments": 29440,
    "maximum_endpoint_fragments": 25,
    "transmitted_bytes": 524288,
    "received_bytes": 24117248,
    "maximum_outgoing_bytes": 4096,
    "maximum_incoming_bytes": 16384,
    "source_staging_max_bytes": 4096,
    "destination_staging_max_bytes": 0,
    "combined_staging_max_bytes": 4096,
    "source_copy_max_bytes": 4096,
    "destination_copy_max_bytes": 0,
    "local_copy_runs": 256,
    "affine_copy_loops": 128,
    "copy_cycles_floor": 512,
    "replay_address_ranges_valid": true
  },
  {
    "phase": 31,
    "mode": "destination",
    "label": "op Some(12) OperatorInputs: TensorShape([32, 577, 64]) F16 Amp(TransposedRight) -> TensorShape([32, 64, 64]) Amp(TransposedRight)",
    "transfers": 256,
    "receive_fragments": 11776,
    "maximum_endpoint_fragments": 10,
    "transmitted_bytes": 524288,
    "received_bytes": 24117248,
    "maximum_outgoing_bytes": 4096,
    "maximum_incoming_bytes": 16384,
    "source_staging_max_bytes": 0,
    "destination_staging_max_bytes": 16384,
    "combined_staging_max_bytes": 16384,
    "source_copy_max_bytes": 0,
    "destination_copy_max_bytes": 16384,
    "local_copy_runs": 29440,
    "affine_copy_loops": 11776,
    "copy_cycles_floor": 2048,
    "replay_address_ranges_valid": true
  },
  {
    "phase": 31,
    "mode": "both",
    "label": "op Some(12) OperatorInputs: TensorShape([32, 577, 64]) F16 Amp(TransposedRight) -> TensorShape([32, 64, 64]) Amp(TransposedRight)",
    "transfers": 128,
    "receive_fragments": 5888,
    "maximum_endpoint_fragments": 5,
    "transmitted_bytes": 524288,
    "received_bytes": 24117248,
    "maximum_outgoing_bytes": 4096,
    "maximum_incoming_bytes": 16384,
    "source_staging_max_bytes": 4096,
    "destination_staging_max_bytes": 16384,
    "combined_staging_max_bytes": 20480,
    "source_copy_max_bytes": 4096,
    "destination_copy_max_bytes": 16384,
    "local_copy_runs": 29696,
    "affine_copy_loops": 11904,
    "copy_cycles_floor": 2560,
    "replay_address_ranges_valid": true
  },
  {
    "phase": 32,
    "mode": "original",
    "label": "op Some(12) OperatorInputs: TensorShape([32, 577, 64]) F16 Amp(TransposedRight) -> TensorShape([32, 64, 64]) Amp(TransposedRight)",
    "transfers": 640,
    "receive_fragments": 29440,
    "maximum_endpoint_fragments": 25,
    "transmitted_bytes": 524288,
    "received_bytes": 24117248,
    "maximum_outgoing_bytes": 4096,
    "maximum_incoming_bytes": 16384,
    "source_staging_max_bytes": 0,
    "destination_staging_max_bytes": 0,
    "combined_staging_max_bytes": 0,
    "source_copy_max_bytes": 0,
    "destination_copy_max_bytes": 0,
    "local_copy_runs": 0,
    "affine_copy_loops": 0,
    "copy_cycles_floor": 0,
    "replay_address_ranges_valid": true
  },
  {
    "phase": 32,
    "mode": "source",
    "label": "op Some(12) OperatorInputs: TensorShape([32, 577, 64]) F16 Amp(TransposedRight) -> TensorShape([32, 64, 64]) Amp(TransposedRight)",
    "transfers": 640,
    "receive_fragments": 29440,
    "maximum_endpoint_fragments": 25,
    "transmitted_bytes": 524288,
    "received_bytes": 24117248,
    "maximum_outgoing_bytes": 4096,
    "maximum_incoming_bytes": 16384,
    "source_staging_max_bytes": 4096,
    "destination_staging_max_bytes": 0,
    "combined_staging_max_bytes": 4096,
    "source_copy_max_bytes": 4096,
    "destination_copy_max_bytes": 0,
    "local_copy_runs": 256,
    "affine_copy_loops": 128,
    "copy_cycles_floor": 512,
    "replay_address_ranges_valid": true
  },
  {
    "phase": 32,
    "mode": "destination",
    "label": "op Some(12) OperatorInputs: TensorShape([32, 577, 64]) F16 Amp(TransposedRight) -> TensorShape([32, 64, 64]) Amp(TransposedRight)",
    "transfers": 256,
    "receive_fragments": 11776,
    "maximum_endpoint_fragments": 10,
    "transmitted_bytes": 524288,
    "received_bytes": 24117248,
    "maximum_outgoing_bytes": 4096,
    "maximum_incoming_bytes": 16384,
    "source_staging_max_bytes": 0,
    "destination_staging_max_bytes": 16384,
    "combined_staging_max_bytes": 16384,
    "source_copy_max_bytes": 0,
    "destination_copy_max_bytes": 16384,
    "local_copy_runs": 29440,
    "affine_copy_loops": 11776,
    "copy_cycles_floor": 2048,
    "replay_address_ranges_valid": true
  },
  {
    "phase": 32,
    "mode": "both",
    "label": "op Some(12) OperatorInputs: TensorShape([32, 577, 64]) F16 Amp(TransposedRight) -> TensorShape([32, 64, 64]) Amp(TransposedRight)",
    "transfers": 128,
    "receive_fragments": 5888,
    "maximum_endpoint_fragments": 5,
    "transmitted_bytes": 524288,
    "received_bytes": 24117248,
    "maximum_outgoing_bytes": 4096,
    "maximum_incoming_bytes": 16384,
    "source_staging_max_bytes": 4096,
    "destination_staging_max_bytes": 16384,
    "combined_staging_max_bytes": 20480,
    "source_copy_max_bytes": 4096,
    "destination_copy_max_bytes": 16384,
    "local_copy_runs": 29696,
    "affine_copy_loops": 11904,
    "copy_cycles_floor": 2560,
    "replay_address_ranges_valid": true
  },
  {
    "phase": 33,
    "mode": "original",
    "label": "op Some(12) OperatorInputs: TensorShape([32, 577, 64]) F16 Amp(TransposedRight) -> TensorShape([32, 64, 64]) Amp(TransposedRight)",
    "transfers": 640,
    "receive_fragments": 29440,
    "maximum_endpoint_fragments": 25,
    "transmitted_bytes": 524288,
    "received_bytes": 24117248,
    "maximum_outgoing_bytes": 4096,
    "maximum_incoming_bytes": 16384,
    "source_staging_max_bytes": 0,
    "destination_staging_max_bytes": 0,
    "combined_staging_max_bytes": 0,
    "source_copy_max_bytes": 0,
    "destination_copy_max_bytes": 0,
    "local_copy_runs": 0,
    "affine_copy_loops": 0,
    "copy_cycles_floor": 0,
    "replay_address_ranges_valid": true
  },
  {
    "phase": 33,
    "mode": "source",
    "label": "op Some(12) OperatorInputs: TensorShape([32, 577, 64]) F16 Amp(TransposedRight) -> TensorShape([32, 64, 64]) Amp(TransposedRight)",
    "transfers": 640,
    "receive_fragments": 29440,
    "maximum_endpoint_fragments": 25,
    "transmitted_bytes": 524288,
    "received_bytes": 24117248,
    "maximum_outgoing_bytes": 4096,
    "maximum_incoming_bytes": 16384,
    "source_staging_max_bytes": 4096,
    "destination_staging_max_bytes": 0,
    "combined_staging_max_bytes": 4096,
    "source_copy_max_bytes": 4096,
    "destination_copy_max_bytes": 0,
    "local_copy_runs": 256,
    "affine_copy_loops": 128,
    "copy_cycles_floor": 512,
    "replay_address_ranges_valid": true
  },
  {
    "phase": 33,
    "mode": "destination",
    "label": "op Some(12) OperatorInputs: TensorShape([32, 577, 64]) F16 Amp(TransposedRight) -> TensorShape([32, 64, 64]) Amp(TransposedRight)",
    "transfers": 256,
    "receive_fragments": 11776,
    "maximum_endpoint_fragments": 10,
    "transmitted_bytes": 524288,
    "received_bytes": 24117248,
    "maximum_outgoing_bytes": 4096,
    "maximum_incoming_bytes": 16384,
    "source_staging_max_bytes": 0,
    "destination_staging_max_bytes": 16384,
    "combined_staging_max_bytes": 16384,
    "source_copy_max_bytes": 0,
    "destination_copy_max_bytes": 16384,
    "local_copy_runs": 29440,
    "affine_copy_loops": 11776,
    "copy_cycles_floor": 2048,
    "replay_address_ranges_valid": true
  },
  {
    "phase": 33,
    "mode": "both",
    "label": "op Some(12) OperatorInputs: TensorShape([32, 577, 64]) F16 Amp(TransposedRight) -> TensorShape([32, 64, 64]) Amp(TransposedRight)",
    "transfers": 128,
    "receive_fragments": 5888,
    "maximum_endpoint_fragments": 5,
    "transmitted_bytes": 524288,
    "received_bytes": 24117248,
    "maximum_outgoing_bytes": 4096,
    "maximum_incoming_bytes": 16384,
    "source_staging_max_bytes": 4096,
    "destination_staging_max_bytes": 16384,
    "combined_staging_max_bytes": 20480,
    "source_copy_max_bytes": 4096,
    "destination_copy_max_bytes": 16384,
    "local_copy_runs": 29696,
    "affine_copy_loops": 11904,
    "copy_cycles_floor": 2560,
    "replay_address_ranges_valid": true
  },
  {
    "phase": 34,
    "mode": "original",
    "label": "op Some(12) OperatorInputs: TensorShape([32, 577, 64]) F16 Amp(TransposedRight) -> TensorShape([32, 64, 64]) Amp(TransposedRight)",
    "transfers": 640,
    "receive_fragments": 29440,
    "maximum_endpoint_fragments": 25,
    "transmitted_bytes": 524288,
    "received_bytes": 24117248,
    "maximum_outgoing_bytes": 4096,
    "maximum_incoming_bytes": 16384,
    "source_staging_max_bytes": 0,
    "destination_staging_max_bytes": 0,
    "combined_staging_max_bytes": 0,
    "source_copy_max_bytes": 0,
    "destination_copy_max_bytes": 0,
    "local_copy_runs": 0,
    "affine_copy_loops": 0,
    "copy_cycles_floor": 0,
    "replay_address_ranges_valid": true
  },
  {
    "phase": 34,
    "mode": "source",
    "label": "op Some(12) OperatorInputs: TensorShape([32, 577, 64]) F16 Amp(TransposedRight) -> TensorShape([32, 64, 64]) Amp(TransposedRight)",
    "transfers": 640,
    "receive_fragments": 29440,
    "maximum_endpoint_fragments": 25,
    "transmitted_bytes": 524288,
    "received_bytes": 24117248,
    "maximum_outgoing_bytes": 4096,
    "maximum_incoming_bytes": 16384,
    "source_staging_max_bytes": 4096,
    "destination_staging_max_bytes": 0,
    "combined_staging_max_bytes": 4096,
    "source_copy_max_bytes": 4096,
    "destination_copy_max_bytes": 0,
    "local_copy_runs": 256,
    "affine_copy_loops": 128,
    "copy_cycles_floor": 512,
    "replay_address_ranges_valid": true
  },
  {
    "phase": 34,
    "mode": "destination",
    "label": "op Some(12) OperatorInputs: TensorShape([32, 577, 64]) F16 Amp(TransposedRight) -> TensorShape([32, 64, 64]) Amp(TransposedRight)",
    "transfers": 256,
    "receive_fragments": 11776,
    "maximum_endpoint_fragments": 10,
    "transmitted_bytes": 524288,
    "received_bytes": 24117248,
    "maximum_outgoing_bytes": 4096,
    "maximum_incoming_bytes": 16384,
    "source_staging_max_bytes": 0,
    "destination_staging_max_bytes": 16384,
    "combined_staging_max_bytes": 16384,
    "source_copy_max_bytes": 0,
    "destination_copy_max_bytes": 16384,
    "local_copy_runs": 29440,
    "affine_copy_loops": 11776,
    "copy_cycles_floor": 2048,
    "replay_address_ranges_valid": true
  },
  {
    "phase": 34,
    "mode": "both",
    "label": "op Some(12) OperatorInputs: TensorShape([32, 577, 64]) F16 Amp(TransposedRight) -> TensorShape([32, 64, 64]) Amp(TransposedRight)",
    "transfers": 128,
    "receive_fragments": 5888,
    "maximum_endpoint_fragments": 5,
    "transmitted_bytes": 524288,
    "received_bytes": 24117248,
    "maximum_outgoing_bytes": 4096,
    "maximum_incoming_bytes": 16384,
    "source_staging_max_bytes": 4096,
    "destination_staging_max_bytes": 16384,
    "combined_staging_max_bytes": 20480,
    "source_copy_max_bytes": 4096,
    "destination_copy_max_bytes": 16384,
    "local_copy_runs": 29696,
    "affine_copy_loops": 11904,
    "copy_cycles_floor": 2560,
    "replay_address_ranges_valid": true
  },
  {
    "phase": 35,
    "mode": "original",
    "label": "op Some(12) OperatorInputs: TensorShape([32, 577, 64]) F16 Amp(TransposedRight) -> TensorShape([32, 64, 64]) Amp(TransposedRight)",
    "transfers": 640,
    "receive_fragments": 29440,
    "maximum_endpoint_fragments": 25,
    "transmitted_bytes": 524288,
    "received_bytes": 24117248,
    "maximum_outgoing_bytes": 4096,
    "maximum_incoming_bytes": 16384,
    "source_staging_max_bytes": 0,
    "destination_staging_max_bytes": 0,
    "combined_staging_max_bytes": 0,
    "source_copy_max_bytes": 0,
    "destination_copy_max_bytes": 0,
    "local_copy_runs": 0,
    "affine_copy_loops": 0,
    "copy_cycles_floor": 0,
    "replay_address_ranges_valid": true
  },
  {
    "phase": 35,
    "mode": "source",
    "label": "op Some(12) OperatorInputs: TensorShape([32, 577, 64]) F16 Amp(TransposedRight) -> TensorShape([32, 64, 64]) Amp(TransposedRight)",
    "transfers": 640,
    "receive_fragments": 29440,
    "maximum_endpoint_fragments": 25,
    "transmitted_bytes": 524288,
    "received_bytes": 24117248,
    "maximum_outgoing_bytes": 4096,
    "maximum_incoming_bytes": 16384,
    "source_staging_max_bytes": 4096,
    "destination_staging_max_bytes": 0,
    "combined_staging_max_bytes": 4096,
    "source_copy_max_bytes": 4096,
    "destination_copy_max_bytes": 0,
    "local_copy_runs": 256,
    "affine_copy_loops": 128,
    "copy_cycles_floor": 512,
    "replay_address_ranges_valid": true
  },
  {
    "phase": 35,
    "mode": "destination",
    "label": "op Some(12) OperatorInputs: TensorShape([32, 577, 64]) F16 Amp(TransposedRight) -> TensorShape([32, 64, 64]) Amp(TransposedRight)",
    "transfers": 256,
    "receive_fragments": 11776,
    "maximum_endpoint_fragments": 10,
    "transmitted_bytes": 524288,
    "received_bytes": 24117248,
    "maximum_outgoing_bytes": 4096,
    "maximum_incoming_bytes": 16384,
    "source_staging_max_bytes": 0,
    "destination_staging_max_bytes": 16384,
    "combined_staging_max_bytes": 16384,
    "source_copy_max_bytes": 0,
    "destination_copy_max_bytes": 16384,
    "local_copy_runs": 29440,
    "affine_copy_loops": 11776,
    "copy_cycles_floor": 2048,
    "replay_address_ranges_valid": true
  },
  {
    "phase": 35,
    "mode": "both",
    "label": "op Some(12) OperatorInputs: TensorShape([32, 577, 64]) F16 Amp(TransposedRight) -> TensorShape([32, 64, 64]) Amp(TransposedRight)",
    "transfers": 128,
    "receive_fragments": 5888,
    "maximum_endpoint_fragments": 5,
    "transmitted_bytes": 524288,
    "received_bytes": 24117248,
    "maximum_outgoing_bytes": 4096,
    "maximum_incoming_bytes": 16384,
    "source_staging_max_bytes": 4096,
    "destination_staging_max_bytes": 16384,
    "combined_staging_max_bytes": 20480,
    "source_copy_max_bytes": 4096,
    "destination_copy_max_bytes": 16384,
    "local_copy_runs": 29696,
    "affine_copy_loops": 11904,
    "copy_cycles_floor": 2560,
    "replay_address_ranges_valid": true
  },
  {
    "phase": 36,
    "mode": "original",
    "label": "op Some(12) OperatorInputs: TensorShape([32, 577, 64]) F16 Amp(TransposedRight) -> TensorShape([32, 1, 64]) Amp(TransposedRight)",
    "transfers": 640,
    "receive_fragments": 29440,
    "maximum_endpoint_fragments": 25,
    "transmitted_bytes": 524288,
    "received_bytes": 24117248,
    "maximum_outgoing_bytes": 4096,
    "maximum_incoming_bytes": 16384,
    "source_staging_max_bytes": 0,
    "destination_staging_max_bytes": 0,
    "combined_staging_max_bytes": 0,
    "source_copy_max_bytes": 0,
    "destination_copy_max_bytes": 0,
    "local_copy_runs": 0,
    "affine_copy_loops": 0,
    "copy_cycles_floor": 0,
    "replay_address_ranges_valid": true
  },
  {
    "phase": 36,
    "mode": "source",
    "label": "op Some(12) OperatorInputs: TensorShape([32, 577, 64]) F16 Amp(TransposedRight) -> TensorShape([32, 1, 64]) Amp(TransposedRight)",
    "transfers": 640,
    "receive_fragments": 29440,
    "maximum_endpoint_fragments": 25,
    "transmitted_bytes": 524288,
    "received_bytes": 24117248,
    "maximum_outgoing_bytes": 4096,
    "maximum_incoming_bytes": 16384,
    "source_staging_max_bytes": 4096,
    "destination_staging_max_bytes": 0,
    "combined_staging_max_bytes": 4096,
    "source_copy_max_bytes": 4096,
    "destination_copy_max_bytes": 0,
    "local_copy_runs": 256,
    "affine_copy_loops": 128,
    "copy_cycles_floor": 512,
    "replay_address_ranges_valid": true
  },
  {
    "phase": 36,
    "mode": "destination",
    "label": "op Some(12) OperatorInputs: TensorShape([32, 577, 64]) F16 Amp(TransposedRight) -> TensorShape([32, 1, 64]) Amp(TransposedRight)",
    "transfers": 256,
    "receive_fragments": 11776,
    "maximum_endpoint_fragments": 10,
    "transmitted_bytes": 524288,
    "received_bytes": 24117248,
    "maximum_outgoing_bytes": 4096,
    "maximum_incoming_bytes": 16384,
    "source_staging_max_bytes": 0,
    "destination_staging_max_bytes": 16384,
    "combined_staging_max_bytes": 16384,
    "source_copy_max_bytes": 0,
    "destination_copy_max_bytes": 16384,
    "local_copy_runs": 29440,
    "affine_copy_loops": 11776,
    "copy_cycles_floor": 2048,
    "replay_address_ranges_valid": true
  },
  {
    "phase": 36,
    "mode": "both",
    "label": "op Some(12) OperatorInputs: TensorShape([32, 577, 64]) F16 Amp(TransposedRight) -> TensorShape([32, 1, 64]) Amp(TransposedRight)",
    "transfers": 128,
    "receive_fragments": 5888,
    "maximum_endpoint_fragments": 5,
    "transmitted_bytes": 524288,
    "received_bytes": 24117248,
    "maximum_outgoing_bytes": 4096,
    "maximum_incoming_bytes": 16384,
    "source_staging_max_bytes": 4096,
    "destination_staging_max_bytes": 16384,
    "combined_staging_max_bytes": 20480,
    "source_copy_max_bytes": 4096,
    "destination_copy_max_bytes": 16384,
    "local_copy_runs": 29696,
    "affine_copy_loops": 11904,
    "copy_cycles_floor": 2560,
    "replay_address_ranges_valid": true
  },
  {
    "phase": 97,
    "mode": "original",
    "label": "op Some(40) LayoutRearrangement: TensorShape([2, 1, 1024]) F16 RowMajor -> TensorShape([2, 1, 1024]) RowMajor",
    "transfers": 22,
    "receive_fragments": 2924,
    "maximum_endpoint_fragments": 22,
    "transmitted_bytes": 4096,
    "received_bytes": 544384,
    "maximum_outgoing_bytes": 4096,
    "maximum_incoming_bytes": 384,
    "source_staging_max_bytes": 0,
    "destination_staging_max_bytes": 0,
    "combined_staging_max_bytes": 0,
    "source_copy_max_bytes": 0,
    "destination_copy_max_bytes": 0,
    "local_copy_runs": 0,
    "affine_copy_loops": 0,
    "copy_cycles_floor": 0,
    "replay_address_ranges_valid": true
  },
  {
    "phase": 97,
    "mode": "source",
    "label": "op Some(40) LayoutRearrangement: TensorShape([2, 1, 1024]) F16 RowMajor -> TensorShape([2, 1, 1024]) RowMajor",
    "transfers": 11,
    "receive_fragments": 1462,
    "maximum_endpoint_fragments": 11,
    "transmitted_bytes": 4096,
    "received_bytes": 544384,
    "maximum_outgoing_bytes": 4096,
    "maximum_incoming_bytes": 384,
    "source_staging_max_bytes": 4096,
    "destination_staging_max_bytes": 0,
    "combined_staging_max_bytes": 4096,
    "source_copy_max_bytes": 4096,
    "destination_copy_max_bytes": 0,
    "local_copy_runs": 22,
    "affine_copy_loops": 11,
    "copy_cycles_floor": 512,
    "replay_address_ranges_valid": true
  },
  {
    "phase": 97,
    "mode": "destination",
    "label": "op Some(40) LayoutRearrangement: TensorShape([2, 1, 1024]) F16 RowMajor -> TensorShape([2, 1, 1024]) RowMajor",
    "transfers": 22,
    "receive_fragments": 2924,
    "maximum_endpoint_fragments": 22,
    "transmitted_bytes": 4096,
    "received_bytes": 544384,
    "maximum_outgoing_bytes": 4096,
    "maximum_incoming_bytes": 384,
    "source_staging_max_bytes": 0,
    "destination_staging_max_bytes": 384,
    "combined_staging_max_bytes": 384,
    "source_copy_max_bytes": 0,
    "destination_copy_max_bytes": 384,
    "local_copy_runs": 1462,
    "affine_copy_loops": 1462,
    "copy_cycles_floor": 48,
    "replay_address_ranges_valid": true
  },
  {
    "phase": 97,
    "mode": "both",
    "label": "op Some(40) LayoutRearrangement: TensorShape([2, 1, 1024]) F16 RowMajor -> TensorShape([2, 1, 1024]) RowMajor",
    "transfers": 11,
    "receive_fragments": 1462,
    "maximum_endpoint_fragments": 11,
    "transmitted_bytes": 4096,
    "received_bytes": 544384,
    "maximum_outgoing_bytes": 4096,
    "maximum_incoming_bytes": 384,
    "source_staging_max_bytes": 4096,
    "destination_staging_max_bytes": 384,
    "combined_staging_max_bytes": 4096,
    "source_copy_max_bytes": 4096,
    "destination_copy_max_bytes": 384,
    "local_copy_runs": 1484,
    "affine_copy_loops": 1473,
    "copy_cycles_floor": 560,
    "replay_address_ranges_valid": true
  },
  {
    "phase": 104,
    "mode": "original",
    "label": "op Some(43) OperatorInputs: TensorShape([2, 1, 1024]) F16 Amp(Left) -> TensorShape([1344]) RowMajor",
    "transfers": 1344,
    "receive_fragments": 1344,
    "maximum_endpoint_fragments": 21,
    "transmitted_bytes": 172032,
    "received_bytes": 172032,
    "maximum_outgoing_bytes": 128,
    "maximum_incoming_bytes": 2688,
    "source_staging_max_bytes": 0,
    "destination_staging_max_bytes": 0,
    "combined_staging_max_bytes": 0,
    "source_copy_max_bytes": 0,
    "destination_copy_max_bytes": 0,
    "local_copy_runs": 0,
    "affine_copy_loops": 0,
    "copy_cycles_floor": 0,
    "replay_address_ranges_valid": true
  },
  {
    "phase": 104,
    "mode": "source",
    "label": "op Some(43) OperatorInputs: TensorShape([2, 1, 1024]) F16 Amp(Left) -> TensorShape([1344]) RowMajor",
    "transfers": 1344,
    "receive_fragments": 1344,
    "maximum_endpoint_fragments": 21,
    "transmitted_bytes": 172032,
    "received_bytes": 172032,
    "maximum_outgoing_bytes": 128,
    "maximum_incoming_bytes": 2688,
    "source_staging_max_bytes": 128,
    "destination_staging_max_bytes": 0,
    "combined_staging_max_bytes": 128,
    "source_copy_max_bytes": 128,
    "destination_copy_max_bytes": 0,
    "local_copy_runs": 1344,
    "affine_copy_loops": 1344,
    "copy_cycles_floor": 16,
    "replay_address_ranges_valid": true
  },
  {
    "phase": 104,
    "mode": "destination",
    "label": "op Some(43) OperatorInputs: TensorShape([2, 1, 1024]) F16 Amp(Left) -> TensorShape([1344]) RowMajor",
    "transfers": 1344,
    "receive_fragments": 1344,
    "maximum_endpoint_fragments": 21,
    "transmitted_bytes": 172032,
    "received_bytes": 172032,
    "maximum_outgoing_bytes": 128,
    "maximum_incoming_bytes": 2688,
    "source_staging_max_bytes": 0,
    "destination_staging_max_bytes": 2688,
    "combined_staging_max_bytes": 2688,
    "source_copy_max_bytes": 0,
    "destination_copy_max_bytes": 2688,
    "local_copy_runs": 64,
    "affine_copy_loops": 64,
    "copy_cycles_floor": 336,
    "replay_address_ranges_valid": true
  },
  {
    "phase": 104,
    "mode": "both",
    "label": "op Some(43) OperatorInputs: TensorShape([2, 1, 1024]) F16 Amp(Left) -> TensorShape([1344]) RowMajor",
    "transfers": 1344,
    "receive_fragments": 1344,
    "maximum_endpoint_fragments": 21,
    "transmitted_bytes": 172032,
    "received_bytes": 172032,
    "maximum_outgoing_bytes": 128,
    "maximum_incoming_bytes": 2688,
    "source_staging_max_bytes": 128,
    "destination_staging_max_bytes": 2688,
    "combined_staging_max_bytes": 2688,
    "source_copy_max_bytes": 128,
    "destination_copy_max_bytes": 2688,
    "local_copy_runs": 1408,
    "affine_copy_loops": 1408,
    "copy_cycles_floor": 352,
    "replay_address_ranges_valid": true
  },
  {
    "phase": 13,
    "mode": "original",
    "label": "op Some(2) OperatorInputs: TensorShape([2, 577, 1024]) F16 RowMajor -> TensorShape([2, 577, 1024]) RowMajor",
    "transfers": 593,
    "receive_fragments": 19041,
    "maximum_endpoint_fragments": 19,
    "transmitted_bytes": 1185792,
    "received_bytes": 5908480,
    "maximum_outgoing_bytes": 2048,
    "maximum_incoming_bytes": 6144,
    "source_staging_max_bytes": 0,
    "destination_staging_max_bytes": 0,
    "combined_staging_max_bytes": 0,
    "source_copy_max_bytes": 0,
    "destination_copy_max_bytes": 0,
    "local_copy_runs": 0,
    "affine_copy_loops": 0,
    "copy_cycles_floor": 0,
    "replay_address_ranges_valid": true
  },
  {
    "phase": 13,
    "mode": "source",
    "label": "op Some(2) OperatorInputs: TensorShape([2, 577, 1024]) F16 RowMajor -> TensorShape([2, 577, 1024]) RowMajor",
    "transfers": 593,
    "receive_fragments": 19041,
    "maximum_endpoint_fragments": 19,
    "transmitted_bytes": 1185792,
    "received_bytes": 5908480,
    "maximum_outgoing_bytes": 2048,
    "maximum_incoming_bytes": 6144,
    "source_staging_max_bytes": 2048,
    "destination_staging_max_bytes": 0,
    "combined_staging_max_bytes": 2048,
    "source_copy_max_bytes": 2048,
    "destination_copy_max_bytes": 0,
    "local_copy_runs": 586,
    "affine_copy_loops": 585,
    "copy_cycles_floor": 256,
    "replay_address_ranges_valid": true
  },
  {
    "phase": 13,
    "mode": "destination",
    "label": "op Some(2) OperatorInputs: TensorShape([2, 577, 1024]) F16 RowMajor -> TensorShape([2, 577, 1024]) RowMajor",
    "transfers": 586,
    "receive_fragments": 10963,
    "maximum_endpoint_fragments": 12,
    "transmitted_bytes": 1185792,
    "received_bytes": 5908480,
    "maximum_outgoing_bytes": 2048,
    "maximum_incoming_bytes": 6144,
    "source_staging_max_bytes": 0,
    "destination_staging_max_bytes": 6144,
    "combined_staging_max_bytes": 6144,
    "source_copy_max_bytes": 0,
    "destination_copy_max_bytes": 6144,
    "local_copy_runs": 18464,
    "affine_copy_loops": 9809,
    "copy_cycles_floor": 768,
    "replay_address_ranges_valid": true
  },
  {
    "phase": 13,
    "mode": "both",
    "label": "op Some(2) OperatorInputs: TensorShape([2, 577, 1024]) F16 RowMajor -> TensorShape([2, 577, 1024]) RowMajor",
    "transfers": 585,
    "receive_fragments": 9809,
    "maximum_endpoint_fragments": 10,
    "transmitted_bytes": 1185792,
    "received_bytes": 5908480,
    "maximum_outgoing_bytes": 2048,
    "maximum_incoming_bytes": 6144,
    "source_staging_max_bytes": 2048,
    "destination_staging_max_bytes": 6144,
    "combined_staging_max_bytes": 6656,
    "source_copy_max_bytes": 2048,
    "destination_copy_max_bytes": 6144,
    "local_copy_runs": 19050,
    "affine_copy_loops": 10394,
    "copy_cycles_floor": 1024,
    "replay_address_ranges_valid": true
  },
  {
    "phase": 15,
    "mode": "original",
    "label": "op Some(3) OperatorInputs: TensorShape([2, 577, 1024]) F16 RowMajor -> TensorShape([2, 577, 1024]) RowMajor",
    "transfers": 593,
    "receive_fragments": 19041,
    "maximum_endpoint_fragments": 19,
    "transmitted_bytes": 1185792,
    "received_bytes": 5908480,
    "maximum_outgoing_bytes": 2560,
    "maximum_incoming_bytes": 6144,
    "source_staging_max_bytes": 0,
    "destination_staging_max_bytes": 0,
    "combined_staging_max_bytes": 0,
    "source_copy_max_bytes": 0,
    "destination_copy_max_bytes": 0,
    "local_copy_runs": 0,
    "affine_copy_loops": 0,
    "copy_cycles_floor": 0,
    "replay_address_ranges_valid": true
  },
  {
    "phase": 15,
    "mode": "source",
    "label": "op Some(3) OperatorInputs: TensorShape([2, 577, 1024]) F16 RowMajor -> TensorShape([2, 577, 1024]) RowMajor",
    "transfers": 593,
    "receive_fragments": 19041,
    "maximum_endpoint_fragments": 19,
    "transmitted_bytes": 1185792,
    "received_bytes": 5908480,
    "maximum_outgoing_bytes": 2560,
    "maximum_incoming_bytes": 6144,
    "source_staging_max_bytes": 2560,
    "destination_staging_max_bytes": 0,
    "combined_staging_max_bytes": 2560,
    "source_copy_max_bytes": 2560,
    "destination_copy_max_bytes": 0,
    "local_copy_runs": 593,
    "affine_copy_loops": 585,
    "copy_cycles_floor": 320,
    "replay_address_ranges_valid": true
  },
  {
    "phase": 15,
    "mode": "destination",
    "label": "op Some(3) OperatorInputs: TensorShape([2, 577, 1024]) F16 RowMajor -> TensorShape([2, 577, 1024]) RowMajor",
    "transfers": 593,
    "receive_fragments": 19041,
    "maximum_endpoint_fragments": 19,
    "transmitted_bytes": 1185792,
    "received_bytes": 5908480,
    "maximum_outgoing_bytes": 2560,
    "maximum_incoming_bytes": 6144,
    "source_staging_max_bytes": 0,
    "destination_staging_max_bytes": 6144,
    "combined_staging_max_bytes": 6144,
    "source_copy_max_bytes": 0,
    "destination_copy_max_bytes": 6144,
    "local_copy_runs": 18981,
    "affine_copy_loops": 9809,
    "copy_cycles_floor": 768,
    "replay_address_ranges_valid": true
  },
  {
    "phase": 15,
    "mode": "both",
    "label": "op Some(3) OperatorInputs: TensorShape([2, 577, 1024]) F16 RowMajor -> TensorShape([2, 577, 1024]) RowMajor",
    "transfers": 585,
    "receive_fragments": 9809,
    "maximum_endpoint_fragments": 10,
    "transmitted_bytes": 1185792,
    "received_bytes": 5908480,
    "maximum_outgoing_bytes": 2560,
    "maximum_incoming_bytes": 6144,
    "source_staging_max_bytes": 2560,
    "destination_staging_max_bytes": 6144,
    "combined_staging_max_bytes": 6656,
    "source_copy_max_bytes": 2560,
    "destination_copy_max_bytes": 6144,
    "local_copy_runs": 19574,
    "affine_copy_loops": 10394,
    "copy_cycles_floor": 1088,
    "replay_address_ranges_valid": true
  },
  {
    "phase": 67,
    "mode": "original",
    "label": "op Some(25) OperatorInputs: TensorShape([2, 577, 1024]) F16 RowMajor -> TensorShape([2, 577, 1024]) RowMajor",
    "transfers": 593,
    "receive_fragments": 19041,
    "maximum_endpoint_fragments": 19,
    "transmitted_bytes": 1185792,
    "received_bytes": 5908480,
    "maximum_outgoing_bytes": 2048,
    "maximum_incoming_bytes": 6144,
    "source_staging_max_bytes": 0,
    "destination_staging_max_bytes": 0,
    "combined_staging_max_bytes": 0,
    "source_copy_max_bytes": 0,
    "destination_copy_max_bytes": 0,
    "local_copy_runs": 0,
    "affine_copy_loops": 0,
    "copy_cycles_floor": 0,
    "replay_address_ranges_valid": true
  },
  {
    "phase": 67,
    "mode": "source",
    "label": "op Some(25) OperatorInputs: TensorShape([2, 577, 1024]) F16 RowMajor -> TensorShape([2, 577, 1024]) RowMajor",
    "transfers": 593,
    "receive_fragments": 19041,
    "maximum_endpoint_fragments": 19,
    "transmitted_bytes": 1185792,
    "received_bytes": 5908480,
    "maximum_outgoing_bytes": 2048,
    "maximum_incoming_bytes": 6144,
    "source_staging_max_bytes": 2048,
    "destination_staging_max_bytes": 0,
    "combined_staging_max_bytes": 2048,
    "source_copy_max_bytes": 2048,
    "destination_copy_max_bytes": 0,
    "local_copy_runs": 585,
    "affine_copy_loops": 585,
    "copy_cycles_floor": 256,
    "replay_address_ranges_valid": true
  },
  {
    "phase": 67,
    "mode": "destination",
    "label": "op Some(25) OperatorInputs: TensorShape([2, 577, 1024]) F16 RowMajor -> TensorShape([2, 577, 1024]) RowMajor",
    "transfers": 585,
    "receive_fragments": 9809,
    "maximum_endpoint_fragments": 10,
    "transmitted_bytes": 1185792,
    "received_bytes": 5908480,
    "maximum_outgoing_bytes": 2048,
    "maximum_incoming_bytes": 6144,
    "source_staging_max_bytes": 0,
    "destination_staging_max_bytes": 6144,
    "combined_staging_max_bytes": 6144,
    "source_copy_max_bytes": 0,
    "destination_copy_max_bytes": 6144,
    "local_copy_runs": 18464,
    "affine_copy_loops": 9809,
    "copy_cycles_floor": 768,
    "replay_address_ranges_valid": true
  },
  {
    "phase": 67,
    "mode": "both",
    "label": "op Some(25) OperatorInputs: TensorShape([2, 577, 1024]) F16 RowMajor -> TensorShape([2, 577, 1024]) RowMajor",
    "transfers": 585,
    "receive_fragments": 9809,
    "maximum_endpoint_fragments": 10,
    "transmitted_bytes": 1185792,
    "received_bytes": 5908480,
    "maximum_outgoing_bytes": 2048,
    "maximum_incoming_bytes": 6144,
    "source_staging_max_bytes": 2048,
    "destination_staging_max_bytes": 6144,
    "combined_staging_max_bytes": 6656,
    "source_copy_max_bytes": 2048,
    "destination_copy_max_bytes": 6144,
    "local_copy_runs": 19049,
    "affine_copy_loops": 10394,
    "copy_cycles_floor": 1024,
    "replay_address_ranges_valid": true
  },
  {
    "phase": 56,
    "mode": "original",
    "label": "op Some(17) OperatorInputs: TensorShape([2, 577, 1024]) F16 RowMajor -> TensorShape([2, 577, 1024]) RowMajor",
    "transfers": 593,
    "receive_fragments": 19041,
    "maximum_endpoint_fragments": 18,
    "transmitted_bytes": 1185792,
    "received_bytes": 5908480,
    "maximum_outgoing_bytes": 2304,
    "maximum_incoming_bytes": 6144,
    "source_staging_max_bytes": 0,
    "destination_staging_max_bytes": 0,
    "combined_staging_max_bytes": 0,
    "source_copy_max_bytes": 0,
    "destination_copy_max_bytes": 0,
    "local_copy_runs": 0,
    "affine_copy_loops": 0,
    "copy_cycles_floor": 0,
    "replay_address_ranges_valid": true
  },
  {
    "phase": 56,
    "mode": "source",
    "label": "op Some(17) OperatorInputs: TensorShape([2, 577, 1024]) F16 RowMajor -> TensorShape([2, 577, 1024]) RowMajor",
    "transfers": 593,
    "receive_fragments": 19041,
    "maximum_endpoint_fragments": 18,
    "transmitted_bytes": 1185792,
    "received_bytes": 5908480,
    "maximum_outgoing_bytes": 2304,
    "maximum_incoming_bytes": 6144,
    "source_staging_max_bytes": 2304,
    "destination_staging_max_bytes": 0,
    "combined_staging_max_bytes": 2304,
    "source_copy_max_bytes": 2304,
    "destination_copy_max_bytes": 0,
    "local_copy_runs": 593,
    "affine_copy_loops": 593,
    "copy_cycles_floor": 288,
    "replay_address_ranges_valid": true
  },
  {
    "phase": 56,
    "mode": "destination",
    "label": "op Some(17) OperatorInputs: TensorShape([2, 577, 1024]) F16 RowMajor -> TensorShape([2, 577, 1024]) RowMajor",
    "transfers": 593,
    "receive_fragments": 19041,
    "maximum_endpoint_fragments": 18,
    "transmitted_bytes": 1185792,
    "received_bytes": 5908480,
    "maximum_outgoing_bytes": 2304,
    "maximum_incoming_bytes": 6144,
    "source_staging_max_bytes": 0,
    "destination_staging_max_bytes": 6144,
    "combined_staging_max_bytes": 6144,
    "source_copy_max_bytes": 0,
    "destination_copy_max_bytes": 6144,
    "local_copy_runs": 1876,
    "affine_copy_loops": 1522,
    "copy_cycles_floor": 768,
    "replay_address_ranges_valid": true
  },
  {
    "phase": 56,
    "mode": "both",
    "label": "op Some(17) OperatorInputs: TensorShape([2, 577, 1024]) F16 RowMajor -> TensorShape([2, 577, 1024]) RowMajor",
    "transfers": 593,
    "receive_fragments": 19041,
    "maximum_endpoint_fragments": 18,
    "transmitted_bytes": 1185792,
    "received_bytes": 5908480,
    "maximum_outgoing_bytes": 2304,
    "maximum_incoming_bytes": 6144,
    "source_staging_max_bytes": 2304,
    "destination_staging_max_bytes": 6144,
    "combined_staging_max_bytes": 6400,
    "source_copy_max_bytes": 2304,
    "destination_copy_max_bytes": 6144,
    "local_copy_runs": 2469,
    "affine_copy_loops": 2115,
    "copy_cycles_floor": 1056,
    "replay_address_ranges_valid": true
  },
  {
    "phase": 63,
    "mode": "original",
    "label": "op Some(21) OperatorInputs: TensorShape([4096, 1024]) F8F143 { scale_exponent: -4 } Amp(TransposedLeft) -> TensorShape([4096, 1024]) BlockMajor(Matrix { row_block: 288, column_block: 16 })",
    "transfers": 2395,
    "receive_fragments": 4315,
    "maximum_endpoint_fragments": 16,
    "transmitted_bytes": 8606208,
    "received_bytes": 26300928,
    "maximum_outgoing_bytes": 21504,
    "maximum_incoming_bytes": 18432,
    "source_staging_max_bytes": 0,
    "destination_staging_max_bytes": 0,
    "combined_staging_max_bytes": 0,
    "source_copy_max_bytes": 0,
    "destination_copy_max_bytes": 0,
    "local_copy_runs": 0,
    "affine_copy_loops": 0,
    "copy_cycles_floor": 0,
    "replay_address_ranges_valid": true
  },
  {
    "phase": 63,
    "mode": "source",
    "label": "op Some(21) OperatorInputs: TensorShape([4096, 1024]) F8F143 { scale_exponent: -4 } Amp(TransposedLeft) -> TensorShape([4096, 1024]) BlockMajor(Matrix { row_block: 288, column_block: 16 })",
    "transfers": 2155,
    "receive_fragments": 3115,
    "maximum_endpoint_fragments": 15,
    "transmitted_bytes": 8606208,
    "received_bytes": 26300928,
    "maximum_outgoing_bytes": 21504,
    "maximum_incoming_bytes": 18432,
    "source_staging_max_bytes": 21504,
    "destination_staging_max_bytes": 0,
    "combined_staging_max_bytes": 21504,
    "source_copy_max_bytes": 21504,
    "destination_copy_max_bytes": 0,
    "local_copy_runs": 1645,
    "affine_copy_loops": 1645,
    "copy_cycles_floor": 2688,
    "replay_address_ranges_valid": true
  },
  {
    "phase": 63,
    "mode": "destination",
    "label": "op Some(21) OperatorInputs: TensorShape([4096, 1024]) F8F143 { scale_exponent: -4 } Amp(TransposedLeft) -> TensorShape([4096, 1024]) BlockMajor(Matrix { row_block: 288, column_block: 16 })",
    "transfers": 2155,
    "receive_fragments": 3115,
    "maximum_endpoint_fragments": 15,
    "transmitted_bytes": 8606208,
    "received_bytes": 26300928,
    "maximum_outgoing_bytes": 21504,
    "maximum_incoming_bytes": 18432,
    "source_staging_max_bytes": 0,
    "destination_staging_max_bytes": 18432,
    "combined_staging_max_bytes": 18432,
    "source_copy_max_bytes": 0,
    "destination_copy_max_bytes": 18432,
    "local_copy_runs": 1490,
    "affine_copy_loops": 1442,
    "copy_cycles_floor": 2304,
    "replay_address_ranges_valid": true
  },
  {
    "phase": 63,
    "mode": "both",
    "label": "op Some(21) OperatorInputs: TensorShape([4096, 1024]) F8F143 { scale_exponent: -4 } Amp(TransposedLeft) -> TensorShape([4096, 1024]) BlockMajor(Matrix { row_block: 288, column_block: 16 })",
    "transfers": 2155,
    "receive_fragments": 3115,
    "maximum_endpoint_fragments": 15,
    "transmitted_bytes": 8606208,
    "received_bytes": 26300928,
    "maximum_outgoing_bytes": 21504,
    "maximum_incoming_bytes": 18432,
    "source_staging_max_bytes": 21504,
    "destination_staging_max_bytes": 18432,
    "combined_staging_max_bytes": 39936,
    "source_copy_max_bytes": 21504,
    "destination_copy_max_bytes": 18432,
    "local_copy_runs": 3135,
    "affine_copy_loops": 3087,
    "copy_cycles_floor": 4992,
    "replay_address_ranges_valid": true
  },
  {
    "phase": 100,
    "mode": "original",
    "label": "op Some(41) OperatorInputs: TensorShape([2, 1, 4096]) F16 Amp(Left) -> TensorShape([2, 1, 4096]) RowMajor",
    "transfers": 1535,
    "receive_fragments": 2047,
    "maximum_endpoint_fragments": 16,
    "transmitted_bytes": 24560,
    "received_bytes": 32752,
    "maximum_outgoing_bytes": 256,
    "maximum_incoming_bytes": 32,
    "source_staging_max_bytes": 0,
    "destination_staging_max_bytes": 0,
    "combined_staging_max_bytes": 0,
    "source_copy_max_bytes": 0,
    "destination_copy_max_bytes": 0,
    "local_copy_runs": 0,
    "affine_copy_loops": 0,
    "copy_cycles_floor": 0,
    "replay_address_ranges_valid": true
  },
  {
    "phase": 100,
    "mode": "source",
    "label": "op Some(41) OperatorInputs: TensorShape([2, 1, 4096]) F16 Amp(Left) -> TensorShape([2, 1, 4096]) RowMajor",
    "transfers": 1535,
    "receive_fragments": 2047,
    "maximum_endpoint_fragments": 16,
    "transmitted_bytes": 24560,
    "received_bytes": 32752,
    "maximum_outgoing_bytes": 256,
    "maximum_incoming_bytes": 32,
    "source_staging_max_bytes": 256,
    "destination_staging_max_bytes": 0,
    "combined_staging_max_bytes": 256,
    "source_copy_max_bytes": 256,
    "destination_copy_max_bytes": 0,
    "local_copy_runs": 544,
    "affine_copy_loops": 289,
    "copy_cycles_floor": 32,
    "replay_address_ranges_valid": true
  },
  {
    "phase": 100,
    "mode": "destination",
    "label": "op Some(41) OperatorInputs: TensorShape([2, 1, 4096]) F16 Amp(Left) -> TensorShape([2, 1, 4096]) RowMajor",
    "transfers": 1535,
    "receive_fragments": 2047,
    "maximum_endpoint_fragments": 16,
    "transmitted_bytes": 24560,
    "received_bytes": 32752,
    "maximum_outgoing_bytes": 256,
    "maximum_incoming_bytes": 32,
    "source_staging_max_bytes": 0,
    "destination_staging_max_bytes": 32,
    "combined_staging_max_bytes": 32,
    "source_copy_max_bytes": 0,
    "destination_copy_max_bytes": 32,
    "local_copy_runs": 1024,
    "affine_copy_loops": 1024,
    "copy_cycles_floor": 4,
    "replay_address_ranges_valid": true
  },
  {
    "phase": 100,
    "mode": "both",
    "label": "op Some(41) OperatorInputs: TensorShape([2, 1, 4096]) F16 Amp(Left) -> TensorShape([2, 1, 4096]) RowMajor",
    "transfers": 1535,
    "receive_fragments": 2047,
    "maximum_endpoint_fragments": 16,
    "transmitted_bytes": 24560,
    "received_bytes": 32752,
    "maximum_outgoing_bytes": 256,
    "maximum_incoming_bytes": 32,
    "source_staging_max_bytes": 256,
    "destination_staging_max_bytes": 32,
    "combined_staging_max_bytes": 256,
    "source_copy_max_bytes": 256,
    "destination_copy_max_bytes": 32,
    "local_copy_runs": 1568,
    "affine_copy_loops": 1313,
    "copy_cycles_floor": 36,
    "replay_address_ranges_valid": true
  },
  {
    "phase": 81,
    "mode": "original",
    "label": "op Some(28) OperatorInputs: TensorShape([2, 1, 1024]) F16 Amp(Left) -> TensorShape([960]) RowMajor",
    "transfers": 960,
    "receive_fragments": 960,
    "maximum_endpoint_fragments": 15,
    "transmitted_bytes": 122880,
    "received_bytes": 122880,
    "maximum_outgoing_bytes": 128,
    "maximum_incoming_bytes": 1920,
    "source_staging_max_bytes": 0,
    "destination_staging_max_bytes": 0,
    "combined_staging_max_bytes": 0,
    "source_copy_max_bytes": 0,
    "destination_copy_max_bytes": 0,
    "local_copy_runs": 0,
    "affine_copy_loops": 0,
    "copy_cycles_floor": 0,
    "replay_address_ranges_valid": true
  },
  {
    "phase": 81,
    "mode": "source",
    "label": "op Some(28) OperatorInputs: TensorShape([2, 1, 1024]) F16 Amp(Left) -> TensorShape([960]) RowMajor",
    "transfers": 960,
    "receive_fragments": 960,
    "maximum_endpoint_fragments": 15,
    "transmitted_bytes": 122880,
    "received_bytes": 122880,
    "maximum_outgoing_bytes": 128,
    "maximum_incoming_bytes": 1920,
    "source_staging_max_bytes": 128,
    "destination_staging_max_bytes": 0,
    "combined_staging_max_bytes": 128,
    "source_copy_max_bytes": 128,
    "destination_copy_max_bytes": 0,
    "local_copy_runs": 960,
    "affine_copy_loops": 960,
    "copy_cycles_floor": 16,
    "replay_address_ranges_valid": true
  },
  {
    "phase": 81,
    "mode": "destination",
    "label": "op Some(28) OperatorInputs: TensorShape([2, 1, 1024]) F16 Amp(Left) -> TensorShape([960]) RowMajor",
    "transfers": 960,
    "receive_fragments": 960,
    "maximum_endpoint_fragments": 15,
    "transmitted_bytes": 122880,
    "received_bytes": 122880,
    "maximum_outgoing_bytes": 128,
    "maximum_incoming_bytes": 1920,
    "source_staging_max_bytes": 0,
    "destination_staging_max_bytes": 1920,
    "combined_staging_max_bytes": 1920,
    "source_copy_max_bytes": 0,
    "destination_copy_max_bytes": 1920,
    "local_copy_runs": 64,
    "affine_copy_loops": 64,
    "copy_cycles_floor": 240,
    "replay_address_ranges_valid": true
  },
  {
    "phase": 81,
    "mode": "both",
    "label": "op Some(28) OperatorInputs: TensorShape([2, 1, 1024]) F16 Amp(Left) -> TensorShape([960]) RowMajor",
    "transfers": 960,
    "receive_fragments": 960,
    "maximum_endpoint_fragments": 15,
    "transmitted_bytes": 122880,
    "received_bytes": 122880,
    "maximum_outgoing_bytes": 128,
    "maximum_incoming_bytes": 1920,
    "source_staging_max_bytes": 128,
    "destination_staging_max_bytes": 1920,
    "combined_staging_max_bytes": 1920,
    "source_copy_max_bytes": 128,
    "destination_copy_max_bytes": 1920,
    "local_copy_runs": 1024,
    "affine_copy_loops": 1024,
    "copy_cycles_floor": 256,
    "replay_address_ranges_valid": true
  },
  {
    "phase": 87,
    "mode": "original",
    "label": "op Some(35) OperatorInputs: TensorShape([16, 1, 128]) F16 Amp(Left) -> TensorShape([16]) RowMajor",
    "transfers": 240,
    "receive_fragments": 240,
    "maximum_endpoint_fragments": 15,
    "transmitted_bytes": 7680,
    "received_bytes": 7680,
    "maximum_outgoing_bytes": 32,
    "maximum_incoming_bytes": 480,
    "source_staging_max_bytes": 0,
    "destination_staging_max_bytes": 0,
    "combined_staging_max_bytes": 0,
    "source_copy_max_bytes": 0,
    "destination_copy_max_bytes": 0,
    "local_copy_runs": 0,
    "affine_copy_loops": 0,
    "copy_cycles_floor": 0,
    "replay_address_ranges_valid": true
  },
  {
    "phase": 87,
    "mode": "source",
    "label": "op Some(35) OperatorInputs: TensorShape([16, 1, 128]) F16 Amp(Left) -> TensorShape([16]) RowMajor",
    "transfers": 240,
    "receive_fragments": 240,
    "maximum_endpoint_fragments": 15,
    "transmitted_bytes": 7680,
    "received_bytes": 7680,
    "maximum_outgoing_bytes": 32,
    "maximum_incoming_bytes": 480,
    "source_staging_max_bytes": 32,
    "destination_staging_max_bytes": 0,
    "combined_staging_max_bytes": 32,
    "source_copy_max_bytes": 32,
    "destination_copy_max_bytes": 0,
    "local_copy_runs": 240,
    "affine_copy_loops": 240,
    "copy_cycles_floor": 4,
    "replay_address_ranges_valid": true
  },
  {
    "phase": 87,
    "mode": "destination",
    "label": "op Some(35) OperatorInputs: TensorShape([16, 1, 128]) F16 Amp(Left) -> TensorShape([16]) RowMajor",
    "transfers": 240,
    "receive_fragments": 240,
    "maximum_endpoint_fragments": 15,
    "transmitted_bytes": 7680,
    "received_bytes": 7680,
    "maximum_outgoing_bytes": 32,
    "maximum_incoming_bytes": 480,
    "source_staging_max_bytes": 0,
    "destination_staging_max_bytes": 480,
    "combined_staging_max_bytes": 480,
    "source_copy_max_bytes": 0,
    "destination_copy_max_bytes": 480,
    "local_copy_runs": 240,
    "affine_copy_loops": 32,
    "copy_cycles_floor": 60,
    "replay_address_ranges_valid": true
  },
  {
    "phase": 87,
    "mode": "both",
    "label": "op Some(35) OperatorInputs: TensorShape([16, 1, 128]) F16 Amp(Left) -> TensorShape([16]) RowMajor",
    "transfers": 240,
    "receive_fragments": 240,
    "maximum_endpoint_fragments": 15,
    "transmitted_bytes": 7680,
    "received_bytes": 7680,
    "maximum_outgoing_bytes": 32,
    "maximum_incoming_bytes": 480,
    "source_staging_max_bytes": 32,
    "destination_staging_max_bytes": 480,
    "combined_staging_max_bytes": 480,
    "source_copy_max_bytes": 32,
    "destination_copy_max_bytes": 480,
    "local_copy_runs": 480,
    "affine_copy_loops": 272,
    "copy_cycles_floor": 64,
    "replay_address_ranges_valid": true
  },
  {
    "phase": 88,
    "mode": "original",
    "label": "op Some(35) LayoutRearrangement: TensorShape([16, 1, 128]) F16 RowMajor -> TensorShape([16, 1, 128]) RowMajor",
    "transfers": 15,
    "receive_fragments": 15,
    "maximum_endpoint_fragments": 15,
    "transmitted_bytes": 3840,
    "received_bytes": 3840,
    "maximum_outgoing_bytes": 256,
    "maximum_incoming_bytes": 3840,
    "source_staging_max_bytes": 0,
    "destination_staging_max_bytes": 0,
    "combined_staging_max_bytes": 0,
    "source_copy_max_bytes": 0,
    "destination_copy_max_bytes": 0,
    "local_copy_runs": 0,
    "affine_copy_loops": 0,
    "copy_cycles_floor": 0,
    "replay_address_ranges_valid": true
  },
  {
    "phase": 88,
    "mode": "source",
    "label": "op Some(35) LayoutRearrangement: TensorShape([16, 1, 128]) F16 RowMajor -> TensorShape([16, 1, 128]) RowMajor",
    "transfers": 15,
    "receive_fragments": 15,
    "maximum_endpoint_fragments": 15,
    "transmitted_bytes": 3840,
    "received_bytes": 3840,
    "maximum_outgoing_bytes": 256,
    "maximum_incoming_bytes": 3840,
    "source_staging_max_bytes": 256,
    "destination_staging_max_bytes": 0,
    "combined_staging_max_bytes": 256,
    "source_copy_max_bytes": 256,
    "destination_copy_max_bytes": 0,
    "local_copy_runs": 15,
    "affine_copy_loops": 15,
    "copy_cycles_floor": 32,
    "replay_address_ranges_valid": true
  },
  {
    "phase": 88,
    "mode": "destination",
    "label": "op Some(35) LayoutRearrangement: TensorShape([16, 1, 128]) F16 RowMajor -> TensorShape([16, 1, 128]) RowMajor",
    "transfers": 15,
    "receive_fragments": 15,
    "maximum_endpoint_fragments": 15,
    "transmitted_bytes": 3840,
    "received_bytes": 3840,
    "maximum_outgoing_bytes": 256,
    "maximum_incoming_bytes": 3840,
    "source_staging_max_bytes": 0,
    "destination_staging_max_bytes": 3840,
    "combined_staging_max_bytes": 3840,
    "source_copy_max_bytes": 0,
    "destination_copy_max_bytes": 3840,
    "local_copy_runs": 1,
    "affine_copy_loops": 1,
    "copy_cycles_floor": 480,
    "replay_address_ranges_valid": true
  },
  {
    "phase": 88,
    "mode": "both",
    "label": "op Some(35) LayoutRearrangement: TensorShape([16, 1, 128]) F16 RowMajor -> TensorShape([16, 1, 128]) RowMajor",
    "transfers": 15,
    "receive_fragments": 15,
    "maximum_endpoint_fragments": 15,
    "transmitted_bytes": 3840,
    "received_bytes": 3840,
    "maximum_outgoing_bytes": 256,
    "maximum_incoming_bytes": 3840,
    "source_staging_max_bytes": 256,
    "destination_staging_max_bytes": 3840,
    "combined_staging_max_bytes": 3840,
    "source_copy_max_bytes": 256,
    "destination_copy_max_bytes": 3840,
    "local_copy_runs": 16,
    "affine_copy_loops": 16,
    "copy_cycles_floor": 512,
    "replay_address_ranges_valid": true
  },
  {
    "phase": 91,
    "mode": "original",
    "label": "op Some(37) OperatorInputs: TensorShape([2, 1, 1024]) F16 Amp(Left) -> TensorShape([960]) RowMajor",
    "transfers": 960,
    "receive_fragments": 960,
    "maximum_endpoint_fragments": 15,
    "transmitted_bytes": 122880,
    "received_bytes": 122880,
    "maximum_outgoing_bytes": 128,
    "maximum_incoming_bytes": 1920,
    "source_staging_max_bytes": 0,
    "destination_staging_max_bytes": 0,
    "combined_staging_max_bytes": 0,
    "source_copy_max_bytes": 0,
    "destination_copy_max_bytes": 0,
    "local_copy_runs": 0,
    "affine_copy_loops": 0,
    "copy_cycles_floor": 0,
    "replay_address_ranges_valid": true
  },
  {
    "phase": 91,
    "mode": "source",
    "label": "op Some(37) OperatorInputs: TensorShape([2, 1, 1024]) F16 Amp(Left) -> TensorShape([960]) RowMajor",
    "transfers": 960,
    "receive_fragments": 960,
    "maximum_endpoint_fragments": 15,
    "transmitted_bytes": 122880,
    "received_bytes": 122880,
    "maximum_outgoing_bytes": 128,
    "maximum_incoming_bytes": 1920,
    "source_staging_max_bytes": 128,
    "destination_staging_max_bytes": 0,
    "combined_staging_max_bytes": 128,
    "source_copy_max_bytes": 128,
    "destination_copy_max_bytes": 0,
    "local_copy_runs": 960,
    "affine_copy_loops": 960,
    "copy_cycles_floor": 16,
    "replay_address_ranges_valid": true
  },
  {
    "phase": 91,
    "mode": "destination",
    "label": "op Some(37) OperatorInputs: TensorShape([2, 1, 1024]) F16 Amp(Left) -> TensorShape([960]) RowMajor",
    "transfers": 960,
    "receive_fragments": 960,
    "maximum_endpoint_fragments": 15,
    "transmitted_bytes": 122880,
    "received_bytes": 122880,
    "maximum_outgoing_bytes": 128,
    "maximum_incoming_bytes": 1920,
    "source_staging_max_bytes": 0,
    "destination_staging_max_bytes": 1920,
    "combined_staging_max_bytes": 1920,
    "source_copy_max_bytes": 0,
    "destination_copy_max_bytes": 1920,
    "local_copy_runs": 64,
    "affine_copy_loops": 64,
    "copy_cycles_floor": 240,
    "replay_address_ranges_valid": true
  },
  {
    "phase": 91,
    "mode": "both",
    "label": "op Some(37) OperatorInputs: TensorShape([2, 1, 1024]) F16 Amp(Left) -> TensorShape([960]) RowMajor",
    "transfers": 960,
    "receive_fragments": 960,
    "maximum_endpoint_fragments": 15,
    "transmitted_bytes": 122880,
    "received_bytes": 122880,
    "maximum_outgoing_bytes": 128,
    "maximum_incoming_bytes": 1920,
    "source_staging_max_bytes": 128,
    "destination_staging_max_bytes": 1920,
    "combined_staging_max_bytes": 1920,
    "source_copy_max_bytes": 128,
    "destination_copy_max_bytes": 1920,
    "local_copy_runs": 1024,
    "affine_copy_loops": 1024,
    "copy_cycles_floor": 256,
    "replay_address_ranges_valid": true
  },
  {
    "phase": 109,
    "mode": "original",
    "label": "op Some(46) OperatorInputs: TensorShape([2, 1, 1024]) F16 Amp(Left) -> TensorShape([960]) RowMajor",
    "transfers": 960,
    "receive_fragments": 960,
    "maximum_endpoint_fragments": 15,
    "transmitted_bytes": 122880,
    "received_bytes": 122880,
    "maximum_outgoing_bytes": 128,
    "maximum_incoming_bytes": 1920,
    "source_staging_max_bytes": 0,
    "destination_staging_max_bytes": 0,
    "combined_staging_max_bytes": 0,
    "source_copy_max_bytes": 0,
    "destination_copy_max_bytes": 0,
    "local_copy_runs": 0,
    "affine_copy_loops": 0,
    "copy_cycles_floor": 0,
    "replay_address_ranges_valid": true
  },
  {
    "phase": 109,
    "mode": "source",
    "label": "op Some(46) OperatorInputs: TensorShape([2, 1, 1024]) F16 Amp(Left) -> TensorShape([960]) RowMajor",
    "transfers": 960,
    "receive_fragments": 960,
    "maximum_endpoint_fragments": 15,
    "transmitted_bytes": 122880,
    "received_bytes": 122880,
    "maximum_outgoing_bytes": 128,
    "maximum_incoming_bytes": 1920,
    "source_staging_max_bytes": 128,
    "destination_staging_max_bytes": 0,
    "combined_staging_max_bytes": 128,
    "source_copy_max_bytes": 128,
    "destination_copy_max_bytes": 0,
    "local_copy_runs": 960,
    "affine_copy_loops": 960,
    "copy_cycles_floor": 16,
    "replay_address_ranges_valid": true
  },
  {
    "phase": 109,
    "mode": "destination",
    "label": "op Some(46) OperatorInputs: TensorShape([2, 1, 1024]) F16 Amp(Left) -> TensorShape([960]) RowMajor",
    "transfers": 960,
    "receive_fragments": 960,
    "maximum_endpoint_fragments": 15,
    "transmitted_bytes": 122880,
    "received_bytes": 122880,
    "maximum_outgoing_bytes": 128,
    "maximum_incoming_bytes": 1920,
    "source_staging_max_bytes": 0,
    "destination_staging_max_bytes": 1920,
    "combined_staging_max_bytes": 1920,
    "source_copy_max_bytes": 0,
    "destination_copy_max_bytes": 1920,
    "local_copy_runs": 64,
    "affine_copy_loops": 64,
    "copy_cycles_floor": 240,
    "replay_address_ranges_valid": true
  },
  {
    "phase": 109,
    "mode": "both",
    "label": "op Some(46) OperatorInputs: TensorShape([2, 1, 1024]) F16 Amp(Left) -> TensorShape([960]) RowMajor",
    "transfers": 960,
    "receive_fragments": 960,
    "maximum_endpoint_fragments": 15,
    "transmitted_bytes": 122880,
    "received_bytes": 122880,
    "maximum_outgoing_bytes": 128,
    "maximum_incoming_bytes": 1920,
    "source_staging_max_bytes": 128,
    "destination_staging_max_bytes": 1920,
    "combined_staging_max_bytes": 1920,
    "source_copy_max_bytes": 128,
    "destination_copy_max_bytes": 1920,
    "local_copy_runs": 1024,
    "affine_copy_loops": 1024,
    "copy_cycles_floor": 256,
    "replay_address_ranges_valid": true
  },
  {
    "phase": 59,
    "mode": "original",
    "label": "op Some(18) OperatorInputs: TensorShape([2, 577, 4096]) F16 Amp(Left) -> TensorShape([3456]) RowMajor",
    "transfers": 5874,
    "receive_fragments": 5874,
    "maximum_endpoint_fragments": 12,
    "transmitted_bytes": 37472896,
    "received_bytes": 37472896,
    "maximum_outgoing_bytes": 27648,
    "maximum_incoming_bytes": 27648,
    "source_staging_max_bytes": 0,
    "destination_staging_max_bytes": 0,
    "combined_staging_max_bytes": 0,
    "source_copy_max_bytes": 0,
    "destination_copy_max_bytes": 0,
    "local_copy_runs": 0,
    "affine_copy_loops": 0,
    "copy_cycles_floor": 0,
    "replay_address_ranges_valid": true
  },
  {
    "phase": 59,
    "mode": "source",
    "label": "op Some(18) OperatorInputs: TensorShape([2, 577, 4096]) F16 Amp(Left) -> TensorShape([3456]) RowMajor",
    "transfers": 5874,
    "receive_fragments": 5874,
    "maximum_endpoint_fragments": 12,
    "transmitted_bytes": 37472896,
    "received_bytes": 37472896,
    "maximum_outgoing_bytes": 27648,
    "maximum_incoming_bytes": 27648,
    "source_staging_max_bytes": 27648,
    "destination_staging_max_bytes": 0,
    "combined_staging_max_bytes": 27648,
    "source_copy_max_bytes": 27648,
    "destination_copy_max_bytes": 0,
    "local_copy_runs": 1452,
    "affine_copy_loops": 1452,
    "copy_cycles_floor": 3456,
    "replay_address_ranges_valid": true
  },
  {
    "phase": 59,
    "mode": "destination",
    "label": "op Some(18) OperatorInputs: TensorShape([2, 577, 4096]) F16 Amp(Left) -> TensorShape([3456]) RowMajor",
    "transfers": 5874,
    "receive_fragments": 5874,
    "maximum_endpoint_fragments": 12,
    "transmitted_bytes": 37472896,
    "received_bytes": 37472896,
    "maximum_outgoing_bytes": 27648,
    "maximum_incoming_bytes": 27648,
    "source_staging_max_bytes": 0,
    "destination_staging_max_bytes": 27648,
    "combined_staging_max_bytes": 27648,
    "source_copy_max_bytes": 0,
    "destination_copy_max_bytes": 27648,
    "local_copy_runs": 3015,
    "affine_copy_loops": 2904,
    "copy_cycles_floor": 3456,
    "replay_address_ranges_valid": true
  },
  {
    "phase": 59,
    "mode": "both",
    "label": "op Some(18) OperatorInputs: TensorShape([2, 577, 4096]) F16 Amp(Left) -> TensorShape([3456]) RowMajor",
    "transfers": 5874,
    "receive_fragments": 5874,
    "maximum_endpoint_fragments": 12,
    "transmitted_bytes": 37472896,
    "received_bytes": 37472896,
    "maximum_outgoing_bytes": 27648,
    "maximum_incoming_bytes": 27648,
    "source_staging_max_bytes": 27648,
    "destination_staging_max_bytes": 27648,
    "combined_staging_max_bytes": 55296,
    "source_copy_max_bytes": 27648,
    "destination_copy_max_bytes": 27648,
    "local_copy_runs": 4467,
    "affine_copy_loops": 4356,
    "copy_cycles_floor": 6912,
    "replay_address_ranges_valid": true
  },
  {
    "phase": 98,
    "mode": "original",
    "label": "op Some(40) LayoutRearrangement: TensorShape([1024, 4096]) F8F143 { scale_exponent: -4 } Amp(TransposedLeft) -> TensorShape([1024, 4096]) BlockMajor(Matrix { row_block: 96, column_block: 16 })",
    "transfers": 8180,
    "receive_fragments": 8180,
    "maximum_endpoint_fragments": 12,
    "transmitted_bytes": 4188160,
    "received_bytes": 4188160,
    "maximum_outgoing_bytes": 3072,
    "maximum_incoming_bytes": 3072,
    "source_staging_max_bytes": 0,
    "destination_staging_max_bytes": 0,
    "combined_staging_max_bytes": 0,
    "source_copy_max_bytes": 0,
    "destination_copy_max_bytes": 0,
    "local_copy_runs": 0,
    "affine_copy_loops": 0,
    "copy_cycles_floor": 0,
    "replay_address_ranges_valid": true
  },
  {
    "phase": 98,
    "mode": "source",
    "label": "op Some(40) LayoutRearrangement: TensorShape([1024, 4096]) F8F143 { scale_exponent: -4 } Amp(TransposedLeft) -> TensorShape([1024, 4096]) BlockMajor(Matrix { row_block: 96, column_block: 16 })",
    "transfers": 5625,
    "receive_fragments": 5625,
    "maximum_endpoint_fragments": 10,
    "transmitted_bytes": 4188160,
    "received_bytes": 4188160,
    "maximum_outgoing_bytes": 3072,
    "maximum_incoming_bytes": 3072,
    "source_staging_max_bytes": 3072,
    "destination_staging_max_bytes": 0,
    "combined_staging_max_bytes": 3072,
    "source_copy_max_bytes": 3072,
    "destination_copy_max_bytes": 0,
    "local_copy_runs": 5713,
    "affine_copy_loops": 3158,
    "copy_cycles_floor": 384,
    "replay_address_ranges_valid": true
  },
  {
    "phase": 98,
    "mode": "destination",
    "label": "op Some(40) LayoutRearrangement: TensorShape([1024, 4096]) F8F143 { scale_exponent: -4 } Amp(TransposedLeft) -> TensorShape([1024, 4096]) BlockMajor(Matrix { row_block: 96, column_block: 16 })",
    "transfers": 5560,
    "receive_fragments": 5560,
    "maximum_endpoint_fragments": 10,
    "transmitted_bytes": 4188160,
    "received_bytes": 4188160,
    "maximum_outgoing_bytes": 3072,
    "maximum_incoming_bytes": 3072,
    "source_staging_max_bytes": 0,
    "destination_staging_max_bytes": 3072,
    "combined_staging_max_bytes": 3072,
    "source_copy_max_bytes": 0,
    "destination_copy_max_bytes": 3072,
    "local_copy_runs": 5843,
    "affine_copy_loops": 3182,
    "copy_cycles_floor": 384,
    "replay_address_ranges_valid": true
  },
  {
    "phase": 98,
    "mode": "both",
    "label": "op Some(40) LayoutRearrangement: TensorShape([1024, 4096]) F8F143 { scale_exponent: -4 } Amp(TransposedLeft) -> TensorShape([1024, 4096]) BlockMajor(Matrix { row_block: 96, column_block: 16 })",
    "transfers": 3823,
    "receive_fragments": 3823,
    "maximum_endpoint_fragments": 10,
    "transmitted_bytes": 4188160,
    "received_bytes": 4188160,
    "maximum_outgoing_bytes": 3072,
    "maximum_incoming_bytes": 3072,
    "source_staging_max_bytes": 3072,
    "destination_staging_max_bytes": 3072,
    "combined_staging_max_bytes": 6144,
    "source_copy_max_bytes": 3072,
    "destination_copy_max_bytes": 3072,
    "local_copy_runs": 9920,
    "affine_copy_loops": 5522,
    "copy_cycles_floor": 768,
    "replay_address_ranges_valid": true
  },
  {
    "phase": 70,
    "mode": "original",
    "label": "op Some(26) OperatorInputs: TensorShape([2, 577, 2048]) F16 Amp(Left) -> TensorShape([1728]) RowMajor",
    "transfers": 2728,
    "receive_fragments": 2728,
    "maximum_endpoint_fragments": 10,
    "transmitted_bytes": 9174016,
    "received_bytes": 9174016,
    "maximum_outgoing_bytes": 27648,
    "maximum_incoming_bytes": 6912,
    "source_staging_max_bytes": 0,
    "destination_staging_max_bytes": 0,
    "combined_staging_max_bytes": 0,
    "source_copy_max_bytes": 0,
    "destination_copy_max_bytes": 0,
    "local_copy_runs": 0,
    "affine_copy_loops": 0,
    "copy_cycles_floor": 0,
    "replay_address_ranges_valid": true
  },
  {
    "phase": 70,
    "mode": "source",
    "label": "op Some(26) OperatorInputs: TensorShape([2, 577, 2048]) F16 Amp(Left) -> TensorShape([1728]) RowMajor",
    "transfers": 2728,
    "receive_fragments": 2728,
    "maximum_endpoint_fragments": 10,
    "transmitted_bytes": 9174016,
    "received_bytes": 9174016,
    "maximum_outgoing_bytes": 27648,
    "maximum_incoming_bytes": 6912,
    "source_staging_max_bytes": 27648,
    "destination_staging_max_bytes": 0,
    "combined_staging_max_bytes": 27648,
    "source_copy_max_bytes": 27648,
    "destination_copy_max_bytes": 0,
    "local_copy_runs": 418,
    "affine_copy_loops": 418,
    "copy_cycles_floor": 3456,
    "replay_address_ranges_valid": true
  },
  {
    "phase": 70,
    "mode": "destination",
    "label": "op Some(26) OperatorInputs: TensorShape([2, 577, 2048]) F16 Amp(Left) -> TensorShape([1728]) RowMajor",
    "transfers": 2728,
    "receive_fragments": 2728,
    "maximum_endpoint_fragments": 10,
    "transmitted_bytes": 9174016,
    "received_bytes": 9174016,
    "maximum_outgoing_bytes": 27648,
    "maximum_incoming_bytes": 6912,
    "source_staging_max_bytes": 0,
    "destination_staging_max_bytes": 6912,
    "combined_staging_max_bytes": 6912,
    "source_copy_max_bytes": 0,
    "destination_copy_max_bytes": 6912,
    "local_copy_runs": 1408,
    "affine_copy_loops": 1408,
    "copy_cycles_floor": 864,
    "replay_address_ranges_valid": true
  },
  {
    "phase": 70,
    "mode": "both",
    "label": "op Some(26) OperatorInputs: TensorShape([2, 577, 2048]) F16 Amp(Left) -> TensorShape([1728]) RowMajor",
    "transfers": 2728,
    "receive_fragments": 2728,
    "maximum_endpoint_fragments": 10,
    "transmitted_bytes": 9174016,
    "received_bytes": 9174016,
    "maximum_outgoing_bytes": 27648,
    "maximum_incoming_bytes": 6912,
    "source_staging_max_bytes": 27648,
    "destination_staging_max_bytes": 6912,
    "combined_staging_max_bytes": 34560,
    "source_copy_max_bytes": 27648,
    "destination_copy_max_bytes": 6912,
    "local_copy_runs": 1826,
    "affine_copy_loops": 1826,
    "copy_cycles_floor": 4320,
    "replay_address_ranges_valid": true
  },
  {
    "phase": 99,
    "mode": "original",
    "label": "op Some(40) OperatorInputs: TensorShape([2, 1, 4096]) F16 Amp(Left) -> TensorShape([1280]) RowMajor",
    "transfers": 1330,
    "receive_fragments": 1330,
    "maximum_endpoint_fragments": 10,
    "transmitted_bytes": 327680,
    "received_bytes": 327680,
    "maximum_outgoing_bytes": 256,
    "maximum_incoming_bytes": 2560,
    "source_staging_max_bytes": 0,
    "destination_staging_max_bytes": 0,
    "combined_staging_max_bytes": 0,
    "source_copy_max_bytes": 0,
    "destination_copy_max_bytes": 0,
    "local_copy_runs": 0,
    "affine_copy_loops": 0,
    "copy_cycles_floor": 0,
    "replay_address_ranges_valid": true
  },
  {
    "phase": 99,
    "mode": "source",
    "label": "op Some(40) OperatorInputs: TensorShape([2, 1, 4096]) F16 Amp(Left) -> TensorShape([1280]) RowMajor",
    "transfers": 1330,
    "receive_fragments": 1330,
    "maximum_endpoint_fragments": 10,
    "transmitted_bytes": 327680,
    "received_bytes": 327680,
    "maximum_outgoing_bytes": 256,
    "maximum_incoming_bytes": 2560,
    "source_staging_max_bytes": 256,
    "destination_staging_max_bytes": 0,
    "combined_staging_max_bytes": 256,
    "source_copy_max_bytes": 256,
    "destination_copy_max_bytes": 0,
    "local_copy_runs": 1330,
    "affine_copy_loops": 1330,
    "copy_cycles_floor": 32,
    "replay_address_ranges_valid": true
  },
  {
    "phase": 99,
    "mode": "destination",
    "label": "op Some(40) OperatorInputs: TensorShape([2, 1, 4096]) F16 Amp(Left) -> TensorShape([1280]) RowMajor",
    "transfers": 1330,
    "receive_fragments": 1330,
    "maximum_endpoint_fragments": 10,
    "transmitted_bytes": 327680,
    "received_bytes": 327680,
    "maximum_outgoing_bytes": 256,
    "maximum_incoming_bytes": 2560,
    "source_staging_max_bytes": 0,
    "destination_staging_max_bytes": 2560,
    "combined_staging_max_bytes": 2560,
    "source_copy_max_bytes": 0,
    "destination_copy_max_bytes": 2560,
    "local_copy_runs": 133,
    "affine_copy_loops": 133,
    "copy_cycles_floor": 320,
    "replay_address_ranges_valid": true
  },
  {
    "phase": 99,
    "mode": "both",
    "label": "op Some(40) OperatorInputs: TensorShape([2, 1, 4096]) F16 Amp(Left) -> TensorShape([1280]) RowMajor",
    "transfers": 1330,
    "receive_fragments": 1330,
    "maximum_endpoint_fragments": 10,
    "transmitted_bytes": 327680,
    "received_bytes": 327680,
    "maximum_outgoing_bytes": 256,
    "maximum_incoming_bytes": 2560,
    "source_staging_max_bytes": 256,
    "destination_staging_max_bytes": 2560,
    "combined_staging_max_bytes": 2560,
    "source_copy_max_bytes": 256,
    "destination_copy_max_bytes": 2560,
    "local_copy_runs": 1463,
    "affine_copy_loops": 1463,
    "copy_cycles_floor": 352,
    "replay_address_ranges_valid": true
  },
  {
    "phase": 71,
    "mode": "original",
    "label": "op Some(26) OperatorInputs: TensorShape([2, 577, 2048]) F16 Amp(Left) -> TensorShape([1728]) RowMajor",
    "transfers": 1408,
    "receive_fragments": 1408,
    "maximum_endpoint_fragments": 9,
    "transmitted_bytes": 4734976,
    "received_bytes": 4734976,
    "maximum_outgoing_bytes": 27648,
    "maximum_incoming_bytes": 3456,
    "source_staging_max_bytes": 0,
    "destination_staging_max_bytes": 0,
    "combined_staging_max_bytes": 0,
    "source_copy_max_bytes": 0,
    "destination_copy_max_bytes": 0,
    "local_copy_runs": 0,
    "affine_copy_loops": 0,
    "copy_cycles_floor": 0,
    "replay_address_ranges_valid": true
  },
  {
    "phase": 71,
    "mode": "source",
    "label": "op Some(26) OperatorInputs: TensorShape([2, 577, 2048]) F16 Amp(Left) -> TensorShape([1728]) RowMajor",
    "transfers": 1408,
    "receive_fragments": 1408,
    "maximum_endpoint_fragments": 9,
    "transmitted_bytes": 4734976,
    "received_bytes": 4734976,
    "maximum_outgoing_bytes": 27648,
    "maximum_incoming_bytes": 3456,
    "source_staging_max_bytes": 27648,
    "destination_staging_max_bytes": 0,
    "combined_staging_max_bytes": 27648,
    "source_copy_max_bytes": 27648,
    "destination_copy_max_bytes": 0,
    "local_copy_runs": 297,
    "affine_copy_loops": 297,
    "copy_cycles_floor": 3456,
    "replay_address_ranges_valid": true
  },
  {
    "phase": 71,
    "mode": "destination",
    "label": "op Some(26) OperatorInputs: TensorShape([2, 577, 2048]) F16 Amp(Left) -> TensorShape([1728]) RowMajor",
    "transfers": 1408,
    "receive_fragments": 1408,
    "maximum_endpoint_fragments": 9,
    "transmitted_bytes": 4734976,
    "received_bytes": 4734976,
    "maximum_outgoing_bytes": 27648,
    "maximum_incoming_bytes": 3456,
    "source_staging_max_bytes": 0,
    "destination_staging_max_bytes": 3456,
    "combined_staging_max_bytes": 3456,
    "source_copy_max_bytes": 0,
    "destination_copy_max_bytes": 3456,
    "local_copy_runs": 1408,
    "affine_copy_loops": 1408,
    "copy_cycles_floor": 432,
    "replay_address_ranges_valid": true
  },
  {
    "phase": 71,
    "mode": "both",
    "label": "op Some(26) OperatorInputs: TensorShape([2, 577, 2048]) F16 Amp(Left) -> TensorShape([1728]) RowMajor",
    "transfers": 1408,
    "receive_fragments": 1408,
    "maximum_endpoint_fragments": 9,
    "transmitted_bytes": 4734976,
    "received_bytes": 4734976,
    "maximum_outgoing_bytes": 27648,
    "maximum_incoming_bytes": 3456,
    "source_staging_max_bytes": 27648,
    "destination_staging_max_bytes": 3456,
    "combined_staging_max_bytes": 31104,
    "source_copy_max_bytes": 27648,
    "destination_copy_max_bytes": 3456,
    "local_copy_runs": 1705,
    "affine_copy_loops": 1705,
    "copy_cycles_floor": 3888,
    "replay_address_ranges_valid": true
  },
  {
    "phase": 72,
    "mode": "original",
    "label": "op Some(26) OperatorInputs: TensorShape([2, 577, 2048]) F16 Amp(Left) -> TensorShape([1728]) RowMajor",
    "transfers": 1408,
    "receive_fragments": 1408,
    "maximum_endpoint_fragments": 9,
    "transmitted_bytes": 4734976,
    "received_bytes": 4734976,
    "maximum_outgoing_bytes": 27648,
    "maximum_incoming_bytes": 3456,
    "source_staging_max_bytes": 0,
    "destination_staging_max_bytes": 0,
    "combined_staging_max_bytes": 0,
    "source_copy_max_bytes": 0,
    "destination_copy_max_bytes": 0,
    "local_copy_runs": 0,
    "affine_copy_loops": 0,
    "copy_cycles_floor": 0,
    "replay_address_ranges_valid": true
  },
  {
    "phase": 72,
    "mode": "source",
    "label": "op Some(26) OperatorInputs: TensorShape([2, 577, 2048]) F16 Amp(Left) -> TensorShape([1728]) RowMajor",
    "transfers": 1408,
    "receive_fragments": 1408,
    "maximum_endpoint_fragments": 9,
    "transmitted_bytes": 4734976,
    "received_bytes": 4734976,
    "maximum_outgoing_bytes": 27648,
    "maximum_incoming_bytes": 3456,
    "source_staging_max_bytes": 27648,
    "destination_staging_max_bytes": 0,
    "combined_staging_max_bytes": 27648,
    "source_copy_max_bytes": 27648,
    "destination_copy_max_bytes": 0,
    "local_copy_runs": 275,
    "affine_copy_loops": 275,
    "copy_cycles_floor": 3456,
    "replay_address_ranges_valid": true
  },
  {
    "phase": 72,
    "mode": "destination",
    "label": "op Some(26) OperatorInputs: TensorShape([2, 577, 2048]) F16 Amp(Left) -> TensorShape([1728]) RowMajor",
    "transfers": 1408,
    "receive_fragments": 1408,
    "maximum_endpoint_fragments": 9,
    "transmitted_bytes": 4734976,
    "received_bytes": 4734976,
    "maximum_outgoing_bytes": 27648,
    "maximum_incoming_bytes": 3456,
    "source_staging_max_bytes": 0,
    "destination_staging_max_bytes": 3456,
    "combined_staging_max_bytes": 3456,
    "source_copy_max_bytes": 0,
    "destination_copy_max_bytes": 3456,
    "local_copy_runs": 1408,
    "affine_copy_loops": 1408,
    "copy_cycles_floor": 432,
    "replay_address_ranges_valid": true
  },
  {
    "phase": 72,
    "mode": "both",
    "label": "op Some(26) OperatorInputs: TensorShape([2, 577, 2048]) F16 Amp(Left) -> TensorShape([1728]) RowMajor",
    "transfers": 1408,
    "receive_fragments": 1408,
    "maximum_endpoint_fragments": 9,
    "transmitted_bytes": 4734976,
    "received_bytes": 4734976,
    "maximum_outgoing_bytes": 27648,
    "maximum_incoming_bytes": 3456,
    "source_staging_max_bytes": 27648,
    "destination_staging_max_bytes": 3456,
    "combined_staging_max_bytes": 31104,
    "source_copy_max_bytes": 27648,
    "destination_copy_max_bytes": 3456,
    "local_copy_runs": 1683,
    "affine_copy_loops": 1683,
    "copy_cycles_floor": 3888,
    "replay_address_ranges_valid": true
  },
  {
    "phase": 73,
    "mode": "original",
    "label": "op Some(26) OperatorInputs: TensorShape([2, 577, 2048]) F16 Amp(Left) -> TensorShape([1728]) RowMajor",
    "transfers": 1408,
    "receive_fragments": 1408,
    "maximum_endpoint_fragments": 9,
    "transmitted_bytes": 4734976,
    "received_bytes": 4734976,
    "maximum_outgoing_bytes": 27648,
    "maximum_incoming_bytes": 3456,
    "source_staging_max_bytes": 0,
    "destination_staging_max_bytes": 0,
    "combined_staging_max_bytes": 0,
    "source_copy_max_bytes": 0,
    "destination_copy_max_bytes": 0,
    "local_copy_runs": 0,
    "affine_copy_loops": 0,
    "copy_cycles_floor": 0,
    "replay_address_ranges_valid": true
  },
  {
    "phase": 73,
    "mode": "source",
    "label": "op Some(26) OperatorInputs: TensorShape([2, 577, 2048]) F16 Amp(Left) -> TensorShape([1728]) RowMajor",
    "transfers": 1408,
    "receive_fragments": 1408,
    "maximum_endpoint_fragments": 9,
    "transmitted_bytes": 4734976,
    "received_bytes": 4734976,
    "maximum_outgoing_bytes": 27648,
    "maximum_incoming_bytes": 3456,
    "source_staging_max_bytes": 27648,
    "destination_staging_max_bytes": 0,
    "combined_staging_max_bytes": 27648,
    "source_copy_max_bytes": 27648,
    "destination_copy_max_bytes": 0,
    "local_copy_runs": 253,
    "affine_copy_loops": 253,
    "copy_cycles_floor": 3456,
    "replay_address_ranges_valid": true
  },
  {
    "phase": 73,
    "mode": "destination",
    "label": "op Some(26) OperatorInputs: TensorShape([2, 577, 2048]) F16 Amp(Left) -> TensorShape([1728]) RowMajor",
    "transfers": 1408,
    "receive_fragments": 1408,
    "maximum_endpoint_fragments": 9,
    "transmitted_bytes": 4734976,
    "received_bytes": 4734976,
    "maximum_outgoing_bytes": 27648,
    "maximum_incoming_bytes": 3456,
    "source_staging_max_bytes": 0,
    "destination_staging_max_bytes": 3456,
    "combined_staging_max_bytes": 3456,
    "source_copy_max_bytes": 0,
    "destination_copy_max_bytes": 3456,
    "local_copy_runs": 1408,
    "affine_copy_loops": 1408,
    "copy_cycles_floor": 432,
    "replay_address_ranges_valid": true
  },
  {
    "phase": 73,
    "mode": "both",
    "label": "op Some(26) OperatorInputs: TensorShape([2, 577, 2048]) F16 Amp(Left) -> TensorShape([1728]) RowMajor",
    "transfers": 1408,
    "receive_fragments": 1408,
    "maximum_endpoint_fragments": 9,
    "transmitted_bytes": 4734976,
    "received_bytes": 4734976,
    "maximum_outgoing_bytes": 27648,
    "maximum_incoming_bytes": 3456,
    "source_staging_max_bytes": 27648,
    "destination_staging_max_bytes": 3456,
    "combined_staging_max_bytes": 31104,
    "source_copy_max_bytes": 27648,
    "destination_copy_max_bytes": 3456,
    "local_copy_runs": 1661,
    "affine_copy_loops": 1661,
    "copy_cycles_floor": 3888,
    "replay_address_ranges_valid": true
  },
  {
    "phase": 74,
    "mode": "original",
    "label": "op Some(26) OperatorInputs: TensorShape([2, 577, 2048]) F16 Amp(Left) -> TensorShape([1728]) RowMajor",
    "transfers": 1408,
    "receive_fragments": 1408,
    "maximum_endpoint_fragments": 9,
    "transmitted_bytes": 4734976,
    "received_bytes": 4734976,
    "maximum_outgoing_bytes": 27648,
    "maximum_incoming_bytes": 3456,
    "source_staging_max_bytes": 0,
    "destination_staging_max_bytes": 0,
    "combined_staging_max_bytes": 0,
    "source_copy_max_bytes": 0,
    "destination_copy_max_bytes": 0,
    "local_copy_runs": 0,
    "affine_copy_loops": 0,
    "copy_cycles_floor": 0,
    "replay_address_ranges_valid": true
  },
  {
    "phase": 74,
    "mode": "source",
    "label": "op Some(26) OperatorInputs: TensorShape([2, 577, 2048]) F16 Amp(Left) -> TensorShape([1728]) RowMajor",
    "transfers": 1408,
    "receive_fragments": 1408,
    "maximum_endpoint_fragments": 9,
    "transmitted_bytes": 4734976,
    "received_bytes": 4734976,
    "maximum_outgoing_bytes": 27648,
    "maximum_incoming_bytes": 3456,
    "source_staging_max_bytes": 27648,
    "destination_staging_max_bytes": 0,
    "combined_staging_max_bytes": 27648,
    "source_copy_max_bytes": 27648,
    "destination_copy_max_bytes": 0,
    "local_copy_runs": 231,
    "affine_copy_loops": 231,
    "copy_cycles_floor": 3456,
    "replay_address_ranges_valid": true
  },
  {
    "phase": 74,
    "mode": "destination",
    "label": "op Some(26) OperatorInputs: TensorShape([2, 577, 2048]) F16 Amp(Left) -> TensorShape([1728]) RowMajor",
    "transfers": 1408,
    "receive_fragments": 1408,
    "maximum_endpoint_fragments": 9,
    "transmitted_bytes": 4734976,
    "received_bytes": 4734976,
    "maximum_outgoing_bytes": 27648,
    "maximum_incoming_bytes": 3456,
    "source_staging_max_bytes": 0,
    "destination_staging_max_bytes": 3456,
    "combined_staging_max_bytes": 3456,
    "source_copy_max_bytes": 0,
    "destination_copy_max_bytes": 3456,
    "local_copy_runs": 1408,
    "affine_copy_loops": 1408,
    "copy_cycles_floor": 432,
    "replay_address_ranges_valid": true
  },
  {
    "phase": 74,
    "mode": "both",
    "label": "op Some(26) OperatorInputs: TensorShape([2, 577, 2048]) F16 Amp(Left) -> TensorShape([1728]) RowMajor",
    "transfers": 1408,
    "receive_fragments": 1408,
    "maximum_endpoint_fragments": 9,
    "transmitted_bytes": 4734976,
    "received_bytes": 4734976,
    "maximum_outgoing_bytes": 27648,
    "maximum_incoming_bytes": 3456,
    "source_staging_max_bytes": 27648,
    "destination_staging_max_bytes": 3456,
    "combined_staging_max_bytes": 31104,
    "source_copy_max_bytes": 27648,
    "destination_copy_max_bytes": 3456,
    "local_copy_runs": 1639,
    "affine_copy_loops": 1639,
    "copy_cycles_floor": 3888,
    "replay_address_ranges_valid": true
  },
  {
    "phase": 75,
    "mode": "original",
    "label": "op Some(26) OperatorInputs: TensorShape([2, 577, 2048]) F16 Amp(Left) -> TensorShape([1728]) RowMajor",
    "transfers": 1408,
    "receive_fragments": 1408,
    "maximum_endpoint_fragments": 9,
    "transmitted_bytes": 4734976,
    "received_bytes": 4734976,
    "maximum_outgoing_bytes": 27648,
    "maximum_incoming_bytes": 3456,
    "source_staging_max_bytes": 0,
    "destination_staging_max_bytes": 0,
    "combined_staging_max_bytes": 0,
    "source_copy_max_bytes": 0,
    "destination_copy_max_bytes": 0,
    "local_copy_runs": 0,
    "affine_copy_loops": 0,
    "copy_cycles_floor": 0,
    "replay_address_ranges_valid": true
  },
  {
    "phase": 75,
    "mode": "source",
    "label": "op Some(26) OperatorInputs: TensorShape([2, 577, 2048]) F16 Amp(Left) -> TensorShape([1728]) RowMajor",
    "transfers": 1408,
    "receive_fragments": 1408,
    "maximum_endpoint_fragments": 9,
    "transmitted_bytes": 4734976,
    "received_bytes": 4734976,
    "maximum_outgoing_bytes": 27648,
    "maximum_incoming_bytes": 3456,
    "source_staging_max_bytes": 27648,
    "destination_staging_max_bytes": 0,
    "combined_staging_max_bytes": 27648,
    "source_copy_max_bytes": 27648,
    "destination_copy_max_bytes": 0,
    "local_copy_runs": 209,
    "affine_copy_loops": 209,
    "copy_cycles_floor": 3456,
    "replay_address_ranges_valid": true
  },
  {
    "phase": 75,
    "mode": "destination",
    "label": "op Some(26) OperatorInputs: TensorShape([2, 577, 2048]) F16 Amp(Left) -> TensorShape([1728]) RowMajor",
    "transfers": 1408,
    "receive_fragments": 1408,
    "maximum_endpoint_fragments": 9,
    "transmitted_bytes": 4734976,
    "received_bytes": 4734976,
    "maximum_outgoing_bytes": 27648,
    "maximum_incoming_bytes": 3456,
    "source_staging_max_bytes": 0,
    "destination_staging_max_bytes": 3456,
    "combined_staging_max_bytes": 3456,
    "source_copy_max_bytes": 0,
    "destination_copy_max_bytes": 3456,
    "local_copy_runs": 1408,
    "affine_copy_loops": 1408,
    "copy_cycles_floor": 432,
    "replay_address_ranges_valid": true
  },
  {
    "phase": 75,
    "mode": "both",
    "label": "op Some(26) OperatorInputs: TensorShape([2, 577, 2048]) F16 Amp(Left) -> TensorShape([1728]) RowMajor",
    "transfers": 1408,
    "receive_fragments": 1408,
    "maximum_endpoint_fragments": 9,
    "transmitted_bytes": 4734976,
    "received_bytes": 4734976,
    "maximum_outgoing_bytes": 27648,
    "maximum_incoming_bytes": 3456,
    "source_staging_max_bytes": 27648,
    "destination_staging_max_bytes": 3456,
    "combined_staging_max_bytes": 31104,
    "source_copy_max_bytes": 27648,
    "destination_copy_max_bytes": 3456,
    "local_copy_runs": 1617,
    "affine_copy_loops": 1617,
    "copy_cycles_floor": 3888,
    "replay_address_ranges_valid": true
  },
  {
    "phase": 76,
    "mode": "original",
    "label": "op Some(26) OperatorInputs: TensorShape([2, 577, 2048]) F16 Amp(Left) -> TensorShape([1728]) RowMajor",
    "transfers": 1408,
    "receive_fragments": 1408,
    "maximum_endpoint_fragments": 9,
    "transmitted_bytes": 4734976,
    "received_bytes": 4734976,
    "maximum_outgoing_bytes": 27648,
    "maximum_incoming_bytes": 3456,
    "source_staging_max_bytes": 0,
    "destination_staging_max_bytes": 0,
    "combined_staging_max_bytes": 0,
    "source_copy_max_bytes": 0,
    "destination_copy_max_bytes": 0,
    "local_copy_runs": 0,
    "affine_copy_loops": 0,
    "copy_cycles_floor": 0,
    "replay_address_ranges_valid": true
  },
  {
    "phase": 76,
    "mode": "source",
    "label": "op Some(26) OperatorInputs: TensorShape([2, 577, 2048]) F16 Amp(Left) -> TensorShape([1728]) RowMajor",
    "transfers": 1408,
    "receive_fragments": 1408,
    "maximum_endpoint_fragments": 9,
    "transmitted_bytes": 4734976,
    "received_bytes": 4734976,
    "maximum_outgoing_bytes": 27648,
    "maximum_incoming_bytes": 3456,
    "source_staging_max_bytes": 27648,
    "destination_staging_max_bytes": 0,
    "combined_staging_max_bytes": 27648,
    "source_copy_max_bytes": 27648,
    "destination_copy_max_bytes": 0,
    "local_copy_runs": 187,
    "affine_copy_loops": 187,
    "copy_cycles_floor": 3456,
    "replay_address_ranges_valid": true
  },
  {
    "phase": 76,
    "mode": "destination",
    "label": "op Some(26) OperatorInputs: TensorShape([2, 577, 2048]) F16 Amp(Left) -> TensorShape([1728]) RowMajor",
    "transfers": 1408,
    "receive_fragments": 1408,
    "maximum_endpoint_fragments": 9,
    "transmitted_bytes": 4734976,
    "received_bytes": 4734976,
    "maximum_outgoing_bytes": 27648,
    "maximum_incoming_bytes": 3456,
    "source_staging_max_bytes": 0,
    "destination_staging_max_bytes": 3456,
    "combined_staging_max_bytes": 3456,
    "source_copy_max_bytes": 0,
    "destination_copy_max_bytes": 3456,
    "local_copy_runs": 1408,
    "affine_copy_loops": 1408,
    "copy_cycles_floor": 432,
    "replay_address_ranges_valid": true
  },
  {
    "phase": 76,
    "mode": "both",
    "label": "op Some(26) OperatorInputs: TensorShape([2, 577, 2048]) F16 Amp(Left) -> TensorShape([1728]) RowMajor",
    "transfers": 1408,
    "receive_fragments": 1408,
    "maximum_endpoint_fragments": 9,
    "transmitted_bytes": 4734976,
    "received_bytes": 4734976,
    "maximum_outgoing_bytes": 27648,
    "maximum_incoming_bytes": 3456,
    "source_staging_max_bytes": 27648,
    "destination_staging_max_bytes": 3456,
    "combined_staging_max_bytes": 31104,
    "source_copy_max_bytes": 27648,
    "destination_copy_max_bytes": 3456,
    "local_copy_runs": 1595,
    "affine_copy_loops": 1595,
    "copy_cycles_floor": 3888,
    "replay_address_ranges_valid": true
  },
  {
    "phase": 19,
    "mode": "original",
    "label": "op Some(4) OperatorInputs: TensorShape([2, 577, 3072]) F16 Amp(Left) -> TensorShape([2560]) RowMajor",
    "transfers": 2790,
    "receive_fragments": 2790,
    "maximum_endpoint_fragments": 8,
    "transmitted_bytes": 13761024,
    "received_bytes": 13761024,
    "maximum_outgoing_bytes": 30720,
    "maximum_incoming_bytes": 10240,
    "source_staging_max_bytes": 0,
    "destination_staging_max_bytes": 0,
    "combined_staging_max_bytes": 0,
    "source_copy_max_bytes": 0,
    "destination_copy_max_bytes": 0,
    "local_copy_runs": 0,
    "affine_copy_loops": 0,
    "copy_cycles_floor": 0,
    "replay_address_ranges_valid": true
  },
  {
    "phase": 19,
    "mode": "source",
    "label": "op Some(4) OperatorInputs: TensorShape([2, 577, 3072]) F16 Amp(Left) -> TensorShape([2560]) RowMajor",
    "transfers": 2790,
    "receive_fragments": 2790,
    "maximum_endpoint_fragments": 8,
    "transmitted_bytes": 13761024,
    "received_bytes": 13761024,
    "maximum_outgoing_bytes": 30720,
    "maximum_incoming_bytes": 10240,
    "source_staging_max_bytes": 30720,
    "destination_staging_max_bytes": 0,
    "combined_staging_max_bytes": 30720,
    "source_copy_max_bytes": 30720,
    "destination_copy_max_bytes": 0,
    "local_copy_runs": 540,
    "affine_copy_loops": 540,
    "copy_cycles_floor": 3840,
    "replay_address_ranges_valid": true
  },
  {
    "phase": 19,
    "mode": "destination",
    "label": "op Some(4) OperatorInputs: TensorShape([2, 577, 3072]) F16 Amp(Left) -> TensorShape([2560]) RowMajor",
    "transfers": 2790,
    "receive_fragments": 2790,
    "maximum_endpoint_fragments": 8,
    "transmitted_bytes": 13761024,
    "received_bytes": 13761024,
    "maximum_outgoing_bytes": 30720,
    "maximum_incoming_bytes": 10240,
    "source_staging_max_bytes": 0,
    "destination_staging_max_bytes": 10240,
    "combined_staging_max_bytes": 10240,
    "source_copy_max_bytes": 0,
    "destination_copy_max_bytes": 10240,
    "local_copy_runs": 1440,
    "affine_copy_loops": 1440,
    "copy_cycles_floor": 1280,
    "replay_address_ranges_valid": true
  },
  {
    "phase": 19,
    "mode": "both",
    "label": "op Some(4) OperatorInputs: TensorShape([2, 577, 3072]) F16 Amp(Left) -> TensorShape([2560]) RowMajor",
    "transfers": 2790,
    "receive_fragments": 2790,
    "maximum_endpoint_fragments": 8,
    "transmitted_bytes": 13761024,
    "received_bytes": 13761024,
    "maximum_outgoing_bytes": 30720,
    "maximum_incoming_bytes": 10240,
    "source_staging_max_bytes": 30720,
    "destination_staging_max_bytes": 10240,
    "combined_staging_max_bytes": 40960,
    "source_copy_max_bytes": 30720,
    "destination_copy_max_bytes": 10240,
    "local_copy_runs": 1980,
    "affine_copy_loops": 1980,
    "copy_cycles_floor": 5120,
    "replay_address_ranges_valid": true
  },
  {
    "phase": 79,
    "mode": "original",
    "label": "op Some(28) LayoutRearrangement: TensorShape([2, 1, 1024]) F16 Amp(Left) -> TensorShape([2, 1, 1024]) RowMajor",
    "transfers": 128,
    "receive_fragments": 8192,
    "maximum_endpoint_fragments": 8,
    "transmitted_bytes": 4096,
    "received_bytes": 262144,
    "maximum_outgoing_bytes": 64,
    "maximum_incoming_bytes": 256,
    "source_staging_max_bytes": 0,
    "destination_staging_max_bytes": 0,
    "combined_staging_max_bytes": 0,
    "source_copy_max_bytes": 0,
    "destination_copy_max_bytes": 0,
    "local_copy_runs": 0,
    "affine_copy_loops": 0,
    "copy_cycles_floor": 0,
    "replay_address_ranges_valid": true
  },
  {
    "phase": 79,
    "mode": "source",
    "label": "op Some(28) LayoutRearrangement: TensorShape([2, 1, 1024]) F16 Amp(Left) -> TensorShape([2, 1, 1024]) RowMajor",
    "transfers": 128,
    "receive_fragments": 8192,
    "maximum_endpoint_fragments": 8,
    "transmitted_bytes": 4096,
    "received_bytes": 262144,
    "maximum_outgoing_bytes": 64,
    "maximum_incoming_bytes": 256,
    "source_staging_max_bytes": 64,
    "destination_staging_max_bytes": 0,
    "combined_staging_max_bytes": 64,
    "source_copy_max_bytes": 64,
    "destination_copy_max_bytes": 0,
    "local_copy_runs": 128,
    "affine_copy_loops": 64,
    "copy_cycles_floor": 8,
    "replay_address_ranges_valid": true
  },
  {
    "phase": 79,
    "mode": "destination",
    "label": "op Some(28) LayoutRearrangement: TensorShape([2, 1, 1024]) F16 Amp(Left) -> TensorShape([2, 1, 1024]) RowMajor",
    "transfers": 128,
    "receive_fragments": 8192,
    "maximum_endpoint_fragments": 8,
    "transmitted_bytes": 4096,
    "received_bytes": 262144,
    "maximum_outgoing_bytes": 64,
    "maximum_incoming_bytes": 256,
    "source_staging_max_bytes": 0,
    "destination_staging_max_bytes": 256,
    "combined_staging_max_bytes": 256,
    "source_copy_max_bytes": 0,
    "destination_copy_max_bytes": 256,
    "local_copy_runs": 8192,
    "affine_copy_loops": 4096,
    "copy_cycles_floor": 32,
    "replay_address_ranges_valid": true
  },
  {
    "phase": 79,
    "mode": "both",
    "label": "op Some(28) LayoutRearrangement: TensorShape([2, 1, 1024]) F16 Amp(Left) -> TensorShape([2, 1, 1024]) RowMajor",
    "transfers": 64,
    "receive_fragments": 4096,
    "maximum_endpoint_fragments": 4,
    "transmitted_bytes": 4096,
    "received_bytes": 262144,
    "maximum_outgoing_bytes": 64,
    "maximum_incoming_bytes": 256,
    "source_staging_max_bytes": 64,
    "destination_staging_max_bytes": 256,
    "combined_staging_max_bytes": 256,
    "source_copy_max_bytes": 64,
    "destination_copy_max_bytes": 256,
    "local_copy_runs": 8320,
    "affine_copy_loops": 4160,
    "copy_cycles_floor": 40,
    "replay_address_ranges_valid": true
  },
  {
    "phase": 20,
    "mode": "original",
    "label": "op Some(4) OperatorInputs: TensorShape([2, 577, 3072]) F16 Amp(Left) -> TensorShape([2560]) RowMajor",
    "transfers": 1440,
    "receive_fragments": 1440,
    "maximum_endpoint_fragments": 7,
    "transmitted_bytes": 7102464,
    "received_bytes": 7102464,
    "maximum_outgoing_bytes": 30720,
    "maximum_incoming_bytes": 5120,
    "source_staging_max_bytes": 0,
    "destination_staging_max_bytes": 0,
    "combined_staging_max_bytes": 0,
    "source_copy_max_bytes": 0,
    "destination_copy_max_bytes": 0,
    "local_copy_runs": 0,
    "affine_copy_loops": 0,
    "copy_cycles_floor": 0,
    "replay_address_ranges_valid": true
  },
  {
    "phase": 20,
    "mode": "source",
    "label": "op Some(4) OperatorInputs: TensorShape([2, 577, 3072]) F16 Amp(Left) -> TensorShape([2560]) RowMajor",
    "transfers": 1440,
    "receive_fragments": 1440,
    "maximum_endpoint_fragments": 7,
    "transmitted_bytes": 7102464,
    "received_bytes": 7102464,
    "maximum_outgoing_bytes": 30720,
    "maximum_incoming_bytes": 5120,
    "source_staging_max_bytes": 30720,
    "destination_staging_max_bytes": 0,
    "combined_staging_max_bytes": 30720,
    "source_copy_max_bytes": 30720,
    "destination_copy_max_bytes": 0,
    "local_copy_runs": 345,
    "affine_copy_loops": 345,
    "copy_cycles_floor": 3840,
    "replay_address_ranges_valid": true
  },
  {
    "phase": 20,
    "mode": "destination",
    "label": "op Some(4) OperatorInputs: TensorShape([2, 577, 3072]) F16 Amp(Left) -> TensorShape([2560]) RowMajor",
    "transfers": 1440,
    "receive_fragments": 1440,
    "maximum_endpoint_fragments": 7,
    "transmitted_bytes": 7102464,
    "received_bytes": 7102464,
    "maximum_outgoing_bytes": 30720,
    "maximum_incoming_bytes": 5120,
    "source_staging_max_bytes": 0,
    "destination_staging_max_bytes": 5120,
    "combined_staging_max_bytes": 5120,
    "source_copy_max_bytes": 0,
    "destination_copy_max_bytes": 5120,
    "local_copy_runs": 1440,
    "affine_copy_loops": 1440,
    "copy_cycles_floor": 640,
    "replay_address_ranges_valid": true
  },
  {
    "phase": 20,
    "mode": "both",
    "label": "op Some(4) OperatorInputs: TensorShape([2, 577, 3072]) F16 Amp(Left) -> TensorShape([2560]) RowMajor",
    "transfers": 1440,
    "receive_fragments": 1440,
    "maximum_endpoint_fragments": 7,
    "transmitted_bytes": 7102464,
    "received_bytes": 7102464,
    "maximum_outgoing_bytes": 30720,
    "maximum_incoming_bytes": 5120,
    "source_staging_max_bytes": 30720,
    "destination_staging_max_bytes": 5120,
    "combined_staging_max_bytes": 35840,
    "source_copy_max_bytes": 30720,
    "destination_copy_max_bytes": 5120,
    "local_copy_runs": 1785,
    "affine_copy_loops": 1785,
    "copy_cycles_floor": 4480,
    "replay_address_ranges_valid": true
  },
  {
    "phase": 21,
    "mode": "original",
    "label": "op Some(4) OperatorInputs: TensorShape([2, 577, 3072]) F16 Amp(Left) -> TensorShape([2560]) RowMajor",
    "transfers": 1440,
    "receive_fragments": 1440,
    "maximum_endpoint_fragments": 7,
    "transmitted_bytes": 7102464,
    "received_bytes": 7102464,
    "maximum_outgoing_bytes": 30720,
    "maximum_incoming_bytes": 5120,
    "source_staging_max_bytes": 0,
    "destination_staging_max_bytes": 0,
    "combined_staging_max_bytes": 0,
    "source_copy_max_bytes": 0,
    "destination_copy_max_bytes": 0,
    "local_copy_runs": 0,
    "affine_copy_loops": 0,
    "copy_cycles_floor": 0,
    "replay_address_ranges_valid": true
  },
  {
    "phase": 21,
    "mode": "source",
    "label": "op Some(4) OperatorInputs: TensorShape([2, 577, 3072]) F16 Amp(Left) -> TensorShape([2560]) RowMajor",
    "transfers": 1440,
    "receive_fragments": 1440,
    "maximum_endpoint_fragments": 7,
    "transmitted_bytes": 7102464,
    "received_bytes": 7102464,
    "maximum_outgoing_bytes": 30720,
    "maximum_incoming_bytes": 5120,
    "source_staging_max_bytes": 30720,
    "destination_staging_max_bytes": 0,
    "combined_staging_max_bytes": 30720,
    "source_copy_max_bytes": 30720,
    "destination_copy_max_bytes": 0,
    "local_copy_runs": 315,
    "affine_copy_loops": 315,
    "copy_cycles_floor": 3840,
    "replay_address_ranges_valid": true
  },
  {
    "phase": 21,
    "mode": "destination",
    "label": "op Some(4) OperatorInputs: TensorShape([2, 577, 3072]) F16 Amp(Left) -> TensorShape([2560]) RowMajor",
    "transfers": 1440,
    "receive_fragments": 1440,
    "maximum_endpoint_fragments": 7,
    "transmitted_bytes": 7102464,
    "received_bytes": 7102464,
    "maximum_outgoing_bytes": 30720,
    "maximum_incoming_bytes": 5120,
    "source_staging_max_bytes": 0,
    "destination_staging_max_bytes": 5120,
    "combined_staging_max_bytes": 5120,
    "source_copy_max_bytes": 0,
    "destination_copy_max_bytes": 5120,
    "local_copy_runs": 1440,
    "affine_copy_loops": 1440,
    "copy_cycles_floor": 640,
    "replay_address_ranges_valid": true
  },
  {
    "phase": 21,
    "mode": "both",
    "label": "op Some(4) OperatorInputs: TensorShape([2, 577, 3072]) F16 Amp(Left) -> TensorShape([2560]) RowMajor",
    "transfers": 1440,
    "receive_fragments": 1440,
    "maximum_endpoint_fragments": 7,
    "transmitted_bytes": 7102464,
    "received_bytes": 7102464,
    "maximum_outgoing_bytes": 30720,
    "maximum_incoming_bytes": 5120,
    "source_staging_max_bytes": 30720,
    "destination_staging_max_bytes": 5120,
    "combined_staging_max_bytes": 35840,
    "source_copy_max_bytes": 30720,
    "destination_copy_max_bytes": 5120,
    "local_copy_runs": 1755,
    "affine_copy_loops": 1755,
    "copy_cycles_floor": 4480,
    "replay_address_ranges_valid": true
  },
  {
    "phase": 22,
    "mode": "original",
    "label": "op Some(4) OperatorInputs: TensorShape([2, 577, 3072]) F16 Amp(Left) -> TensorShape([2560]) RowMajor",
    "transfers": 1440,
    "receive_fragments": 1440,
    "maximum_endpoint_fragments": 7,
    "transmitted_bytes": 7102464,
    "received_bytes": 7102464,
    "maximum_outgoing_bytes": 30720,
    "maximum_incoming_bytes": 5120,
    "source_staging_max_bytes": 0,
    "destination_staging_max_bytes": 0,
    "combined_staging_max_bytes": 0,
    "source_copy_max_bytes": 0,
    "destination_copy_max_bytes": 0,
    "local_copy_runs": 0,
    "affine_copy_loops": 0,
    "copy_cycles_floor": 0,
    "replay_address_ranges_valid": true
  },
  {
    "phase": 22,
    "mode": "source",
    "label": "op Some(4) OperatorInputs: TensorShape([2, 577, 3072]) F16 Amp(Left) -> TensorShape([2560]) RowMajor",
    "transfers": 1440,
    "receive_fragments": 1440,
    "maximum_endpoint_fragments": 7,
    "transmitted_bytes": 7102464,
    "received_bytes": 7102464,
    "maximum_outgoing_bytes": 30720,
    "maximum_incoming_bytes": 5120,
    "source_staging_max_bytes": 30720,
    "destination_staging_max_bytes": 0,
    "combined_staging_max_bytes": 30720,
    "source_copy_max_bytes": 30720,
    "destination_copy_max_bytes": 0,
    "local_copy_runs": 285,
    "affine_copy_loops": 285,
    "copy_cycles_floor": 3840,
    "replay_address_ranges_valid": true
  },
  {
    "phase": 22,
    "mode": "destination",
    "label": "op Some(4) OperatorInputs: TensorShape([2, 577, 3072]) F16 Amp(Left) -> TensorShape([2560]) RowMajor",
    "transfers": 1440,
    "receive_fragments": 1440,
    "maximum_endpoint_fragments": 7,
    "transmitted_bytes": 7102464,
    "received_bytes": 7102464,
    "maximum_outgoing_bytes": 30720,
    "maximum_incoming_bytes": 5120,
    "source_staging_max_bytes": 0,
    "destination_staging_max_bytes": 5120,
    "combined_staging_max_bytes": 5120,
    "source_copy_max_bytes": 0,
    "destination_copy_max_bytes": 5120,
    "local_copy_runs": 1440,
    "affine_copy_loops": 1440,
    "copy_cycles_floor": 640,
    "replay_address_ranges_valid": true
  },
  {
    "phase": 22,
    "mode": "both",
    "label": "op Some(4) OperatorInputs: TensorShape([2, 577, 3072]) F16 Amp(Left) -> TensorShape([2560]) RowMajor",
    "transfers": 1440,
    "receive_fragments": 1440,
    "maximum_endpoint_fragments": 7,
    "transmitted_bytes": 7102464,
    "received_bytes": 7102464,
    "maximum_outgoing_bytes": 30720,
    "maximum_incoming_bytes": 5120,
    "source_staging_max_bytes": 30720,
    "destination_staging_max_bytes": 5120,
    "combined_staging_max_bytes": 35840,
    "source_copy_max_bytes": 30720,
    "destination_copy_max_bytes": 5120,
    "local_copy_runs": 1725,
    "affine_copy_loops": 1725,
    "copy_cycles_floor": 4480,
    "replay_address_ranges_valid": true
  },
  {
    "phase": 23,
    "mode": "original",
    "label": "op Some(4) OperatorInputs: TensorShape([2, 577, 3072]) F16 Amp(Left) -> TensorShape([2560]) RowMajor",
    "transfers": 1440,
    "receive_fragments": 1440,
    "maximum_endpoint_fragments": 7,
    "transmitted_bytes": 7102464,
    "received_bytes": 7102464,
    "maximum_outgoing_bytes": 30720,
    "maximum_incoming_bytes": 5120,
    "source_staging_max_bytes": 0,
    "destination_staging_max_bytes": 0,
    "combined_staging_max_bytes": 0,
    "source_copy_max_bytes": 0,
    "destination_copy_max_bytes": 0,
    "local_copy_runs": 0,
    "affine_copy_loops": 0,
    "copy_cycles_floor": 0,
    "replay_address_ranges_valid": true
  },
  {
    "phase": 23,
    "mode": "source",
    "label": "op Some(4) OperatorInputs: TensorShape([2, 577, 3072]) F16 Amp(Left) -> TensorShape([2560]) RowMajor",
    "transfers": 1440,
    "receive_fragments": 1440,
    "maximum_endpoint_fragments": 7,
    "transmitted_bytes": 7102464,
    "received_bytes": 7102464,
    "maximum_outgoing_bytes": 30720,
    "maximum_incoming_bytes": 5120,
    "source_staging_max_bytes": 30720,
    "destination_staging_max_bytes": 0,
    "combined_staging_max_bytes": 30720,
    "source_copy_max_bytes": 30720,
    "destination_copy_max_bytes": 0,
    "local_copy_runs": 255,
    "affine_copy_loops": 255,
    "copy_cycles_floor": 3840,
    "replay_address_ranges_valid": true
  },
  {
    "phase": 23,
    "mode": "destination",
    "label": "op Some(4) OperatorInputs: TensorShape([2, 577, 3072]) F16 Amp(Left) -> TensorShape([2560]) RowMajor",
    "transfers": 1440,
    "receive_fragments": 1440,
    "maximum_endpoint_fragments": 7,
    "transmitted_bytes": 7102464,
    "received_bytes": 7102464,
    "maximum_outgoing_bytes": 30720,
    "maximum_incoming_bytes": 5120,
    "source_staging_max_bytes": 0,
    "destination_staging_max_bytes": 5120,
    "combined_staging_max_bytes": 5120,
    "source_copy_max_bytes": 0,
    "destination_copy_max_bytes": 5120,
    "local_copy_runs": 1440,
    "affine_copy_loops": 1440,
    "copy_cycles_floor": 640,
    "replay_address_ranges_valid": true
  },
  {
    "phase": 23,
    "mode": "both",
    "label": "op Some(4) OperatorInputs: TensorShape([2, 577, 3072]) F16 Amp(Left) -> TensorShape([2560]) RowMajor",
    "transfers": 1440,
    "receive_fragments": 1440,
    "maximum_endpoint_fragments": 7,
    "transmitted_bytes": 7102464,
    "received_bytes": 7102464,
    "maximum_outgoing_bytes": 30720,
    "maximum_incoming_bytes": 5120,
    "source_staging_max_bytes": 30720,
    "destination_staging_max_bytes": 5120,
    "combined_staging_max_bytes": 35840,
    "source_copy_max_bytes": 30720,
    "destination_copy_max_bytes": 5120,
    "local_copy_runs": 1695,
    "affine_copy_loops": 1695,
    "copy_cycles_floor": 4480,
    "replay_address_ranges_valid": true
  },
  {
    "phase": 12,
    "mode": "original",
    "label": "op Some(1) LayoutRearrangement: TensorShape([1, 577, 1024]) F16 RowMajor -> TensorShape([1, 577, 1024]) RowMajor",
    "transfers": 2031,
    "receive_fragments": 2031,
    "maximum_endpoint_fragments": 6,
    "transmitted_bytes": 1180888,
    "received_bytes": 1180888,
    "maximum_outgoing_bytes": 808,
    "maximum_incoming_bytes": 2048,
    "source_staging_max_bytes": 0,
    "destination_staging_max_bytes": 0,
    "combined_staging_max_bytes": 0,
    "source_copy_max_bytes": 0,
    "destination_copy_max_bytes": 0,
    "local_copy_runs": 0,
    "affine_copy_loops": 0,
    "copy_cycles_floor": 0,
    "replay_address_ranges_valid": true
  },
  {
    "phase": 12,
    "mode": "source",
    "label": "op Some(1) LayoutRearrangement: TensorShape([1, 577, 1024]) F16 RowMajor -> TensorShape([1, 577, 1024]) RowMajor",
    "transfers": 2031,
    "receive_fragments": 2031,
    "maximum_endpoint_fragments": 6,
    "transmitted_bytes": 1180888,
    "received_bytes": 1180888,
    "maximum_outgoing_bytes": 808,
    "maximum_incoming_bytes": 2048,
    "source_staging_max_bytes": 808,
    "destination_staging_max_bytes": 0,
    "combined_staging_max_bytes": 808,
    "source_copy_max_bytes": 808,
    "destination_copy_max_bytes": 0,
    "local_copy_runs": 1897,
    "affine_copy_loops": 1897,
    "copy_cycles_floor": 101,
    "replay_address_ranges_valid": true
  },
  {
    "phase": 12,
    "mode": "destination",
    "label": "op Some(1) LayoutRearrangement: TensorShape([1, 577, 1024]) F16 RowMajor -> TensorShape([1, 577, 1024]) RowMajor",
    "transfers": 2031,
    "receive_fragments": 2031,
    "maximum_endpoint_fragments": 6,
    "transmitted_bytes": 1180888,
    "received_bytes": 1180888,
    "maximum_outgoing_bytes": 808,
    "maximum_incoming_bytes": 2048,
    "source_staging_max_bytes": 0,
    "destination_staging_max_bytes": 2048,
    "combined_staging_max_bytes": 2048,
    "source_copy_max_bytes": 0,
    "destination_copy_max_bytes": 2048,
    "local_copy_runs": 577,
    "affine_copy_loops": 577,
    "copy_cycles_floor": 256,
    "replay_address_ranges_valid": true
  },
  {
    "phase": 12,
    "mode": "both",
    "label": "op Some(1) LayoutRearrangement: TensorShape([1, 577, 1024]) F16 RowMajor -> TensorShape([1, 577, 1024]) RowMajor",
    "transfers": 2031,
    "receive_fragments": 2031,
    "maximum_endpoint_fragments": 6,
    "transmitted_bytes": 1180888,
    "received_bytes": 1180888,
    "maximum_outgoing_bytes": 808,
    "maximum_incoming_bytes": 2048,
    "source_staging_max_bytes": 808,
    "destination_staging_max_bytes": 2048,
    "combined_staging_max_bytes": 2856,
    "source_copy_max_bytes": 808,
    "destination_copy_max_bytes": 2048,
    "local_copy_runs": 2474,
    "affine_copy_loops": 2474,
    "copy_cycles_floor": 357,
    "replay_address_ranges_valid": true
  },
  {
    "phase": 103,
    "mode": "original",
    "label": "op Some(43) LayoutRearrangement: TensorShape([4096, 1024]) F8F143 { scale_exponent: -4 } Amp(TransposedLeft) -> TensorShape([4096, 1024]) BlockMajor(Matrix { row_block: 192, column_block: 16 })",
    "transfers": 1578,
    "receive_fragments": 1578,
    "maximum_endpoint_fragments": 4,
    "transmitted_bytes": 4131840,
    "received_bytes": 4131840,
    "maximum_outgoing_bytes": 3072,
    "maximum_incoming_bytes": 3072,
    "source_staging_max_bytes": 0,
    "destination_staging_max_bytes": 0,
    "combined_staging_max_bytes": 0,
    "source_copy_max_bytes": 0,
    "destination_copy_max_bytes": 0,
    "local_copy_runs": 0,
    "affine_copy_loops": 0,
    "copy_cycles_floor": 0,
    "replay_address_ranges_valid": true
  },
  {
    "phase": 103,
    "mode": "source",
    "label": "op Some(43) LayoutRearrangement: TensorShape([4096, 1024]) F8F143 { scale_exponent: -4 } Amp(TransposedLeft) -> TensorShape([4096, 1024]) BlockMajor(Matrix { row_block: 192, column_block: 16 })",
    "transfers": 1578,
    "receive_fragments": 1578,
    "maximum_endpoint_fragments": 4,
    "transmitted_bytes": 4131840,
    "received_bytes": 4131840,
    "maximum_outgoing_bytes": 3072,
    "maximum_incoming_bytes": 3072,
    "source_staging_max_bytes": 3072,
    "destination_staging_max_bytes": 0,
    "combined_staging_max_bytes": 3072,
    "source_copy_max_bytes": 3072,
    "destination_copy_max_bytes": 0,
    "local_copy_runs": 1389,
    "affine_copy_loops": 1389,
    "copy_cycles_floor": 384,
    "replay_address_ranges_valid": true
  },
  {
    "phase": 103,
    "mode": "destination",
    "label": "op Some(43) LayoutRearrangement: TensorShape([4096, 1024]) F8F143 { scale_exponent: -4 } Amp(TransposedLeft) -> TensorShape([4096, 1024]) BlockMajor(Matrix { row_block: 192, column_block: 16 })",
    "transfers": 1578,
    "receive_fragments": 1578,
    "maximum_endpoint_fragments": 4,
    "transmitted_bytes": 4131840,
    "received_bytes": 4131840,
    "maximum_outgoing_bytes": 3072,
    "maximum_incoming_bytes": 3072,
    "source_staging_max_bytes": 0,
    "destination_staging_max_bytes": 3072,
    "combined_staging_max_bytes": 3072,
    "source_copy_max_bytes": 0,
    "destination_copy_max_bytes": 3072,
    "local_copy_runs": 1389,
    "affine_copy_loops": 1389,
    "copy_cycles_floor": 384,
    "replay_address_ranges_valid": true
  },
  {
    "phase": 103,
    "mode": "both",
    "label": "op Some(43) LayoutRearrangement: TensorShape([4096, 1024]) F8F143 { scale_exponent: -4 } Amp(TransposedLeft) -> TensorShape([4096, 1024]) BlockMajor(Matrix { row_block: 192, column_block: 16 })",
    "transfers": 1578,
    "receive_fragments": 1578,
    "maximum_endpoint_fragments": 4,
    "transmitted_bytes": 4131840,
    "received_bytes": 4131840,
    "maximum_outgoing_bytes": 3072,
    "maximum_incoming_bytes": 3072,
    "source_staging_max_bytes": 3072,
    "destination_staging_max_bytes": 3072,
    "combined_staging_max_bytes": 6144,
    "source_copy_max_bytes": 3072,
    "destination_copy_max_bytes": 3072,
    "local_copy_runs": 2778,
    "affine_copy_loops": 2778,
    "copy_cycles_floor": 768,
    "replay_address_ranges_valid": true
  },
  {
    "phase": 86,
    "mode": "original",
    "label": "op Some(35) OperatorInputs: TensorShape([16, 1, 656]) F16 Amp(Left) -> TensorShape([16, 1, 608]) Amp(Left)",
    "transfers": 160,
    "receive_fragments": 368,
    "maximum_endpoint_fragments": 3,
    "transmitted_bytes": 1264640,
    "received_bytes": 1391104,
    "maximum_outgoing_bytes": 10944,
    "maximum_incoming_bytes": 10336,
    "source_staging_max_bytes": 0,
    "destination_staging_max_bytes": 0,
    "combined_staging_max_bytes": 0,
    "source_copy_max_bytes": 0,
    "destination_copy_max_bytes": 0,
    "local_copy_runs": 0,
    "affine_copy_loops": 0,
    "copy_cycles_floor": 0,
    "replay_address_ranges_valid": true
  },
  {
    "phase": 86,
    "mode": "source",
    "label": "op Some(35) OperatorInputs: TensorShape([16, 1, 656]) F16 Amp(Left) -> TensorShape([16, 1, 608]) Amp(Left)",
    "transfers": 160,
    "receive_fragments": 368,
    "maximum_endpoint_fragments": 3,
    "transmitted_bytes": 1264640,
    "received_bytes": 1391104,
    "maximum_outgoing_bytes": 10944,
    "maximum_incoming_bytes": 10336,
    "source_staging_max_bytes": 10944,
    "destination_staging_max_bytes": 0,
    "combined_staging_max_bytes": 10944,
    "source_copy_max_bytes": 10944,
    "destination_copy_max_bytes": 0,
    "local_copy_runs": 160,
    "affine_copy_loops": 160,
    "copy_cycles_floor": 1368,
    "replay_address_ranges_valid": true
  },
  {
    "phase": 86,
    "mode": "destination",
    "label": "op Some(35) OperatorInputs: TensorShape([16, 1, 656]) F16 Amp(Left) -> TensorShape([16, 1, 608]) Amp(Left)",
    "transfers": 160,
    "receive_fragments": 368,
    "maximum_endpoint_fragments": 3,
    "transmitted_bytes": 1264640,
    "received_bytes": 1391104,
    "maximum_outgoing_bytes": 10944,
    "maximum_incoming_bytes": 10336,
    "source_staging_max_bytes": 0,
    "destination_staging_max_bytes": 10336,
    "combined_staging_max_bytes": 10336,
    "source_copy_max_bytes": 0,
    "destination_copy_max_bytes": 10336,
    "local_copy_runs": 368,
    "affine_copy_loops": 368,
    "copy_cycles_floor": 1292,
    "replay_address_ranges_valid": true
  },
  {
    "phase": 86,
    "mode": "both",
    "label": "op Some(35) OperatorInputs: TensorShape([16, 1, 656]) F16 Amp(Left) -> TensorShape([16, 1, 608]) Amp(Left)",
    "transfers": 160,
    "receive_fragments": 368,
    "maximum_endpoint_fragments": 3,
    "transmitted_bytes": 1264640,
    "received_bytes": 1391104,
    "maximum_outgoing_bytes": 10944,
    "maximum_incoming_bytes": 10336,
    "source_staging_max_bytes": 10944,
    "destination_staging_max_bytes": 10336,
    "combined_staging_max_bytes": 10944,
    "source_copy_max_bytes": 10944,
    "destination_copy_max_bytes": 10336,
    "local_copy_runs": 528,
    "affine_copy_loops": 528,
    "copy_cycles_floor": 2660,
    "replay_address_ranges_valid": true
  },
  {
    "phase": 80,
    "mode": "original",
    "label": "op Some(28) LayoutRearrangement: TensorShape([1024, 1024]) F8F143 { scale_exponent: -4 } Amp(TransposedLeft) -> TensorShape([1024, 1024]) BlockMajor(Matrix { row_block: 64, column_block: 16 })",
    "transfers": 1020,
    "receive_fragments": 1020,
    "maximum_endpoint_fragments": 2,
    "transmitted_bytes": 1044480,
    "received_bytes": 1044480,
    "maximum_outgoing_bytes": 1024,
    "maximum_incoming_bytes": 1024,
    "source_staging_max_bytes": 0,
    "destination_staging_max_bytes": 0,
    "combined_staging_max_bytes": 0,
    "source_copy_max_bytes": 0,
    "destination_copy_max_bytes": 0,
    "local_copy_runs": 0,
    "affine_copy_loops": 0,
    "copy_cycles_floor": 0,
    "replay_address_ranges_valid": true
  },
  {
    "phase": 80,
    "mode": "source",
    "label": "op Some(28) LayoutRearrangement: TensorShape([1024, 1024]) F8F143 { scale_exponent: -4 } Amp(TransposedLeft) -> TensorShape([1024, 1024]) BlockMajor(Matrix { row_block: 64, column_block: 16 })",
    "transfers": 1020,
    "receive_fragments": 1020,
    "maximum_endpoint_fragments": 2,
    "transmitted_bytes": 1044480,
    "received_bytes": 1044480,
    "maximum_outgoing_bytes": 1024,
    "maximum_incoming_bytes": 1024,
    "source_staging_max_bytes": 1024,
    "destination_staging_max_bytes": 0,
    "combined_staging_max_bytes": 1024,
    "source_copy_max_bytes": 1024,
    "destination_copy_max_bytes": 0,
    "local_copy_runs": 1020,
    "affine_copy_loops": 1020,
    "copy_cycles_floor": 128,
    "replay_address_ranges_valid": true
  },
  {
    "phase": 80,
    "mode": "destination",
    "label": "op Some(28) LayoutRearrangement: TensorShape([1024, 1024]) F8F143 { scale_exponent: -4 } Amp(TransposedLeft) -> TensorShape([1024, 1024]) BlockMajor(Matrix { row_block: 64, column_block: 16 })",
    "transfers": 1020,
    "receive_fragments": 1020,
    "maximum_endpoint_fragments": 2,
    "transmitted_bytes": 1044480,
    "received_bytes": 1044480,
    "maximum_outgoing_bytes": 1024,
    "maximum_incoming_bytes": 1024,
    "source_staging_max_bytes": 0,
    "destination_staging_max_bytes": 1024,
    "combined_staging_max_bytes": 1024,
    "source_copy_max_bytes": 0,
    "destination_copy_max_bytes": 1024,
    "local_copy_runs": 1020,
    "affine_copy_loops": 1020,
    "copy_cycles_floor": 128,
    "replay_address_ranges_valid": true
  },
  {
    "phase": 80,
    "mode": "both",
    "label": "op Some(28) LayoutRearrangement: TensorShape([1024, 1024]) F8F143 { scale_exponent: -4 } Amp(TransposedLeft) -> TensorShape([1024, 1024]) BlockMajor(Matrix { row_block: 64, column_block: 16 })",
    "transfers": 1020,
    "receive_fragments": 1020,
    "maximum_endpoint_fragments": 2,
    "transmitted_bytes": 1044480,
    "received_bytes": 1044480,
    "maximum_outgoing_bytes": 1024,
    "maximum_incoming_bytes": 1024,
    "source_staging_max_bytes": 1024,
    "destination_staging_max_bytes": 1024,
    "combined_staging_max_bytes": 2048,
    "source_copy_max_bytes": 1024,
    "destination_copy_max_bytes": 1024,
    "local_copy_runs": 2040,
    "affine_copy_loops": 2040,
    "copy_cycles_floor": 256,
    "replay_address_ranges_valid": true
  },
  {
    "phase": 90,
    "mode": "original",
    "label": "op Some(37) LayoutRearrangement: TensorShape([1024, 1024]) F8F143 { scale_exponent: -4 } Amp(TransposedLeft) -> TensorShape([1024, 1024]) BlockMajor(Matrix { row_block: 64, column_block: 16 })",
    "transfers": 1020,
    "receive_fragments": 1020,
    "maximum_endpoint_fragments": 2,
    "transmitted_bytes": 1044480,
    "received_bytes": 1044480,
    "maximum_outgoing_bytes": 1024,
    "maximum_incoming_bytes": 1024,
    "source_staging_max_bytes": 0,
    "destination_staging_max_bytes": 0,
    "combined_staging_max_bytes": 0,
    "source_copy_max_bytes": 0,
    "destination_copy_max_bytes": 0,
    "local_copy_runs": 0,
    "affine_copy_loops": 0,
    "copy_cycles_floor": 0,
    "replay_address_ranges_valid": true
  },
  {
    "phase": 90,
    "mode": "source",
    "label": "op Some(37) LayoutRearrangement: TensorShape([1024, 1024]) F8F143 { scale_exponent: -4 } Amp(TransposedLeft) -> TensorShape([1024, 1024]) BlockMajor(Matrix { row_block: 64, column_block: 16 })",
    "transfers": 1020,
    "receive_fragments": 1020,
    "maximum_endpoint_fragments": 2,
    "transmitted_bytes": 1044480,
    "received_bytes": 1044480,
    "maximum_outgoing_bytes": 1024,
    "maximum_incoming_bytes": 1024,
    "source_staging_max_bytes": 1024,
    "destination_staging_max_bytes": 0,
    "combined_staging_max_bytes": 1024,
    "source_copy_max_bytes": 1024,
    "destination_copy_max_bytes": 0,
    "local_copy_runs": 1020,
    "affine_copy_loops": 1020,
    "copy_cycles_floor": 128,
    "replay_address_ranges_valid": true
  },
  {
    "phase": 90,
    "mode": "destination",
    "label": "op Some(37) LayoutRearrangement: TensorShape([1024, 1024]) F8F143 { scale_exponent: -4 } Amp(TransposedLeft) -> TensorShape([1024, 1024]) BlockMajor(Matrix { row_block: 64, column_block: 16 })",
    "transfers": 1020,
    "receive_fragments": 1020,
    "maximum_endpoint_fragments": 2,
    "transmitted_bytes": 1044480,
    "received_bytes": 1044480,
    "maximum_outgoing_bytes": 1024,
    "maximum_incoming_bytes": 1024,
    "source_staging_max_bytes": 0,
    "destination_staging_max_bytes": 1024,
    "combined_staging_max_bytes": 1024,
    "source_copy_max_bytes": 0,
    "destination_copy_max_bytes": 1024,
    "local_copy_runs": 1020,
    "affine_copy_loops": 1020,
    "copy_cycles_floor": 128,
    "replay_address_ranges_valid": true
  },
  {
    "phase": 90,
    "mode": "both",
    "label": "op Some(37) LayoutRearrangement: TensorShape([1024, 1024]) F8F143 { scale_exponent: -4 } Amp(TransposedLeft) -> TensorShape([1024, 1024]) BlockMajor(Matrix { row_block: 64, column_block: 16 })",
    "transfers": 1020,
    "receive_fragments": 1020,
    "maximum_endpoint_fragments": 2,
    "transmitted_bytes": 1044480,
    "received_bytes": 1044480,
    "maximum_outgoing_bytes": 1024,
    "maximum_incoming_bytes": 1024,
    "source_staging_max_bytes": 1024,
    "destination_staging_max_bytes": 1024,
    "combined_staging_max_bytes": 2048,
    "source_copy_max_bytes": 1024,
    "destination_copy_max_bytes": 1024,
    "local_copy_runs": 2040,
    "affine_copy_loops": 2040,
    "copy_cycles_floor": 256,
    "replay_address_ranges_valid": true
  },
  {
    "phase": 108,
    "mode": "original",
    "label": "op Some(46) LayoutRearrangement: TensorShape([1024, 1024]) F8F143 { scale_exponent: -4 } Amp(TransposedLeft) -> TensorShape([1024, 1024]) BlockMajor(Matrix { row_block: 64, column_block: 16 })",
    "transfers": 1020,
    "receive_fragments": 1020,
    "maximum_endpoint_fragments": 2,
    "transmitted_bytes": 1044480,
    "received_bytes": 1044480,
    "maximum_outgoing_bytes": 1024,
    "maximum_incoming_bytes": 1024,
    "source_staging_max_bytes": 0,
    "destination_staging_max_bytes": 0,
    "combined_staging_max_bytes": 0,
    "source_copy_max_bytes": 0,
    "destination_copy_max_bytes": 0,
    "local_copy_runs": 0,
    "affine_copy_loops": 0,
    "copy_cycles_floor": 0,
    "replay_address_ranges_valid": true
  },
  {
    "phase": 108,
    "mode": "source",
    "label": "op Some(46) LayoutRearrangement: TensorShape([1024, 1024]) F8F143 { scale_exponent: -4 } Amp(TransposedLeft) -> TensorShape([1024, 1024]) BlockMajor(Matrix { row_block: 64, column_block: 16 })",
    "transfers": 1020,
    "receive_fragments": 1020,
    "maximum_endpoint_fragments": 2,
    "transmitted_bytes": 1044480,
    "received_bytes": 1044480,
    "maximum_outgoing_bytes": 1024,
    "maximum_incoming_bytes": 1024,
    "source_staging_max_bytes": 1024,
    "destination_staging_max_bytes": 0,
    "combined_staging_max_bytes": 1024,
    "source_copy_max_bytes": 1024,
    "destination_copy_max_bytes": 0,
    "local_copy_runs": 1020,
    "affine_copy_loops": 1020,
    "copy_cycles_floor": 128,
    "replay_address_ranges_valid": true
  },
  {
    "phase": 108,
    "mode": "destination",
    "label": "op Some(46) LayoutRearrangement: TensorShape([1024, 1024]) F8F143 { scale_exponent: -4 } Amp(TransposedLeft) -> TensorShape([1024, 1024]) BlockMajor(Matrix { row_block: 64, column_block: 16 })",
    "transfers": 1020,
    "receive_fragments": 1020,
    "maximum_endpoint_fragments": 2,
    "transmitted_bytes": 1044480,
    "received_bytes": 1044480,
    "maximum_outgoing_bytes": 1024,
    "maximum_incoming_bytes": 1024,
    "source_staging_max_bytes": 0,
    "destination_staging_max_bytes": 1024,
    "combined_staging_max_bytes": 1024,
    "source_copy_max_bytes": 0,
    "destination_copy_max_bytes": 1024,
    "local_copy_runs": 1020,
    "affine_copy_loops": 1020,
    "copy_cycles_floor": 128,
    "replay_address_ranges_valid": true
  },
  {
    "phase": 108,
    "mode": "both",
    "label": "op Some(46) LayoutRearrangement: TensorShape([1024, 1024]) F8F143 { scale_exponent: -4 } Amp(TransposedLeft) -> TensorShape([1024, 1024]) BlockMajor(Matrix { row_block: 64, column_block: 16 })",
    "transfers": 1020,
    "receive_fragments": 1020,
    "maximum_endpoint_fragments": 2,
    "transmitted_bytes": 1044480,
    "received_bytes": 1044480,
    "maximum_outgoing_bytes": 1024,
    "maximum_incoming_bytes": 1024,
    "source_staging_max_bytes": 1024,
    "destination_staging_max_bytes": 1024,
    "combined_staging_max_bytes": 2048,
    "source_copy_max_bytes": 1024,
    "destination_copy_max_bytes": 1024,
    "local_copy_runs": 2040,
    "affine_copy_loops": 2040,
    "copy_cycles_floor": 256,
    "replay_address_ranges_valid": true
  },
  {
    "phase": 14,
    "mode": "original",
    "label": "op Some(2) LayoutRearrangement: TensorShape([2, 577, 1024]) F16 RowMajor -> TensorShape([2, 577, 1024]) RowMajor",
    "transfers": 577,
    "receive_fragments": 577,
    "maximum_endpoint_fragments": 1,
    "transmitted_bytes": 1181696,
    "received_bytes": 1181696,
    "maximum_outgoing_bytes": 2048,
    "maximum_incoming_bytes": 2048,
    "source_staging_max_bytes": 0,
    "destination_staging_max_bytes": 0,
    "combined_staging_max_bytes": 0,
    "source_copy_max_bytes": 0,
    "destination_copy_max_bytes": 0,
    "local_copy_runs": 0,
    "affine_copy_loops": 0,
    "copy_cycles_floor": 0,
    "replay_address_ranges_valid": true
  },
  {
    "phase": 14,
    "mode": "source",
    "label": "op Some(2) LayoutRearrangement: TensorShape([2, 577, 1024]) F16 RowMajor -> TensorShape([2, 577, 1024]) RowMajor",
    "transfers": 577,
    "receive_fragments": 577,
    "maximum_endpoint_fragments": 1,
    "transmitted_bytes": 1181696,
    "received_bytes": 1181696,
    "maximum_outgoing_bytes": 2048,
    "maximum_incoming_bytes": 2048,
    "source_staging_max_bytes": 2048,
    "destination_staging_max_bytes": 0,
    "combined_staging_max_bytes": 2048,
    "source_copy_max_bytes": 2048,
    "destination_copy_max_bytes": 0,
    "local_copy_runs": 577,
    "affine_copy_loops": 577,
    "copy_cycles_floor": 256,
    "replay_address_ranges_valid": true
  },
  {
    "phase": 14,
    "mode": "destination",
    "label": "op Some(2) LayoutRearrangement: TensorShape([2, 577, 1024]) F16 RowMajor -> TensorShape([2, 577, 1024]) RowMajor",
    "transfers": 577,
    "receive_fragments": 577,
    "maximum_endpoint_fragments": 1,
    "transmitted_bytes": 1181696,
    "received_bytes": 1181696,
    "maximum_outgoing_bytes": 2048,
    "maximum_incoming_bytes": 2048,
    "source_staging_max_bytes": 0,
    "destination_staging_max_bytes": 2048,
    "combined_staging_max_bytes": 2048,
    "source_copy_max_bytes": 0,
    "destination_copy_max_bytes": 2048,
    "local_copy_runs": 577,
    "affine_copy_loops": 577,
    "copy_cycles_floor": 256,
    "replay_address_ranges_valid": true
  },
  {
    "phase": 14,
    "mode": "both",
    "label": "op Some(2) LayoutRearrangement: TensorShape([2, 577, 1024]) F16 RowMajor -> TensorShape([2, 577, 1024]) RowMajor",
    "transfers": 577,
    "receive_fragments": 577,
    "maximum_endpoint_fragments": 1,
    "transmitted_bytes": 1181696,
    "received_bytes": 1181696,
    "maximum_outgoing_bytes": 2048,
    "maximum_incoming_bytes": 2048,
    "source_staging_max_bytes": 2048,
    "destination_staging_max_bytes": 2048,
    "combined_staging_max_bytes": 2048,
    "source_copy_max_bytes": 2048,
    "destination_copy_max_bytes": 2048,
    "local_copy_runs": 1154,
    "affine_copy_loops": 1154,
    "copy_cycles_floor": 512,
    "replay_address_ranges_valid": true
  },
  {
    "phase": 16,
    "mode": "original",
    "label": "op Some(3) LayoutRearrangement: TensorShape([2, 577, 1024]) F16 RowMajor -> TensorShape([2, 577, 1024]) RowMajor",
    "transfers": 577,
    "receive_fragments": 577,
    "maximum_endpoint_fragments": 1,
    "transmitted_bytes": 1181696,
    "received_bytes": 1181696,
    "maximum_outgoing_bytes": 2048,
    "maximum_incoming_bytes": 2048,
    "source_staging_max_bytes": 0,
    "destination_staging_max_bytes": 0,
    "combined_staging_max_bytes": 0,
    "source_copy_max_bytes": 0,
    "destination_copy_max_bytes": 0,
    "local_copy_runs": 0,
    "affine_copy_loops": 0,
    "copy_cycles_floor": 0,
    "replay_address_ranges_valid": true
  },
  {
    "phase": 16,
    "mode": "source",
    "label": "op Some(3) LayoutRearrangement: TensorShape([2, 577, 1024]) F16 RowMajor -> TensorShape([2, 577, 1024]) RowMajor",
    "transfers": 577,
    "receive_fragments": 577,
    "maximum_endpoint_fragments": 1,
    "transmitted_bytes": 1181696,
    "received_bytes": 1181696,
    "maximum_outgoing_bytes": 2048,
    "maximum_incoming_bytes": 2048,
    "source_staging_max_bytes": 2048,
    "destination_staging_max_bytes": 0,
    "combined_staging_max_bytes": 2048,
    "source_copy_max_bytes": 2048,
    "destination_copy_max_bytes": 0,
    "local_copy_runs": 577,
    "affine_copy_loops": 577,
    "copy_cycles_floor": 256,
    "replay_address_ranges_valid": true
  },
  {
    "phase": 16,
    "mode": "destination",
    "label": "op Some(3) LayoutRearrangement: TensorShape([2, 577, 1024]) F16 RowMajor -> TensorShape([2, 577, 1024]) RowMajor",
    "transfers": 577,
    "receive_fragments": 577,
    "maximum_endpoint_fragments": 1,
    "transmitted_bytes": 1181696,
    "received_bytes": 1181696,
    "maximum_outgoing_bytes": 2048,
    "maximum_incoming_bytes": 2048,
    "source_staging_max_bytes": 0,
    "destination_staging_max_bytes": 2048,
    "combined_staging_max_bytes": 2048,
    "source_copy_max_bytes": 0,
    "destination_copy_max_bytes": 2048,
    "local_copy_runs": 577,
    "affine_copy_loops": 577,
    "copy_cycles_floor": 256,
    "replay_address_ranges_valid": true
  },
  {
    "phase": 16,
    "mode": "both",
    "label": "op Some(3) LayoutRearrangement: TensorShape([2, 577, 1024]) F16 RowMajor -> TensorShape([2, 577, 1024]) RowMajor",
    "transfers": 577,
    "receive_fragments": 577,
    "maximum_endpoint_fragments": 1,
    "transmitted_bytes": 1181696,
    "received_bytes": 1181696,
    "maximum_outgoing_bytes": 2048,
    "maximum_incoming_bytes": 2048,
    "source_staging_max_bytes": 2048,
    "destination_staging_max_bytes": 2048,
    "combined_staging_max_bytes": 2048,
    "source_copy_max_bytes": 2048,
    "destination_copy_max_bytes": 2048,
    "local_copy_runs": 1154,
    "affine_copy_loops": 1154,
    "copy_cycles_floor": 512,
    "replay_address_ranges_valid": true
  },
  {
    "phase": 25,
    "mode": "original",
    "label": "op Some(5) LayoutRearrangement: TensorShape([2, 577, 3072]) F16 RowMajor -> TensorShape([2, 577, 3072]) RowMajor",
    "transfers": 577,
    "receive_fragments": 577,
    "maximum_endpoint_fragments": 1,
    "transmitted_bytes": 3545088,
    "received_bytes": 3545088,
    "maximum_outgoing_bytes": 6144,
    "maximum_incoming_bytes": 6144,
    "source_staging_max_bytes": 0,
    "destination_staging_max_bytes": 0,
    "combined_staging_max_bytes": 0,
    "source_copy_max_bytes": 0,
    "destination_copy_max_bytes": 0,
    "local_copy_runs": 0,
    "affine_copy_loops": 0,
    "copy_cycles_floor": 0,
    "replay_address_ranges_valid": true
  },
  {
    "phase": 25,
    "mode": "source",
    "label": "op Some(5) LayoutRearrangement: TensorShape([2, 577, 3072]) F16 RowMajor -> TensorShape([2, 577, 3072]) RowMajor",
    "transfers": 577,
    "receive_fragments": 577,
    "maximum_endpoint_fragments": 1,
    "transmitted_bytes": 3545088,
    "received_bytes": 3545088,
    "maximum_outgoing_bytes": 6144,
    "maximum_incoming_bytes": 6144,
    "source_staging_max_bytes": 6144,
    "destination_staging_max_bytes": 0,
    "combined_staging_max_bytes": 6144,
    "source_copy_max_bytes": 6144,
    "destination_copy_max_bytes": 0,
    "local_copy_runs": 577,
    "affine_copy_loops": 577,
    "copy_cycles_floor": 768,
    "replay_address_ranges_valid": true
  },
  {
    "phase": 25,
    "mode": "destination",
    "label": "op Some(5) LayoutRearrangement: TensorShape([2, 577, 3072]) F16 RowMajor -> TensorShape([2, 577, 3072]) RowMajor",
    "transfers": 577,
    "receive_fragments": 577,
    "maximum_endpoint_fragments": 1,
    "transmitted_bytes": 3545088,
    "received_bytes": 3545088,
    "maximum_outgoing_bytes": 6144,
    "maximum_incoming_bytes": 6144,
    "source_staging_max_bytes": 0,
    "destination_staging_max_bytes": 6144,
    "combined_staging_max_bytes": 6144,
    "source_copy_max_bytes": 0,
    "destination_copy_max_bytes": 6144,
    "local_copy_runs": 577,
    "affine_copy_loops": 577,
    "copy_cycles_floor": 768,
    "replay_address_ranges_valid": true
  },
  {
    "phase": 25,
    "mode": "both",
    "label": "op Some(5) LayoutRearrangement: TensorShape([2, 577, 3072]) F16 RowMajor -> TensorShape([2, 577, 3072]) RowMajor",
    "transfers": 577,
    "receive_fragments": 577,
    "maximum_endpoint_fragments": 1,
    "transmitted_bytes": 3545088,
    "received_bytes": 3545088,
    "maximum_outgoing_bytes": 6144,
    "maximum_incoming_bytes": 6144,
    "source_staging_max_bytes": 6144,
    "destination_staging_max_bytes": 6144,
    "combined_staging_max_bytes": 6144,
    "source_copy_max_bytes": 6144,
    "destination_copy_max_bytes": 6144,
    "local_copy_runs": 1154,
    "affine_copy_loops": 1154,
    "copy_cycles_floor": 1536,
    "replay_address_ranges_valid": true
  },
  {
    "phase": 55,
    "mode": "original",
    "label": "op Some(15) LayoutRearrangement: TensorShape([2, 577, 1024]) F16 RowMajor -> TensorShape([2, 577, 1024]) RowMajor",
    "transfers": 577,
    "receive_fragments": 577,
    "maximum_endpoint_fragments": 1,
    "transmitted_bytes": 1181696,
    "received_bytes": 1181696,
    "maximum_outgoing_bytes": 2048,
    "maximum_incoming_bytes": 2048,
    "source_staging_max_bytes": 0,
    "destination_staging_max_bytes": 0,
    "combined_staging_max_bytes": 0,
    "source_copy_max_bytes": 0,
    "destination_copy_max_bytes": 0,
    "local_copy_runs": 0,
    "affine_copy_loops": 0,
    "copy_cycles_floor": 0,
    "replay_address_ranges_valid": true
  },
  {
    "phase": 55,
    "mode": "source",
    "label": "op Some(15) LayoutRearrangement: TensorShape([2, 577, 1024]) F16 RowMajor -> TensorShape([2, 577, 1024]) RowMajor",
    "transfers": 577,
    "receive_fragments": 577,
    "maximum_endpoint_fragments": 1,
    "transmitted_bytes": 1181696,
    "received_bytes": 1181696,
    "maximum_outgoing_bytes": 2048,
    "maximum_incoming_bytes": 2048,
    "source_staging_max_bytes": 2048,
    "destination_staging_max_bytes": 0,
    "combined_staging_max_bytes": 2048,
    "source_copy_max_bytes": 2048,
    "destination_copy_max_bytes": 0,
    "local_copy_runs": 577,
    "affine_copy_loops": 577,
    "copy_cycles_floor": 256,
    "replay_address_ranges_valid": true
  },
  {
    "phase": 55,
    "mode": "destination",
    "label": "op Some(15) LayoutRearrangement: TensorShape([2, 577, 1024]) F16 RowMajor -> TensorShape([2, 577, 1024]) RowMajor",
    "transfers": 577,
    "receive_fragments": 577,
    "maximum_endpoint_fragments": 1,
    "transmitted_bytes": 1181696,
    "received_bytes": 1181696,
    "maximum_outgoing_bytes": 2048,
    "maximum_incoming_bytes": 2048,
    "source_staging_max_bytes": 0,
    "destination_staging_max_bytes": 2048,
    "combined_staging_max_bytes": 2048,
    "source_copy_max_bytes": 0,
    "destination_copy_max_bytes": 2048,
    "local_copy_runs": 577,
    "affine_copy_loops": 577,
    "copy_cycles_floor": 256,
    "replay_address_ranges_valid": true
  },
  {
    "phase": 55,
    "mode": "both",
    "label": "op Some(15) LayoutRearrangement: TensorShape([2, 577, 1024]) F16 RowMajor -> TensorShape([2, 577, 1024]) RowMajor",
    "transfers": 577,
    "receive_fragments": 577,
    "maximum_endpoint_fragments": 1,
    "transmitted_bytes": 1181696,
    "received_bytes": 1181696,
    "maximum_outgoing_bytes": 2048,
    "maximum_incoming_bytes": 2048,
    "source_staging_max_bytes": 2048,
    "destination_staging_max_bytes": 2048,
    "combined_staging_max_bytes": 2048,
    "source_copy_max_bytes": 2048,
    "destination_copy_max_bytes": 2048,
    "local_copy_runs": 1154,
    "affine_copy_loops": 1154,
    "copy_cycles_floor": 512,
    "replay_address_ranges_valid": true
  },
  {
    "phase": 57,
    "mode": "original",
    "label": "op Some(17) LayoutRearrangement: TensorShape([2, 577, 1024]) F16 RowMajor -> TensorShape([2, 577, 1024]) RowMajor",
    "transfers": 577,
    "receive_fragments": 577,
    "maximum_endpoint_fragments": 1,
    "transmitted_bytes": 1181696,
    "received_bytes": 1181696,
    "maximum_outgoing_bytes": 2048,
    "maximum_incoming_bytes": 2048,
    "source_staging_max_bytes": 0,
    "destination_staging_max_bytes": 0,
    "combined_staging_max_bytes": 0,
    "source_copy_max_bytes": 0,
    "destination_copy_max_bytes": 0,
    "local_copy_runs": 0,
    "affine_copy_loops": 0,
    "copy_cycles_floor": 0,
    "replay_address_ranges_valid": true
  },
  {
    "phase": 57,
    "mode": "source",
    "label": "op Some(17) LayoutRearrangement: TensorShape([2, 577, 1024]) F16 RowMajor -> TensorShape([2, 577, 1024]) RowMajor",
    "transfers": 577,
    "receive_fragments": 577,
    "maximum_endpoint_fragments": 1,
    "transmitted_bytes": 1181696,
    "received_bytes": 1181696,
    "maximum_outgoing_bytes": 2048,
    "maximum_incoming_bytes": 2048,
    "source_staging_max_bytes": 2048,
    "destination_staging_max_bytes": 0,
    "combined_staging_max_bytes": 2048,
    "source_copy_max_bytes": 2048,
    "destination_copy_max_bytes": 0,
    "local_copy_runs": 577,
    "affine_copy_loops": 577,
    "copy_cycles_floor": 256,
    "replay_address_ranges_valid": true
  },
  {
    "phase": 57,
    "mode": "destination",
    "label": "op Some(17) LayoutRearrangement: TensorShape([2, 577, 1024]) F16 RowMajor -> TensorShape([2, 577, 1024]) RowMajor",
    "transfers": 577,
    "receive_fragments": 577,
    "maximum_endpoint_fragments": 1,
    "transmitted_bytes": 1181696,
    "received_bytes": 1181696,
    "maximum_outgoing_bytes": 2048,
    "maximum_incoming_bytes": 2048,
    "source_staging_max_bytes": 0,
    "destination_staging_max_bytes": 2048,
    "combined_staging_max_bytes": 2048,
    "source_copy_max_bytes": 0,
    "destination_copy_max_bytes": 2048,
    "local_copy_runs": 577,
    "affine_copy_loops": 577,
    "copy_cycles_floor": 256,
    "replay_address_ranges_valid": true
  },
  {
    "phase": 57,
    "mode": "both",
    "label": "op Some(17) LayoutRearrangement: TensorShape([2, 577, 1024]) F16 RowMajor -> TensorShape([2, 577, 1024]) RowMajor",
    "transfers": 577,
    "receive_fragments": 577,
    "maximum_endpoint_fragments": 1,
    "transmitted_bytes": 1181696,
    "received_bytes": 1181696,
    "maximum_outgoing_bytes": 2048,
    "maximum_incoming_bytes": 2048,
    "source_staging_max_bytes": 2048,
    "destination_staging_max_bytes": 2048,
    "combined_staging_max_bytes": 2048,
    "source_copy_max_bytes": 2048,
    "destination_copy_max_bytes": 2048,
    "local_copy_runs": 1154,
    "affine_copy_loops": 1154,
    "copy_cycles_floor": 512,
    "replay_address_ranges_valid": true
  },
  {
    "phase": 61,
    "mode": "original",
    "label": "op Some(19) LayoutRearrangement: TensorShape([2, 577, 4096]) F16 RowMajor -> TensorShape([2, 577, 4096]) RowMajor",
    "transfers": 577,
    "receive_fragments": 577,
    "maximum_endpoint_fragments": 1,
    "transmitted_bytes": 4726784,
    "received_bytes": 4726784,
    "maximum_outgoing_bytes": 8192,
    "maximum_incoming_bytes": 8192,
    "source_staging_max_bytes": 0,
    "destination_staging_max_bytes": 0,
    "combined_staging_max_bytes": 0,
    "source_copy_max_bytes": 0,
    "destination_copy_max_bytes": 0,
    "local_copy_runs": 0,
    "affine_copy_loops": 0,
    "copy_cycles_floor": 0,
    "replay_address_ranges_valid": true
  },
  {
    "phase": 61,
    "mode": "source",
    "label": "op Some(19) LayoutRearrangement: TensorShape([2, 577, 4096]) F16 RowMajor -> TensorShape([2, 577, 4096]) RowMajor",
    "transfers": 577,
    "receive_fragments": 577,
    "maximum_endpoint_fragments": 1,
    "transmitted_bytes": 4726784,
    "received_bytes": 4726784,
    "maximum_outgoing_bytes": 8192,
    "maximum_incoming_bytes": 8192,
    "source_staging_max_bytes": 8192,
    "destination_staging_max_bytes": 0,
    "combined_staging_max_bytes": 8192,
    "source_copy_max_bytes": 8192,
    "destination_copy_max_bytes": 0,
    "local_copy_runs": 577,
    "affine_copy_loops": 577,
    "copy_cycles_floor": 1024,
    "replay_address_ranges_valid": true
  },
  {
    "phase": 61,
    "mode": "destination",
    "label": "op Some(19) LayoutRearrangement: TensorShape([2, 577, 4096]) F16 RowMajor -> TensorShape([2, 577, 4096]) RowMajor",
    "transfers": 577,
    "receive_fragments": 577,
    "maximum_endpoint_fragments": 1,
    "transmitted_bytes": 4726784,
    "received_bytes": 4726784,
    "maximum_outgoing_bytes": 8192,
    "maximum_incoming_bytes": 8192,
    "source_staging_max_bytes": 0,
    "destination_staging_max_bytes": 8192,
    "combined_staging_max_bytes": 8192,
    "source_copy_max_bytes": 0,
    "destination_copy_max_bytes": 8192,
    "local_copy_runs": 577,
    "affine_copy_loops": 577,
    "copy_cycles_floor": 1024,
    "replay_address_ranges_valid": true
  },
  {
    "phase": 61,
    "mode": "both",
    "label": "op Some(19) LayoutRearrangement: TensorShape([2, 577, 4096]) F16 RowMajor -> TensorShape([2, 577, 4096]) RowMajor",
    "transfers": 577,
    "receive_fragments": 577,
    "maximum_endpoint_fragments": 1,
    "transmitted_bytes": 4726784,
    "received_bytes": 4726784,
    "maximum_outgoing_bytes": 8192,
    "maximum_incoming_bytes": 8192,
    "source_staging_max_bytes": 8192,
    "destination_staging_max_bytes": 8192,
    "combined_staging_max_bytes": 8192,
    "source_copy_max_bytes": 8192,
    "destination_copy_max_bytes": 8192,
    "local_copy_runs": 1154,
    "affine_copy_loops": 1154,
    "copy_cycles_floor": 2048,
    "replay_address_ranges_valid": true
  },
  {
    "phase": 66,
    "mode": "original",
    "label": "op Some(22) LayoutRearrangement: TensorShape([2, 577, 1024]) F16 RowMajor -> TensorShape([2, 577, 1024]) RowMajor",
    "transfers": 577,
    "receive_fragments": 577,
    "maximum_endpoint_fragments": 1,
    "transmitted_bytes": 1181696,
    "received_bytes": 1181696,
    "maximum_outgoing_bytes": 2048,
    "maximum_incoming_bytes": 2048,
    "source_staging_max_bytes": 0,
    "destination_staging_max_bytes": 0,
    "combined_staging_max_bytes": 0,
    "source_copy_max_bytes": 0,
    "destination_copy_max_bytes": 0,
    "local_copy_runs": 0,
    "affine_copy_loops": 0,
    "copy_cycles_floor": 0,
    "replay_address_ranges_valid": true
  },
  {
    "phase": 66,
    "mode": "source",
    "label": "op Some(22) LayoutRearrangement: TensorShape([2, 577, 1024]) F16 RowMajor -> TensorShape([2, 577, 1024]) RowMajor",
    "transfers": 577,
    "receive_fragments": 577,
    "maximum_endpoint_fragments": 1,
    "transmitted_bytes": 1181696,
    "received_bytes": 1181696,
    "maximum_outgoing_bytes": 2048,
    "maximum_incoming_bytes": 2048,
    "source_staging_max_bytes": 2048,
    "destination_staging_max_bytes": 0,
    "combined_staging_max_bytes": 2048,
    "source_copy_max_bytes": 2048,
    "destination_copy_max_bytes": 0,
    "local_copy_runs": 577,
    "affine_copy_loops": 577,
    "copy_cycles_floor": 256,
    "replay_address_ranges_valid": true
  },
  {
    "phase": 66,
    "mode": "destination",
    "label": "op Some(22) LayoutRearrangement: TensorShape([2, 577, 1024]) F16 RowMajor -> TensorShape([2, 577, 1024]) RowMajor",
    "transfers": 577,
    "receive_fragments": 577,
    "maximum_endpoint_fragments": 1,
    "transmitted_bytes": 1181696,
    "received_bytes": 1181696,
    "maximum_outgoing_bytes": 2048,
    "maximum_incoming_bytes": 2048,
    "source_staging_max_bytes": 0,
    "destination_staging_max_bytes": 2048,
    "combined_staging_max_bytes": 2048,
    "source_copy_max_bytes": 0,
    "destination_copy_max_bytes": 2048,
    "local_copy_runs": 577,
    "affine_copy_loops": 577,
    "copy_cycles_floor": 256,
    "replay_address_ranges_valid": true
  },
  {
    "phase": 66,
    "mode": "both",
    "label": "op Some(22) LayoutRearrangement: TensorShape([2, 577, 1024]) F16 RowMajor -> TensorShape([2, 577, 1024]) RowMajor",
    "transfers": 577,
    "receive_fragments": 577,
    "maximum_endpoint_fragments": 1,
    "transmitted_bytes": 1181696,
    "received_bytes": 1181696,
    "maximum_outgoing_bytes": 2048,
    "maximum_incoming_bytes": 2048,
    "source_staging_max_bytes": 2048,
    "destination_staging_max_bytes": 2048,
    "combined_staging_max_bytes": 2048,
    "source_copy_max_bytes": 2048,
    "destination_copy_max_bytes": 2048,
    "local_copy_runs": 1154,
    "affine_copy_loops": 1154,
    "copy_cycles_floor": 512,
    "replay_address_ranges_valid": true
  },
  {
    "phase": 68,
    "mode": "original",
    "label": "op Some(25) LayoutRearrangement: TensorShape([2, 577, 1024]) F16 RowMajor -> TensorShape([2, 577, 1024]) RowMajor",
    "transfers": 577,
    "receive_fragments": 577,
    "maximum_endpoint_fragments": 1,
    "transmitted_bytes": 1181696,
    "received_bytes": 1181696,
    "maximum_outgoing_bytes": 2048,
    "maximum_incoming_bytes": 2048,
    "source_staging_max_bytes": 0,
    "destination_staging_max_bytes": 0,
    "combined_staging_max_bytes": 0,
    "source_copy_max_bytes": 0,
    "destination_copy_max_bytes": 0,
    "local_copy_runs": 0,
    "affine_copy_loops": 0,
    "copy_cycles_floor": 0,
    "replay_address_ranges_valid": true
  },
  {
    "phase": 68,
    "mode": "source",
    "label": "op Some(25) LayoutRearrangement: TensorShape([2, 577, 1024]) F16 RowMajor -> TensorShape([2, 577, 1024]) RowMajor",
    "transfers": 577,
    "receive_fragments": 577,
    "maximum_endpoint_fragments": 1,
    "transmitted_bytes": 1181696,
    "received_bytes": 1181696,
    "maximum_outgoing_bytes": 2048,
    "maximum_incoming_bytes": 2048,
    "source_staging_max_bytes": 2048,
    "destination_staging_max_bytes": 0,
    "combined_staging_max_bytes": 2048,
    "source_copy_max_bytes": 2048,
    "destination_copy_max_bytes": 0,
    "local_copy_runs": 577,
    "affine_copy_loops": 577,
    "copy_cycles_floor": 256,
    "replay_address_ranges_valid": true
  },
  {
    "phase": 68,
    "mode": "destination",
    "label": "op Some(25) LayoutRearrangement: TensorShape([2, 577, 1024]) F16 RowMajor -> TensorShape([2, 577, 1024]) RowMajor",
    "transfers": 577,
    "receive_fragments": 577,
    "maximum_endpoint_fragments": 1,
    "transmitted_bytes": 1181696,
    "received_bytes": 1181696,
    "maximum_outgoing_bytes": 2048,
    "maximum_incoming_bytes": 2048,
    "source_staging_max_bytes": 0,
    "destination_staging_max_bytes": 2048,
    "combined_staging_max_bytes": 2048,
    "source_copy_max_bytes": 0,
    "destination_copy_max_bytes": 2048,
    "local_copy_runs": 577,
    "affine_copy_loops": 577,
    "copy_cycles_floor": 256,
    "replay_address_ranges_valid": true
  },
  {
    "phase": 68,
    "mode": "both",
    "label": "op Some(25) LayoutRearrangement: TensorShape([2, 577, 1024]) F16 RowMajor -> TensorShape([2, 577, 1024]) RowMajor",
    "transfers": 577,
    "receive_fragments": 577,
    "maximum_endpoint_fragments": 1,
    "transmitted_bytes": 1181696,
    "received_bytes": 1181696,
    "maximum_outgoing_bytes": 2048,
    "maximum_incoming_bytes": 2048,
    "source_staging_max_bytes": 2048,
    "destination_staging_max_bytes": 2048,
    "combined_staging_max_bytes": 2048,
    "source_copy_max_bytes": 2048,
    "destination_copy_max_bytes": 2048,
    "local_copy_runs": 1154,
    "affine_copy_loops": 1154,
    "copy_cycles_floor": 512,
    "replay_address_ranges_valid": true
  },
  {
    "phase": 78,
    "mode": "original",
    "label": "op Some(27) LayoutRearrangement: TensorShape([2, 577, 2048]) F16 RowMajor -> TensorShape([2, 577, 2048]) RowMajor",
    "transfers": 577,
    "receive_fragments": 577,
    "maximum_endpoint_fragments": 1,
    "transmitted_bytes": 2363392,
    "received_bytes": 2363392,
    "maximum_outgoing_bytes": 4096,
    "maximum_incoming_bytes": 4096,
    "source_staging_max_bytes": 0,
    "destination_staging_max_bytes": 0,
    "combined_staging_max_bytes": 0,
    "source_copy_max_bytes": 0,
    "destination_copy_max_bytes": 0,
    "local_copy_runs": 0,
    "affine_copy_loops": 0,
    "copy_cycles_floor": 0,
    "replay_address_ranges_valid": true
  },
  {
    "phase": 78,
    "mode": "source",
    "label": "op Some(27) LayoutRearrangement: TensorShape([2, 577, 2048]) F16 RowMajor -> TensorShape([2, 577, 2048]) RowMajor",
    "transfers": 577,
    "receive_fragments": 577,
    "maximum_endpoint_fragments": 1,
    "transmitted_bytes": 2363392,
    "received_bytes": 2363392,
    "maximum_outgoing_bytes": 4096,
    "maximum_incoming_bytes": 4096,
    "source_staging_max_bytes": 4096,
    "destination_staging_max_bytes": 0,
    "combined_staging_max_bytes": 4096,
    "source_copy_max_bytes": 4096,
    "destination_copy_max_bytes": 0,
    "local_copy_runs": 577,
    "affine_copy_loops": 577,
    "copy_cycles_floor": 512,
    "replay_address_ranges_valid": true
  },
  {
    "phase": 78,
    "mode": "destination",
    "label": "op Some(27) LayoutRearrangement: TensorShape([2, 577, 2048]) F16 RowMajor -> TensorShape([2, 577, 2048]) RowMajor",
    "transfers": 577,
    "receive_fragments": 577,
    "maximum_endpoint_fragments": 1,
    "transmitted_bytes": 2363392,
    "received_bytes": 2363392,
    "maximum_outgoing_bytes": 4096,
    "maximum_incoming_bytes": 4096,
    "source_staging_max_bytes": 0,
    "destination_staging_max_bytes": 4096,
    "combined_staging_max_bytes": 4096,
    "source_copy_max_bytes": 0,
    "destination_copy_max_bytes": 4096,
    "local_copy_runs": 577,
    "affine_copy_loops": 577,
    "copy_cycles_floor": 512,
    "replay_address_ranges_valid": true
  },
  {
    "phase": 78,
    "mode": "both",
    "label": "op Some(27) LayoutRearrangement: TensorShape([2, 577, 2048]) F16 RowMajor -> TensorShape([2, 577, 2048]) RowMajor",
    "transfers": 577,
    "receive_fragments": 577,
    "maximum_endpoint_fragments": 1,
    "transmitted_bytes": 2363392,
    "received_bytes": 2363392,
    "maximum_outgoing_bytes": 4096,
    "maximum_incoming_bytes": 4096,
    "source_staging_max_bytes": 4096,
    "destination_staging_max_bytes": 4096,
    "combined_staging_max_bytes": 4096,
    "source_copy_max_bytes": 4096,
    "destination_copy_max_bytes": 4096,
    "local_copy_runs": 1154,
    "affine_copy_loops": 1154,
    "copy_cycles_floor": 1024,
    "replay_address_ranges_valid": true
  }
]