None {" copy_u64": 1}
Some((1, 8)) {"operation.3 copy_u64": 1, "operation.3 fill_zero_u64": 1, "operation.3 layer_norm_f16": 729}
Some((1, 9)) {"operation.4 cast_f16_f8": 183, "operation.4 copy_u64": 1, "operation.4 fill_zero_u64": 1944}
Some((1, 10)) {"operation.4 copy_u64": 3}
Some((1, 11)) {"operation.4 gemm_f8_init_large_rows_interleaved_k128_c64_r242_r244": 972, "operation.4 gemm_f8_init_small_rows_interleaved_k128_c64_r242_r244": 486}
Some((1, 12)) {"operation.4 copy_u64": 1458, "operation.4 reduce_sum_f16": 1458, "operation.5 copy_strided_u64": 560, "operation.5 copy_u64": 1566}
Some((1, 13)) {"operation.5 add_f16": 1458, "operation.5 copy_u64": 1458, "operation.5 fill_zero_u64": 1}
Some((1, 14)) {"operation.7 fill_zero_u64": 1, "operation.8 fill_zero_u64": 1, "operation.9 copy_u64": 1}
Some((1, 15)) {"operation.12 fill_zero_u64": 448, "operation.9 copy_strided_u64": 16, "operation.9 copy_u64": 9}
Some((1, 16)) {"operation.12 fill_zero_u64": 7296, "operation.12 rearrange_row_major_to_amp_f16_o1_r0_p64_c16_p16": 80, "operation.12 rearrange_row_major_to_amp_f16_o1_r64_p64_c16_p16": 880, "operation.12 rearrange_row_major_to_amp_f16_o2_r0_p64_c16_p16_b64x16": 80, "operation.12 rearrange_row_major_to_amp_f16_o2_r64_p64_c16_p16_b64x16": 880, "operation.9 copy_strided_u64": 19, "operation.9 rearrange_row_major_to_amp_f16_o0_r7_p7_c72_p80": 112, "operation.9 rearrange_row_major_to_amp_f16_o0_r8_p8_c72_p80": 1360}
Some((1, 17)) {"operation.12 copy_u64": 4, "operation.12 fill_zero_u64": 1472, "operation.12 gemm_f16_init_large_rows_interleaved_k80_c64_r91_r92": 144, "operation.12 gemm_f16_init_large_rows_interleaved_k80_c80_r91_r92": 32, "operation.12 gemm_f16_init_small_rows_interleaved_k80_c64_r91_r92": 1008, "operation.12 gemm_f16_init_small_rows_interleaved_k80_c80_r91_r92": 224}
Some((1, 18)) {"operation.12 attention_softmax_d72_p768_tail_f16": 1472, "operation.12 copy_strided_u64": 48, "operation.12 copy_u64": 1232, "operation.12 fill_zero_u64": 1024}
Some((1, 19)) {"operation.12 copy_strided_u64": 240, "operation.12 gemm_f16_init_large_rows_interleaved_k96_c80_r91_r92": 128, "operation.12 gemm_f16_init_small_rows_interleaved_k96_c80_r91_r92": 896}
Some((1, 20)) {"operation.12 attention_merge_v72_p80_k768_out16_f16": 1472, "operation.12 copy_strided_u64": 1568, "operation.12 copy_u64": 3168, "operation.12 fill_zero_u64": 1472, "operation.12 reduce_sum_f16": 1472, "operation.13 fill_zero_u64": 1}
Some((1, 21)) {"operation.14 cast_f16_f8": 183, "operation.14 copy_u64": 2, "operation.14 fill_zero_u64": 864}
Some((1, 22)) {"operation.14 copy_u64": 3}
Some((1, 23)) {"operation.14 gemm_f8_init_small_rows_interleaved_k96_c48_r146_r146": 1440}
Some((1, 24)) {"operation.14 copy_u64": 80, "operation.14 reduce_sum_f16": 1440, "operation.15 copy_u64": 1728, "operation.15 static_copy_u32": 550}
Some((1, 25)) {"operation.15 add_f16": 1458, "operation.15 copy_u64": 1458}
Some((1, 26)) {"operation.16 add_f16": 1458, "operation.16 copy_u64": 1458}
Some((1, 27)) {"operation.17 copy_u64": 1, "operation.17 fill_zero_u64": 1, "operation.17 layer_norm_f16": 729}
Some((1, 28)) {"operation.18 copy_u64": 1, "operation.18 fill_zero_u64": 162}
Some((1, 29)) {"operation.18 cast_f16_f8": 1458}
Some((1, 30)) {"operation.18 copy_strided_u64": 1}
Some((1, 31)) {"operation.18 gemm_f8_init_large_rows_interleaved_k192_c144_r80_r82": 30, "operation.18 gemm_f8_init_large_rows_interleaved_k192_c160_r80_r82": 780, "operation.18 gemm_f8_init_small_rows_interleaved_k192_c144_r80_r82": 24, "operation.18 gemm_f8_init_small_rows_interleaved_k192_c160_r80_r82": 624}
Some((1, 32)) {"operation.18 copy_u64": 54, "operation.18 reduce_sum_f16": 1458, "operation.19 copy_strided_u64": 453, "operation.19 copy_u64": 1458, "operation.19 static_copy_u32": 270}
Some((1, 33)) {"operation.19 add_f16": 1458, "operation.19 copy_u64": 1458}
Some((1, 34)) {"operation.20 copy_u64": 106, "operation.20 gelu_tanh_approx_f16": 1472, "operation.20 static_copy_u32": 1}
Some((1, 35)) {"operation.21 copy_u64": 1, "operation.21 fill_zero_u64": 288}
Some((1, 36)) {"operation.21 cast_f16_f8": 1458, "operation.21 fill_zero_u64": 72}
Some((1, 37)) {"operation.21 copy_strided_u64": 3}
Some((1, 38)) {"operation.21 gemm_f8_init_large_rows_interleaved_k160_c128_r120_r122": 1215, "operation.21 gemm_f8_init_small_rows_interleaved_k160_c128_r120_r122": 243}
Some((1, 39)) {"operation.21 copy_strided_u64": 4, "operation.21 copy_u64": 204, "operation.21 reduce_sum_f16": 1458, "operation.22 copy_strided_u64": 341, "operation.22 copy_u64": 1679}
Some((1, 40)) {"operation.22 add_f16": 1458, "operation.22 copy_u64": 1458, "operation.23 add_f16": 1458, "operation.23 copy_u64": 1458, "operation.23 fill_zero_u64": 1}