{
 "generated_at": "2026-09-10T16:08:20+0800",
 "label": "ledger_capture_704x1280_v1",
 "steady_gpu_ms_total": 38096.377424,
 "windows_ns": {
  "measured": [
   [
    278487043706,
    305240051817
   ]
  ],
  "steady": [
   [
    292640032241,
    295180904169
   ],
   [
    295180904169,
    297661369715
   ],
   [
    297661369715,
    300149129456
   ],
   [
    300149129456,
    302632831461
   ]
  ]
 },
 "dit_steady": {
  "gpu_ms": 33525.468339,
  "groups": 768,
  "by_unit": [
   {
    "unit": "vllm_omni.ar_diffusion_paged_write_attn: ar_diffusion_paged_write_attn",
    "ms": 18582.07990500001,
    "share": 0.5542675710627903,
    "count": 57600,
    "families": {
     "attention_core": 17555.812551000003,
     "gather_index_cat": 1009.091524,
     "elementwise_reduce": 9.446836,
     "attention_prep": 7.728994
    },
    "top_kernels": [
     {
      "name": "void cutlass::device_kernel<flash::enable_sm90_or_later<flash::FlashAttnFwdSm90<flash::Col",
      "ms": 17555.812551000003
     },
     {
      "name": "void at::native::vectorized_gather_kernel<(int)16, long>(char *, char *, T2 *, int, long, ",
      "ms": 551.392408
     },
     {
      "name": "void at::native::index_elementwise_kernel<(int)128, (int)4, void at::native::gpu_index_ker",
      "ms": 452.83913099999995
     },
     {
      "name": "void flash::prepare_varlen_num_blocks_kernel<(int)1, (bool)0>(int, int, int, const int *, ",
      "ms": 7.728994
     }
    ]
   },
   {
    "unit": "addmm: float_1, modulation, chunk, float_2, layer_norm, scale_msa, add_1, mul",
    "ms": 1825.2225440000002,
    "share": 0.05444286491523009,
    "count": 6400,
    "families": {
     "gemm": 1825.2225440000002
    },
    "top_kernels": [
     {
      "name": "nvjet_sm90_tst_192x208_64x4_2x1_v_bz_coopB_bias_TNT",
      "ms": 1825.2225440000002
     }
    ]
   },
   {
    "unit": "addmm: output_parallel_6, input_1, output_parallel_7",
    "ms": 1673.030364,
    "share": 0.04990326599117998,
    "count": 6400,
    "families": {
     "gemm": 1673.030364
    },
    "top_kernels": [
     {
      "name": "nvjet_sm90_tst_192x192_64x4_1x2_h_bz_coopB_bias_TNN",
      "ms": 1673.030364
     }
    ]
   },
   {
    "unit": "addmm: output_parallel_6",
    "ms": 1579.8352940000002,
    "share": 0.04712343696515005,
    "count": 6400,
    "families": {
     "gemm": 1579.8352940000002
    },
    "top_kernels": [
     {
      "name": "nvjet_sm90_tst_128x240_64x4_2x1_v_bz_coopA_bias_TNT",
      "ms": 1579.8352940000002
     }
    ]
   },
   {
    "unit": "_c10d_functional.all_to_all_single: output_parallel, split, value_7, stack, reshape, transpose, input_t, t",
    "ms": 894.959071,
    "share": 0.026694901379167277,
    "count": 6400,
    "families": {
     "nccl": 894.959071
    },
    "top_kernels": [
     {
      "name": "ncclDevKernel_SendRecv(ncclDevKernelArgsStorage<(unsigned long)4096>)",
      "ms": 894.959071
     }
    ]
   },
   {
    "unit": "_c10d_functional.all_to_all_single: output_5, reshape_3, transpose_3, transpose_4, contiguous_7, tensor_1",
    "ms": 876.1906189999997,
    "share": 0.026135074688299936,
    "count": 6400,
    "families": {
     "nccl": 498.73746,
     "gemm": 377.45315899999997
    },
    "top_kernels": [
     {
      "name": "ncclDevKernel_SendRecv(ncclDevKernelArgsStorage<(unsigned long)4096>)",
      "ms": 498.73746
     },
     {
      "name": "nvjet_sm90_tst_192x192_64x4_1x2_h_bz_coopB_bias_TNN",
      "ms": 377.45315899999997
     }
    ]
   },
   {
    "unit": "addmm: output_parallel_2",
    "ms": 861.1579569999999,
    "share": 0.025686679401230598,
    "count": 6400,
    "families": {
     "nccl": 465.97354199999995,
     "gemm": 395.18441500000006
    },
    "top_kernels": [
     {
      "name": "ncclDevKernel_SendRecv(ncclDevKernelArgsStorage<(unsigned long)4096>)",
      "ms": 465.97354199999995
     },
     {
      "name": "nvjet_sm90_tst_192x192_64x4_1x2_h_bz_coopB_bias_TNN",
      "ms": 395.18441500000006
     }
    ]
   },
   {
    "unit": "addmm: output_parallel_2, input_parallel_1, output_parallel_3",
    "ms": 804.324631,
    "share": 0.023991451002768942,
    "count": 6400,
    "families": {
     "gemm": 804.324631
    },
    "top_kernels": [
     {
      "name": "nvjet_sm90_tst_192x192_64x4_1x2_h_bz_coopB_bias_TNN",
      "ms": 804.324631
     }
    ]
   },
   {
    "unit": "_c10d_functional.all_to_all_single: output_parallel_4, value_float_2, truediv_4, rms_2, truediv_5, float_1",
    "ms": 724.5251430000001,
    "share": 0.021611186327773497,
    "count": 6400,
    "families": {
     "nccl": 724.5251430000001
    },
    "top_kernels": [
     {
      "name": "ncclDevKernel_SendRecv(ncclDevKernelArgsStorage<(unsigned long)4096>)",
      "ms": 724.5251430000001
     }
    ]
   },
   {
    "unit": "addmm: contiguous_12, input_parallel_2, output_parallel_5",
    "ms": 667.680875,
    "share": 0.01991563155057525,
    "count": 6400,
    "families": {
     "gemm": 667.680875
    },
    "top_kernels": [
     {
      "name": "nvjet_sm90_tst_192x192_64x4_1x2_h_bz_coopB_bias_TNN",
      "ms": 667.680875
     }
    ]
   },
   {
    "unit": "addmm: output_parallel_3, camera_features, camera_scale",
    "ms": 665.3169279999998,
    "share": 0.019845119575139243,
    "count": 6400,
    "families": {
     "gemm": 665.3169279999998
    },
    "top_kernels": [
     {
      "name": "nvjet_sm90_tst_192x192_64x4_1x2_h_bz_coopB_bias_TNN",
      "ms": 665.3169279999998
     }
    ]
   },
   {
    "unit": "addmm: contiguous_8, input_parallel, output_parallel_1",
    "ms": 653.668824,
    "share": 0.019497679119357462,
    "count": 6400,
    "families": {
     "gemm": 653.668824
    },
    "top_kernels": [
     {
      "name": "nvjet_sm90_tst_192x192_64x4_1x2_h_bz_coopB_bias_TNN",
      "ms": 653.668824
     }
    ]
   },
   {
    "unit": "addmm: float_8, layer_norm_1, float_9, float_10, to_6, output_parallel_4",
    "ms": 650.8669850000001,
    "share": 0.019414105670907213,
    "count": 6400,
    "families": {
     "gemm": 650.8669850000001
    },
    "top_kernels": [
     {
      "name": "nvjet_sm90_tst_192x192_64x4_1x2_h_bz_coopB_bias_TNN",
      "ms": 650.8669850000001
     }
    ]
   },
   {
    "unit": "addmm: camera_shift",
    "ms": 642.361356,
    "share": 0.01916039917786158,
    "count": 6400,
    "families": {
     "gemm": 642.361356
    },
    "top_kernels": [
     {
      "name": "nvjet_sm90_tst_192x192_64x4_1x2_h_bz_coopB_bias_TNN",
      "ms": 642.361356
     }
    ]
   },
   {
    "unit": "_c10d_functional.all_to_all_single: reshape_10, transpose_8, transpose_9, contiguous_11, tensor_3",
    "ms": 297.137911,
    "share": 0.008863050263621254,
    "count": 6400,
    "families": {
     "nccl": 297.137911
    },
    "top_kernels": [
     {
      "name": "ncclDevKernel_SendRecv(ncclDevKernelArgsStorage<(unsigned long)4096>)",
      "ms": 297.137911
     }
    ]
   },
   {
    "unit": "fa3_fwd._flash_attn_forward: _flash_attn_forward_default, key_4, value_9",
    "ms": 205.287215,
    "share": 0.006123321318711915,
    "count": 6400,
    "families": {
     "attention_core": 205.287215
    },
    "top_kernels": [
     {
      "name": "void cutlass::device_kernel<flash::enable_sm90_or_later<flash::FlashAttnFwdSm90<flash::Col",
      "ms": 205.287215
     }
    ]
   },
   {
    "unit": "triton_red_fused__to_copy_add_mul_native_layer_norm_split_squeeze_view_9: float_1, modulation, chunk, camera_scale, add_11, output_parallel_1, g",
    "ms": 178.94312400000004,
    "share": 0.005337527941163359,
    "count": 6400,
    "families": {
     "triton_fused": 178.94312400000004
    },
    "top_kernels": [
     {
      "name": "triton_red_fused__to_copy_add_mul_native_layer_norm_split_squeeze_view_9",
      "ms": 178.94312400000004
     }
    ]
   },
   {
    "unit": "triton_poi_fused_gelu_view_13: output_parallel_6, input_1",
    "ms": 167.14757899999998,
    "share": 0.00498568960498482,
    "count": 6400,
    "families": {
     "triton_fused": 167.14757899999998
    },
    "top_kernels": [
     {
      "name": "triton_poi_fused_gelu_view_13",
      "ms": 167.14757899999998
     }
    ]
   },
   {
    "unit": "triton_red_fused__to_copy_add_mul_native_layer_norm_split_squeeze_1: float_1, modulation, chunk, float_2, layer_norm, scale_msa, add_1, mul",
    "ms": 154.37514600000003,
    "share": 0.0046047125856379535,
    "count": 6400,
    "families": {
     "triton_fused": 154.37514600000003
    },
    "top_kernels": [
     {
      "name": "triton_red_fused__to_copy_add_mul_native_layer_norm_split_squeeze_1",
      "ms": 154.37514600000003
     }
    ]
   },
   {
    "unit": "triton_red_fused__to_copy_add_mul_native_layer_norm_split_squeeze_view_12: float_1, modulation, chunk, output_parallel_5, hidden_states_2, float_",
    "ms": 147.877554,
    "share": 0.004410901959808712,
    "count": 6400,
    "families": {
     "triton_fused": 147.877554
    },
    "top_kernels": [
     {
      "name": "triton_red_fused__to_copy_add_mul_native_layer_norm_split_squeeze_view_12",
      "ms": 147.877554
     }
    ]
   },
   {
    "unit": "triton_poi_fused_all_to_all_single_clone_split_with_sizes_stack_transpose_view_4: output_parallel, split, value_7, stack, reshape, transpose, input_t, t",
    "ms": 130.831945,
    "share": 0.003902464349701683,
    "count": 6400,
    "families": {
     "triton_fused": 130.831945
    },
    "top_kernels": [
     {
      "name": "triton_poi_fused_all_to_all_single_clone_split_with_sizes_stack_transpose_view_4",
      "ms": 130.831945
     }
    ]
   },
   {
    "unit": "triton_poi_fused__to_copy_add_mul_split_squeeze_view_14: float_1, modulation, chunk, output_parallel_5, hidden_states_2, output",
    "ms": 102.65756100000002,
    "share": 0.003062076865323877,
    "count": 6400,
    "families": {
     "triton_fused": 102.65756100000002
    },
    "top_kernels": [
     {
      "name": "triton_poi_fused__to_copy_add_mul_split_squeeze_view_14",
      "ms": 102.65756100000002
     }
    ]
   },
   {
    "unit": "triton_poi_fused__to_copy_all_to_all_single_clone_div_mul_sqrt_transpose_view_11: output_parallel_4, value_float_2, truediv_4, rms_2, truediv_5, float_1",
    "ms": 79.903088,
    "share": 0.0023833548629967733,
    "count": 6400,
    "families": {
     "triton_fused": 79.903088
    },
    "top_kernels": [
     {
      "name": "triton_poi_fused__to_copy_all_to_all_single_clone_div_mul_sqrt_transpose_view_11",
      "ms": 79.903088
     }
    ]
   },
   {
    "unit": "triton_poi_fused_silu_view_6: output_parallel_2, input_parallel_1",
    "ms": 75.43051399999999,
    "share": 0.0022499466148322845,
    "count": 6400,
    "families": {
     "triton_fused": 75.43051399999999
    },
    "top_kernels": [
     {
      "name": "triton_poi_fused_silu_view_6",
      "ms": 75.43051399999999
     }
    ]
   },
   {
    "unit": "rotary_kernel_0: output_parallel, split, value_float_1, truediv_2, rms_1, truediv_3, fl",
    "ms": 74.51557,
    "share": 0.002222655601601736,
    "count": 6400,
    "families": {
     "triton_fused": 74.51557
    },
    "top_kernels": [
     {
      "name": "rotary_kernel",
      "ms": 74.51557
     }
    ]
   },
   {
    "unit": "rotary_kernel_0: output_parallel, split, value_float, truediv, rms, truediv_1, float_5,",
    "ms": 73.464416,
    "share": 0.002191301707023112,
    "count": 6400,
    "families": {
     "triton_fused": 73.464416
    },
    "top_kernels": [
     {
      "name": "rotary_kernel",
      "ms": 73.464416
     }
    ]
   },
   {
    "unit": "triton_red_fused__to_copy_div_mul_pow_split_with_sizes_sqrt_sum_view_3: output_parallel, split, value_float_1, pow_2, sum_of_squares_1, truedi",
    "ms": 67.446052,
    "share": 0.002011785527289424,
    "count": 6400,
    "families": {
     "triton_fused": 67.446052
    },
    "top_kernels": [
     {
      "name": "triton_red_fused__to_copy_div_mul_pow_split_with_sizes_sqrt_sum_view_3",
      "ms": 67.446052
     }
    ]
   },
   {
    "unit": "triton_poi_fused_add_view_7: output_parallel_3, camera_features",
    "ms": 66.119455,
    "share": 0.001972215699760519,
    "count": 6400,
    "families": {
     "triton_fused": 66.119455
    },
    "top_kernels": [
     {
      "name": "triton_poi_fused_add_view_7",
      "ms": 66.119455
     }
    ]
   },
   {
    "unit": "non-graph: tfm_sub/head/_LingBotHead",
    "ms": 65.45939200000001,
    "share": 0.001952527294714969,
    "count": 1440,
    "families": {
     "elementwise_reduce": 62.818355999999994,
     "gemm": 2.641036
    },
    "top_kernels": [
     {
      "name": "void at::native::elementwise_kernel<(int)128, (int)2, void at::native::gpu_kernel_impl_noc",
      "ms": 29.886645
     },
     {
      "name": "void at::native::unrolled_elementwise_kernel<at::native::direct_copy_kernel_cuda(at::Tenso",
      "ms": 13.731138000000001
     },
     {
      "name": "void at::native::<unnamed>::vectorized_layer_norm_kernel<float, float, (bool)0>(int, T2, c",
      "ms": 12.74339
     },
     {
      "name": "void at::native::vectorized_elementwise_kernel<(int)8, at::native::bfloat16_copy_kernel_cu",
      "ms": 6.024444
     }
    ]
   },
   {
    "unit": "triton_red_fused__to_copy_div_mul_pow_split_with_sizes_sqrt_sum_view_2: output_parallel, split, value_float, pow_1, sum_of_squares, truediv, r",
    "ms": 59.509965,
    "share": 0.0017750673726091507,
    "count": 6400,
    "families": {
     "triton_fused": 59.509965
    },
    "top_kernels": [
     {
      "name": "triton_red_fused__to_copy_div_mul_pow_split_with_sizes_sqrt_sum_view_2",
      "ms": 59.509965
     }
    ]
   },
   {
    "unit": "non-graph: tfm_sub/sp_prepare/_LingBotSPPrepare",
    "ms": 59.25756499999999,
    "share": 0.0017675387678646082,
    "count": 160,
    "families": {
     "elementwise_reduce": 59.25756499999999
    },
    "top_kernels": [
     {
      "name": "void at::native::elementwise_kernel<(int)128, (int)4, void at::native::gpu_kernel_impl_noc",
      "ms": 59.25756499999999
     }
    ]
   },
   {
    "unit": "non-graph: tfm_sub/c2ws_hidden_states_layer1/ColumnParallelLinear",
    "ms": 57.56840900000001,
    "share": 0.0017171545052819138,
    "count": 160,
    "families": {
     "gemm": 57.56840900000001
    },
    "top_kernels": [
     {
      "name": "nvjet_sm90_tst_256x128_64x4_1x2_h_bz_coopA_bias_TNT",
      "ms": 57.56840900000001
     }
    ]
   },
   {
    "unit": "non-graph: tfm_sub/c2ws_hidden_states_layer2/RowParallelLinear",
    "ms": 57.490714,
    "share": 0.0017148370134212668,
    "count": 160,
    "families": {
     "gemm": 57.490714
    },
    "top_kernels": [
     {
      "name": "nvjet_sm90_tst_256x128_64x4_1x2_h_bz_coopA_bias_TNT",
      "ms": 57.490714
     }
    ]
   },
   {
    "unit": "triton_poi_fused_all_to_all_single_clone_transpose_unsqueeze_view_5: output_5, reshape_3, transpose_3, transpose_4, contiguous_7, tensor_1",
    "ms": 48.680484,
    "share": 0.0014520448605745576,
    "count": 6400,
    "families": {
     "triton_fused": 48.680484
    },
    "top_kernels": [
     {
      "name": "triton_poi_fused_all_to_all_single_clone_transpose_unsqueeze_view_5",
      "ms": 48.680484
     }
    ]
   },
   {
    "unit": "triton_poi_fused_all_to_all_single_clone_transpose_unsqueeze_view_5: reshape_10, transpose_8, transpose_9, contiguous_11, tensor_3",
    "ms": 48.468555,
    "share": 0.0014457234276311895,
    "count": 6400,
    "families": {
     "triton_fused": 48.468555
    },
    "top_kernels": [
     {
      "name": "triton_poi_fused_all_to_all_single_clone_transpose_unsqueeze_view_5",
      "ms": 48.468555
     }
    ]
   },
   {
    "unit": "triton_poi_fused_clone_transpose_view_8: contiguous_8",
    "ms": 43.368869000000004,
    "share": 0.001293609639139604,
    "count": 6400,
    "families": {
     "triton_fused": 43.368869000000004
    },
    "top_kernels": [
     {
      "name": "triton_poi_fused_clone_transpose_view_8",
      "ms": 43.368869000000004
     }
    ]
   },
   {
    "unit": "triton_poi_fused_clone_transpose_view_8: contiguous_12",
    "ms": 40.367169999999994,
    "share": 0.0012040747527169093,
    "count": 6400,
    "families": {
     "triton_fused": 40.367169999999994
    },
    "top_kernels": [
     {
      "name": "triton_poi_fused_clone_transpose_view_8",
      "ms": 40.367169999999994
     }
    ]
   },
   {
    "unit": "triton_red_fused__to_copy_pow_sum_view_10: output_parallel_4, value_float_2, pow_3, sum_of_squares_2",
    "ms": 30.626336,
    "share": 0.000913524479071111,
    "count": 6400,
    "families": {
     "triton_fused": 30.626336
    },
    "top_kernels": [
     {
      "name": "triton_red_fused__to_copy_pow_sum_view_10",
      "ms": 30.626336
     }
    ]
   },
   {
    "unit": "non-graph: ncclAllGather",
    "ms": 24.741336,
    "share": 0.0007379862899996697,
    "count": 160,
    "families": {
     "nccl": 24.741336
    },
    "top_kernels": [
     {
      "name": "ncclDevKernel_AllGather_RING_LL(ncclDevKernelArgsStorage<(unsigned long)4096>)",
      "ms": 24.741336
     }
    ]
   },
   {
    "unit": "non-graph: tfm_sub/patch_embedding_wancamctrl/_LingBotCameraPatchEmbedding",
    "ms": 24.193209,
    "share": 0.0007216367197428878,
    "count": 320,
    "families": {
     "gemm": 18.686588999999998,
     "elementwise_reduce": 5.50662
    },
    "top_kernels": [
     {
      "name": "nvjet_sm90_tst_256x128_64x4_1x2_h_bz_coopA_bias_TNT",
      "ms": 18.686588999999998
     },
     {
      "name": "void at::native::elementwise_kernel<(int)128, (int)4, void at::native::gpu_kernel_impl_noc",
      "ms": 5.50662
     }
    ]
   },
   {
    "unit": "non-graph: no context",
    "ms": 14.080965,
    "share": 0.000420008002800059,
    "count": 5864,
    "families": {
     "elementwise_reduce": 7.196713999999998,
     "gather_index_cat": 4.095344,
     "conv": 2.737323,
     "gemm": 0.051584000000000005
    },
    "top_kernels": [
     {
      "name": "void at::native::elementwise_kernel<(int)128, (int)2, void at::native::gpu_kernel_impl_noc",
      "ms": 3.215056
     },
     {
      "name": "sm80_xmma_gemm_f32f32_f32f32_f32_tn_n_tilesize128x32x8_stage3_warpsize2x2x1_ffma_aligna4_a",
      "ms": 2.658315
     },
     {
      "name": "void at::native::<unnamed>::CatArrayBatchedCopy<at::native::<unnamed>::OpaqueType<(unsigne",
      "ms": 2.028394
     },
     {
      "name": "void at::native::<unnamed>::CatArrayBatchedCopy_alignedK_contig<at::native::<unnamed>::Opa",
      "ms": 1.948261
     }
    ]
   },
   {
    "unit": "non-graph: tfm_sub/_timestep_embeddings",
    "ms": 10.031028000000006,
    "share": 0.00029920620044943455,
    "count": 2560,
    "families": {
     "gemm": 8.183562,
     "elementwise_reduce": 1.6904420000000007,
     "gather_index_cat": 0.15702399999999997
    },
    "top_kernels": [
     {
      "name": "nvjet_sm90_tst_64x8_64x16_2x1_v_bz_bias_TNT",
      "ms": 6.237406
     },
     {
      "name": "nvjet_sm90_tst_64x8_64x16_4x1_v_bz_bias_TNT",
      "ms": 1.9461559999999998
     },
     {
      "name": "void at::native::elementwise_kernel<(int)128, (int)4, void at::native::gpu_kernel_impl<at:",
      "ms": 0.424803
     },
     {
      "name": "void at::native::vectorized_elementwise_kernel<(int)8, at::native::<unnamed>::silu_kernel(",
      "ms": 0.253156
     }
    ]
   },
   {
    "unit": "non-graph: tfm_sub/patch_embedding/Conv3dLayer",
    "ms": 5.351960999999999,
    "share": 0.00015963866472743922,
    "count": 320,
    "families": {
     "gemm": 4.8020380000000005,
     "elementwise_reduce": 0.549923
    },
    "top_kernels": [
     {
      "name": "nvjet_sm90_tst_128x192_64x5_2x1_v_bz_coopB_bias_TNN",
      "ms": 4.8020380000000005
     },
     {
      "name": "void at::native::elementwise_kernel<(int)128, (int)4, void at::native::gpu_kernel_impl_noc",
      "ms": 0.549923
     }
    ]
   },
   {
    "unit": "non-graph: tfm/f=15/c=0/s=*",
    "ms": 2.103434,
    "share": 6.274137556351708e-05,
    "count": 64,
    "families": {
     "elementwise_reduce": 2.103434
    },
    "top_kernels": [
     {
      "name": "void at::native::vectorized_elementwise_kernel<(int)8, at::native::CUDAFunctor_add<c10::BF",
      "ms": 1.159013
     },
     {
      "name": "void at::native::vectorized_elementwise_kernel<(int)8, at::native::<unnamed>::silu_kernel(",
      "ms": 0.944421
     }
    ]
   },
   {
    "unit": "non-graph: tfm/f=18/c=0/s=*",
    "ms": 2.102921,
    "share": 6.2726073763858e-05,
    "count": 64,
    "families": {
     "elementwise_reduce": 2.102921
    },
    "top_kernels": [
     {
      "name": "void at::native::vectorized_elementwise_kernel<(int)8, at::native::CUDAFunctor_add<c10::BF",
      "ms": 1.159363
     },
     {
      "name": "void at::native::vectorized_elementwise_kernel<(int)8, at::native::<unnamed>::silu_kernel(",
      "ms": 0.943558
     }
    ]
   },
   {
    "unit": "non-graph: tfm/f=21/c=0/s=*",
    "ms": 2.102378,
    "share": 6.27098771221136e-05,
    "count": 64,
    "families": {
     "elementwise_reduce": 2.102378
    },
    "top_kernels": [
     {
      "name": "void at::native::vectorized_elementwise_kernel<(int)8, at::native::CUDAFunctor_add<c10::BF",
      "ms": 1.159239
     },
     {
      "name": "void at::native::vectorized_elementwise_kernel<(int)8, at::native::<unnamed>::silu_kernel(",
      "ms": 0.943139
     }
    ]
   },
   {
    "unit": "non-graph: tfm/f=24/c=0/s=*",
    "ms": 2.100078,
    "share": 6.26412725622386e-05,
    "count": 64,
    "families": {
     "elementwise_reduce": 2.100078
    },
    "top_kernels": [
     {
      "name": "void at::native::vectorized_elementwise_kernel<(int)8, at::native::CUDAFunctor_add<c10::BF",
      "ms": 1.160614
     },
     {
      "name": "void at::native::vectorized_elementwise_kernel<(int)8, at::native::<unnamed>::silu_kernel(",
      "ms": 0.939464
     }
    ]
   },
   {
    "unit": "non-graph: gen/f=*",
    "ms": 1.707688,
    "share": 5.0937036366870246e-05,
    "count": 3200,
    "families": {
     "elementwise_reduce": 1.707688
    },
    "top_kernels": [
     {
      "name": "void at::native::<unnamed>::distribution_elementwise_grid_stride_kernel<float, (int)4, voi",
      "ms": 0.361251
     },
     {
      "name": "void at::native::elementwise_kernel<(int)128, (int)4, void at::native::gpu_kernel_impl_noc",
      "ms": 0.316065
     },
     {
      "name": "void at::native::vectorized_elementwise_kernel<(int)4, at::native::AUnaryFunctor<float, fl",
      "ms": 0.305443
     },
     {
      "name": "void at::native::vectorized_elementwise_kernel<(int)4, at::native::CUDAFunctor_add<float>,",
      "ms": 0.271553
     }
    ]
   },
   {
    "unit": "non-graph: tfm_sub/_rotary_embedding",
    "ms": 1.302791,
    "share": 3.8859740506129494e-05,
    "count": 640,
    "families": {
     "gather_index_cat": 1.007141,
     "elementwise_reduce": 0.29564999999999997
    },
    "top_kernels": [
     {
      "name": "void at::native::<unnamed>::CatArrayBatchedCopy<at::native::<unnamed>::OpaqueType<(unsigne",
      "ms": 1.007141
     },
     {
      "name": "void at::native::vectorized_elementwise_kernel<(int)8, at::native::bfloat16_copy_kernel_cu",
      "ms": 0.29564999999999997
     }
    ]
   },
   {
    "unit": "non-graph: tfm/f=15/c=1/s=None",
    "ms": 0.532224,
    "share": 1.587521446735068e-05,
    "count": 16,
    "families": {
     "elementwise_reduce": 0.532224
    },
    "top_kernels": [
     {
      "name": "void at::native::vectorized_elementwise_kernel<(int)8, at::native::CUDAFunctor_add<c10::BF",
      "ms": 0.292064
     },
     {
      "name": "void at::native::vectorized_elementwise_kernel<(int)8, at::native::<unnamed>::silu_kernel(",
      "ms": 0.24016
     }
    ]
   },
   {
    "unit": "non-graph: tfm/f=21/c=1/s=None",
    "ms": 0.530948,
    "share": 1.5837153850654816e-05,
    "count": 16,
    "families": {
     "elementwise_reduce": 0.530948
    },
    "top_kernels": [
     {
      "name": "void at::native::vectorized_elementwise_kernel<(int)8, at::native::CUDAFunctor_add<c10::BF",
      "ms": 0.290114
     },
     {
      "name": "void at::native::vectorized_elementwise_kernel<(int)8, at::native::<unnamed>::silu_kernel(",
      "ms": 0.240834
     }
    ]
   },
   {
    "unit": "non-graph: tfm/f=24/c=1/s=None",
    "ms": 0.5290250000000001,
    "share": 1.5779794472985425e-05,
    "count": 16,
    "families": {
     "elementwise_reduce": 0.5290250000000001
    },
    "top_kernels": [
     {
      "name": "void at::native::vectorized_elementwise_kernel<(int)8, at::native::CUDAFunctor_add<c10::BF",
      "ms": 0.289409
     },
     {
      "name": "void at::native::vectorized_elementwise_kernel<(int)8, at::native::<unnamed>::silu_kernel(",
      "ms": 0.239616
     }
    ]
   },
   {
    "unit": "non-graph: tfm/f=18/c=1/s=None",
    "ms": 0.526466,
    "share": 1.570346444310712e-05,
    "count": 16,
    "families": {
     "elementwise_reduce": 0.526466
    },
    "top_kernels": [
     {
      "name": "void at::native::vectorized_elementwise_kernel<(int)8, at::native::CUDAFunctor_add<c10::BF",
      "ms": 0.287809
     },
     {
      "name": "void at::native::vectorized_elementwise_kernel<(int)8, at::native::<unnamed>::silu_kernel(",
      "ms": 0.238657
     }
    ]
   },
   {
    "unit": "non-graph: tfm_sub/_unpatchify",
    "ms": 0.376802,
    "share": 1.1239276247833003e-05,
    "count": 160,
    "families": {
     "elementwise_reduce": 0.376802
    },
    "top_kernels": [
     {
      "name": "void at::native::elementwise_kernel<(int)128, (int)4, void at::native::gpu_kernel_impl_noc",
      "ms": 0.376802
     }
    ]
   }
  ],
  "by_family": {
   "attention_core": 17761.099766000007,
   "gemm": 10084.369307,
   "nccl": 2906.0744629999995,
   "triton_fused": 1589.7333830000005,
   "gather_index_cat": 1014.3510330000003,
   "elementwise_reduce": 159.3740700000002,
   "attention_prep": 7.728994,
   "conv": 2.737323
  }
 },
 "vae_steady": {
  "gpu_ms": 4570.909085,
  "groups": 1418,
  "by_unit": [
   {
    "unit": "vae_frame/*",
    "ms": 563.5872800000003,
    "share": 0.12329872887856842,
    "count": 4364,
    "families": {
     "attention_core": 430.418209,
     "elementwise_reduce": 133.169071
    },
    "top_kernels": [
     {
      "name": "fmha_cutlassF_f32_aligned_32x128_gmem_sm80(PyTorchMemEffAttention::AttentionKernel<float, ",
      "ms": 430.418209
     },
     {
      "name": "void at::native::elementwise_kernel<(int)128, (int)2, void at::native::gpu_kernel_impl_noc",
      "ms": 92.06099800000001
     },
     {
      "name": "void at::native::vectorized_elementwise_kernel<(int)4, at::native::CUDAFunctor_add<float>,",
      "ms": 41.108073000000005
     }
    ]
   },
   {
    "unit": "vae/up_blocks.3.resnets.0.conv1/WanDistCausalConv3d",
    "ms": 245.099703,
    "share": 0.05362165347027461,
    "count": 1347,
    "families": {
     "conv": 166.07522400000002,
     "elementwise_reduce": 76.61983099999999,
     "nccl": 2.404648
    },
    "top_kernels": [
     {
      "name": "sm90_xmma_fprop_implicit_gemm_f32f32_tf32f32_f32_nhwckrsc_nhwc_tilesize256x32x32_warpgroup",
      "ms": 141.209718
     },
     {
      "name": "void at::native::elementwise_kernel<(int)128, (int)2, void at::native::gpu_kernel_impl_noc",
      "ms": 71.233532
     },
     {
      "name": "void cudnn::engines_precompiled::nchwToNhwcKernel<float, float, float, (bool)0, (bool)1, (",
      "ms": 15.704882
     },
     {
      "name": "void cudnn::engines_precompiled::nhwcToNchwKernel<float, float, float, (bool)1, (bool)0, (",
      "ms": 9.160623999999999
     }
    ]
   },
   {
    "unit": "vae/up_blocks.3.resnets.0.conv2/WanDistCausalConv3d",
    "ms": 244.21076900000003,
    "share": 0.05342717705793092,
    "count": 1347,
    "families": {
     "conv": 166.344438,
     "elementwise_reduce": 75.826643,
     "nccl": 2.039688
    },
    "top_kernels": [
     {
      "name": "sm90_xmma_fprop_implicit_gemm_f32f32_tf32f32_f32_nhwckrsc_nhwc_tilesize256x32x32_warpgroup",
      "ms": 141.416633
     },
     {
      "name": "void at::native::elementwise_kernel<(int)128, (int)2, void at::native::gpu_kernel_impl_noc",
      "ms": 70.44274700000001
     },
     {
      "name": "void cudnn::engines_precompiled::nchwToNhwcKernel<float, float, float, (bool)0, (bool)1, (",
      "ms": 15.765712
     },
     {
      "name": "void cudnn::engines_precompiled::nhwcToNchwKernel<float, float, float, (bool)1, (bool)0, (",
      "ms": 9.162093
     }
    ]
   },
   {
    "unit": "vae/up_blocks.3.resnets.1.conv2/WanDistCausalConv3d",
    "ms": 244.10245500000002,
    "share": 0.0534034806776298,
    "count": 1347,
    "families": {
     "conv": 166.18037299999997,
     "elementwise_reduce": 75.77339599999999,
     "nccl": 2.148686
    },
    "top_kernels": [
     {
      "name": "sm90_xmma_fprop_implicit_gemm_f32f32_tf32f32_f32_nhwckrsc_nhwc_tilesize256x32x32_warpgroup",
      "ms": 141.280599
     },
     {
      "name": "void at::native::elementwise_kernel<(int)128, (int)2, void at::native::gpu_kernel_impl_noc",
      "ms": 70.39531999999998
     },
     {
      "name": "void cudnn::engines_precompiled::nchwToNhwcKernel<float, float, float, (bool)0, (bool)1, (",
      "ms": 15.754059999999999
     },
     {
      "name": "void cudnn::engines_precompiled::nhwcToNchwKernel<float, float, float, (bool)1, (bool)0, (",
      "ms": 9.145714
     }
    ]
   },
   {
    "unit": "vae/up_blocks.3.resnets.2.conv2/WanDistCausalConv3d",
    "ms": 244.089324,
    "share": 0.053400607944929186,
    "count": 1347,
    "families": {
     "conv": 166.20612500000001,
     "elementwise_reduce": 75.72158399999998,
     "nccl": 2.1616150000000003
    },
    "top_kernels": [
     {
      "name": "sm90_xmma_fprop_implicit_gemm_f32f32_tf32f32_f32_nhwckrsc_nhwc_tilesize256x32x32_warpgroup",
      "ms": 141.324144
     },
     {
      "name": "void at::native::elementwise_kernel<(int)128, (int)2, void at::native::gpu_kernel_impl_noc",
      "ms": 70.33023299999999
     },
     {
      "name": "void cudnn::engines_precompiled::nchwToNhwcKernel<float, float, float, (bool)0, (bool)1, (",
      "ms": 15.730379999999998
     },
     {
      "name": "void cudnn::engines_precompiled::nhwcToNchwKernel<float, float, float, (bool)1, (bool)0, (",
      "ms": 9.151601
     }
    ]
   },
   {
    "unit": "vae/up_blocks.3.resnets.1.conv1/WanDistCausalConv3d",
    "ms": 243.827444,
    "share": 0.05334331518431415,
    "count": 1348,
    "families": {
     "conv": 166.0698,
     "elementwise_reduce": 75.79949199999999,
     "nccl": 1.9581520000000001
    },
    "top_kernels": [
     {
      "name": "sm90_xmma_fprop_implicit_gemm_f32f32_tf32f32_f32_nhwckrsc_nhwc_tilesize256x32x32_warpgroup",
      "ms": 141.24946
     },
     {
      "name": "void at::native::elementwise_kernel<(int)128, (int)2, void at::native::gpu_kernel_impl_noc",
      "ms": 70.423207
     },
     {
      "name": "void cudnn::engines_precompiled::nchwToNhwcKernel<float, float, float, (bool)0, (bool)1, (",
      "ms": 15.663911
     },
     {
      "name": "void cudnn::engines_precompiled::nhwcToNchwKernel<float, float, float, (bool)1, (bool)0, (",
      "ms": 9.156429
     }
    ]
   },
   {
    "unit": "vae/up_blocks.3.resnets.2.conv1/WanDistCausalConv3d",
    "ms": 243.516771,
    "share": 0.05327534774190329,
    "count": 1347,
    "families": {
     "conv": 165.897677,
     "elementwise_reduce": 75.724585,
     "nccl": 1.894509
    },
    "top_kernels": [
     {
      "name": "sm90_xmma_fprop_implicit_gemm_f32f32_tf32f32_f32_nhwckrsc_nhwc_tilesize256x32x32_warpgroup",
      "ms": 141.040787
     },
     {
      "name": "void at::native::elementwise_kernel<(int)128, (int)2, void at::native::gpu_kernel_impl_noc",
      "ms": 70.340875
     },
     {
      "name": "void cudnn::engines_precompiled::nchwToNhwcKernel<float, float, float, (bool)0, (bool)1, (",
      "ms": 15.694219000000002
     },
     {
      "name": "void cudnn::engines_precompiled::nhwcToNchwKernel<float, float, float, (bool)1, (bool)0, (",
      "ms": 9.162671
     }
    ]
   },
   {
    "unit": "vae/up_blocks.2.upsamplers.0.resample.1/WanDistConv2d",
    "ms": 176.824485,
    "share": 0.03868475213831561,
    "count": 1091,
    "families": {
     "conv": 99.794861,
     "elementwise_reduce": 45.78201599999999,
     "gather_index_cat": 28.176172,
     "nccl": 3.0714360000000003
    },
    "top_kernels": [
     {
      "name": "sm90_xmma_fprop_implicit_gemm_f32f32_tf32f32_f32_nhwckrsc_nhwc_tilesize64x192x32_warpgroup",
      "ms": 70.899704
     },
     {
      "name": "void at::native::elementwise_kernel<(int)128, (int)2, void at::native::gpu_kernel_impl_noc",
      "ms": 38.621824
     },
     {
      "name": "void at::native::<unnamed>::CatArrayBatchedCopy_alignedK_contig<at::native::<unnamed>::Opa",
      "ms": 28.176172
     },
     {
      "name": "void cudnn::engines_precompiled::nchwToNhwcKernel<float, float, float, (bool)0, (bool)1, (",
      "ms": 20.103657000000002
     }
    ]
   },
   {
    "unit": "vae/up_blocks.2.resnets.0.conv1/WanDistCausalConv3d",
    "ms": 139.07559899999995,
    "share": 0.03042624484840305,
    "count": 1347,
    "families": {
     "conv": 95.62677099999999,
     "elementwise_reduce": 40.68846,
     "nccl": 2.760368
    },
    "top_kernels": [
     {
      "name": "sm90_xmma_fprop_implicit_gemm_f32f32_tf32f32_f32_nhwckrsc_nhwc_tilesize128x128x32_warpgrou",
      "ms": 83.297431
     },
     {
      "name": "void at::native::elementwise_kernel<(int)128, (int)2, void at::native::gpu_kernel_impl_noc",
      "ms": 37.947483
     },
     {
      "name": "void cudnn::engines_precompiled::nchwToNhwcKernel<float, float, float, (bool)0, (bool)1, (",
      "ms": 7.973482000000001
     },
     {
      "name": "void cudnn::engines_precompiled::nhwcToNchwKernel<float, float, float, (bool)1, (bool)0, (",
      "ms": 4.355858
     }
    ]
   },
   {
    "unit": "vae/up_blocks.2.resnets.2.conv1/WanDistCausalConv3d",
    "ms": 138.688908,
    "share": 0.030341646578603956,
    "count": 1347,
    "families": {
     "conv": 95.85212200000001,
     "elementwise_reduce": 40.18634199999999,
     "nccl": 2.6504440000000002
    },
    "top_kernels": [
     {
      "name": "sm90_xmma_fprop_implicit_gemm_f32f32_tf32f32_f32_nhwckrsc_nhwc_tilesize128x128x32_warpgrou",
      "ms": 83.528829
     },
     {
      "name": "void at::native::elementwise_kernel<(int)128, (int)2, void at::native::gpu_kernel_impl_noc",
      "ms": 37.444824999999994
     },
     {
      "name": "void cudnn::engines_precompiled::nchwToNhwcKernel<float, float, float, (bool)0, (bool)1, (",
      "ms": 7.974532999999999
     },
     {
      "name": "void cudnn::engines_precompiled::nhwcToNchwKernel<float, float, float, (bool)1, (bool)0, (",
      "ms": 4.34876
     }
    ]
   },
   {
    "unit": "vae/up_blocks.2.resnets.1.conv2/WanDistCausalConv3d",
    "ms": 138.67357999999996,
    "share": 0.030338293197533582,
    "count": 1347,
    "families": {
     "conv": 95.71474399999998,
     "elementwise_reduce": 40.19763700000001,
     "nccl": 2.761199
    },
    "top_kernels": [
     {
      "name": "sm90_xmma_fprop_implicit_gemm_f32f32_tf32f32_f32_nhwckrsc_nhwc_tilesize128x128x32_warpgrou",
      "ms": 83.409498
     },
     {
      "name": "void at::native::elementwise_kernel<(int)128, (int)2, void at::native::gpu_kernel_impl_noc",
      "ms": 37.463164000000006
     },
     {
      "name": "void cudnn::engines_precompiled::nchwToNhwcKernel<float, float, float, (bool)0, (bool)1, (",
      "ms": 7.955655
     },
     {
      "name": "void cudnn::engines_precompiled::nhwcToNchwKernel<float, float, float, (bool)1, (bool)0, (",
      "ms": 4.349591
     }
    ]
   },
   {
    "unit": "vae/up_blocks.2.resnets.2.conv2/WanDistCausalConv3d",
    "ms": 138.43373400000007,
    "share": 0.030285820922207222,
    "count": 1348,
    "families": {
     "conv": 95.61080399999999,
     "elementwise_reduce": 40.186657,
     "nccl": 2.636273
    },
    "top_kernels": [
     {
      "name": "sm90_xmma_fprop_implicit_gemm_f32f32_tf32f32_f32_nhwckrsc_nhwc_tilesize128x128x32_warpgrou",
      "ms": 83.29487499999999
     },
     {
      "name": "void at::native::elementwise_kernel<(int)128, (int)2, void at::native::gpu_kernel_impl_noc",
      "ms": 37.44872699999999
     },
     {
      "name": "void cudnn::engines_precompiled::nchwToNhwcKernel<float, float, float, (bool)0, (bool)1, (",
      "ms": 7.9649350000000005
     },
     {
      "name": "void cudnn::engines_precompiled::nhwcToNchwKernel<float, float, float, (bool)1, (bool)0, (",
      "ms": 4.350994
     }
    ]
   },
   {
    "unit": "vae/conv_out/WanDistCausalConv3d",
    "ms": 138.35416000000006,
    "share": 0.030268412131413035,
    "count": 1343,
    "families": {
     "conv": 72.54535200000001,
     "elementwise_reduce": 63.83542399999999,
     "nccl": 1.973384
    },
    "top_kernels": [
     {
      "name": "void at::native::elementwise_kernel<(int)128, (int)2, void at::native::gpu_kernel_impl_noc",
      "ms": 58.451939
     },
     {
      "name": "sm80_xmma_fprop_implicit_gemm_indexed_wo_smem_tf32f32_tf32f32_f32_nhwckrsc_nhwc_tilesize12",
      "ms": 54.756045
     },
     {
      "name": "void cudnn::engines_precompiled::nchwToNhwcKernel<float, float, float, (bool)0, (bool)1, (",
      "ms": 15.859722
     },
     {
      "name": "void at::native::vectorized_elementwise_kernel<(int)4, at::native::FillFunctor<float>, std",
      "ms": 5.383485
     }
    ]
   },
   {
    "unit": "vae/up_blocks.2.resnets.1.conv1/WanDistCausalConv3d",
    "ms": 138.33613100000005,
    "share": 0.03026446783944294,
    "count": 1347,
    "families": {
     "conv": 95.558741,
     "elementwise_reduce": 40.210432999999995,
     "nccl": 2.5669570000000004
    },
    "top_kernels": [
     {
      "name": "sm90_xmma_fprop_implicit_gemm_f32f32_tf32f32_f32_nhwckrsc_nhwc_tilesize128x128x32_warpgrou",
      "ms": 83.231871
     },
     {
      "name": "void at::native::elementwise_kernel<(int)128, (int)2, void at::native::gpu_kernel_impl_noc",
      "ms": 37.462709
     },
     {
      "name": "void cudnn::engines_precompiled::nchwToNhwcKernel<float, float, float, (bool)0, (bool)1, (",
      "ms": 7.9772170000000004
     },
     {
      "name": "void cudnn::engines_precompiled::nhwcToNchwKernel<float, float, float, (bool)1, (bool)0, (",
      "ms": 4.349653
     }
    ]
   },
   {
    "unit": "vae/up_blocks.2.resnets.0.conv2/WanDistCausalConv3d",
    "ms": 137.88205500000004,
    "share": 0.030165127425631138,
    "count": 1347,
    "families": {
     "conv": 95.40961900000002,
     "elementwise_reduce": 40.234373999999995,
     "nccl": 2.2380619999999998
    },
    "top_kernels": [
     {
      "name": "sm90_xmma_fprop_implicit_gemm_f32f32_tf32f32_f32_nhwckrsc_nhwc_tilesize128x128x32_warpgrou",
      "ms": 83.085364
     },
     {
      "name": "void at::native::elementwise_kernel<(int)128, (int)2, void at::native::gpu_kernel_impl_noc",
      "ms": 37.486458
     },
     {
      "name": "void cudnn::engines_precompiled::nchwToNhwcKernel<float, float, float, (bool)0, (bool)1, (",
      "ms": 7.969928
     },
     {
      "name": "void cudnn::engines_precompiled::nhwcToNchwKernel<float, float, float, (bool)1, (bool)0, (",
      "ms": 4.354327
     }
    ]
   },
   {
    "unit": "vae/up_blocks.1.upsamplers.0.resample.1/WanDistConv2d",
    "ms": 104.53778900000002,
    "share": 0.022870240264251506,
    "count": 1163,
    "families": {
     "conv": 61.590250000000005,
     "elementwise_reduce": 23.928855,
     "gather_index_cat": 16.265075,
     "nccl": 2.753609
    },
    "top_kernels": [
     {
      "name": "sm90_xmma_fprop_implicit_gemm_f32f32_tf32f32_f32_nhwckrsc_nhwc_tilesize128x128x32_warpgrou",
      "ms": 46.996231
     },
     {
      "name": "void at::native::elementwise_kernel<(int)128, (int)2, void at::native::gpu_kernel_impl_noc",
      "ms": 20.294247
     },
     {
      "name": "void at::native::<unnamed>::CatArrayBatchedCopy_alignedK_contig<at::native::<unnamed>::Opa",
      "ms": 16.265075
     },
     {
      "name": "void cudnn::engines_precompiled::nchwToNhwcKernel<float, float, float, (bool)0, (bool)1, (",
      "ms": 10.175885999999998
     }
    ]
   },
   {
    "unit": "vae/up_blocks.1.resnets.0.conv2/WanDistCausalConv3d",
    "ms": 71.718564,
    "share": 0.01569021887469897,
    "count": 1343,
    "families": {
     "conv": 53.641316,
     "elementwise_reduce": 14.395887999999998,
     "nccl": 3.6813599999999997
    },
    "top_kernels": [
     {
      "name": "sm90_xmma_fprop_implicit_gemm_f32f32_tf32f32_f32_nhwckrsc_nhwc_tilesize128x128x32_warpgrou",
      "ms": 49.337264
     },
     {
      "name": "void at::native::elementwise_kernel<(int)128, (int)2, void at::native::gpu_kernel_impl_noc",
      "ms": 13.427468999999999
     },
     {
      "name": "ncclDevKernel_SendRecv(ncclDevKernelArgsStorage<(unsigned long)4096>)",
      "ms": 3.6813599999999997
     },
     {
      "name": "void cudnn::engines_precompiled::nchwToNhwcKernel<float, float, float, (bool)0, (bool)1, (",
      "ms": 3.2454870000000002
     }
    ]
   },
   {
    "unit": "vae/up_blocks.1.resnets.1.conv1/WanDistCausalConv3d",
    "ms": 70.75471800000003,
    "share": 0.01547935359996337,
    "count": 1343,
    "families": {
     "conv": 53.57779999999999,
     "elementwise_reduce": 14.389673000000004,
     "nccl": 2.787245
    },
    "top_kernels": [
     {
      "name": "sm90_xmma_fprop_implicit_gemm_f32f32_tf32f32_f32_nhwckrsc_nhwc_tilesize128x128x32_warpgrou",
      "ms": 49.280018
     },
     {
      "name": "void at::native::elementwise_kernel<(int)128, (int)2, void at::native::gpu_kernel_impl_noc",
      "ms": 13.421508000000003
     },
     {
      "name": "void cudnn::engines_precompiled::nchwToNhwcKernel<float, float, float, (bool)0, (bool)1, (",
      "ms": 3.2411350000000003
     },
     {
      "name": "ncclDevKernel_SendRecv(ncclDevKernelArgsStorage<(unsigned long)4096>)",
      "ms": 2.787245
     }
    ]
   },
   {
    "unit": "vae/up_blocks.1.resnets.2.conv1/WanDistCausalConv3d",
    "ms": 70.257846,
    "share": 0.015370650497197538,
    "count": 1343,
    "families": {
     "conv": 53.656519,
     "elementwise_reduce": 14.383874999999998,
     "nccl": 2.2174519999999998
    },
    "top_kernels": [
     {
      "name": "sm90_xmma_fprop_implicit_gemm_f32f32_tf32f32_f32_nhwckrsc_nhwc_tilesize128x128x32_warpgrou",
      "ms": 49.365557
     },
     {
      "name": "void at::native::elementwise_kernel<(int)128, (int)2, void at::native::gpu_kernel_impl_noc",
      "ms": 13.414977
     },
     {
      "name": "void cudnn::engines_precompiled::nchwToNhwcKernel<float, float, float, (bool)0, (bool)1, (",
      "ms": 3.2306369999999998
     },
     {
      "name": "ncclDevKernel_SendRecv(ncclDevKernelArgsStorage<(unsigned long)4096>)",
      "ms": 2.2174519999999998
     }
    ]
   },
   {
    "unit": "vae/up_blocks.1.resnets.1.conv2/WanDistCausalConv3d",
    "ms": 70.17208799999999,
    "share": 0.015351888802662542,
    "count": 1343,
    "families": {
     "conv": 53.539333,
     "elementwise_reduce": 14.384163999999998,
     "nccl": 2.2485910000000002
    },
    "top_kernels": [
     {
      "name": "sm90_xmma_fprop_implicit_gemm_f32f32_tf32f32_f32_nhwckrsc_nhwc_tilesize128x128x32_warpgrou",
      "ms": 49.251029
     },
     {
      "name": "void at::native::elementwise_kernel<(int)128, (int)2, void at::native::gpu_kernel_impl_noc",
      "ms": 13.410300999999999
     },
     {
      "name": "void cudnn::engines_precompiled::nchwToNhwcKernel<float, float, float, (bool)0, (bool)1, (",
      "ms": 3.228523
     },
     {
      "name": "ncclDevKernel_SendRecv(ncclDevKernelArgsStorage<(unsigned long)4096>)",
      "ms": 2.2485910000000002
     }
    ]
   },
   {
    "unit": "vae/up_blocks.1.resnets.2.conv2/WanDistCausalConv3d",
    "ms": 70.12342100000001,
    "share": 0.015341241686499219,
    "count": 1343,
    "families": {
     "conv": 53.47175099999999,
     "elementwise_reduce": 14.387659000000003,
     "nccl": 2.264011
    },
    "top_kernels": [
     {
      "name": "sm90_xmma_fprop_implicit_gemm_f32f32_tf32f32_f32_nhwckrsc_nhwc_tilesize128x128x32_warpgrou",
      "ms": 49.168243
     },
     {
      "name": "void at::native::elementwise_kernel<(int)128, (int)2, void at::native::gpu_kernel_impl_noc",
      "ms": 13.414693000000002
     },
     {
      "name": "void cudnn::engines_precompiled::nchwToNhwcKernel<float, float, float, (bool)0, (bool)1, (",
      "ms": 3.241967
     },
     {
      "name": "ncclDevKernel_SendRecv(ncclDevKernelArgsStorage<(unsigned long)4096>)",
      "ms": 2.264011
     }
    ]
   },
   {
    "unit": "vae/up_blocks.3.resnets.0.norm1/WanRMS_norm",
    "ms": 47.017766,
    "share": 0.01028630522411714,
    "count": 408,
    "families": {
     "elementwise_reduce": 47.017766
    },
    "top_kernels": [
     {
      "name": "void at::native::elementwise_kernel<(int)128, (int)2, void at::native::gpu_kernel_impl_noc",
      "ms": 26.897322
     },
     {
      "name": "void at::native::vectorized_elementwise_kernel<(int)4, at::native::CUDAFunctorOnSelf_add<f",
      "ms": 7.770178
     },
     {
      "name": "void at::native::vectorized_elementwise_kernel<(int)4, at::native::AUnaryFunctor<float, fl",
      "ms": 7.744614
     },
     {
      "name": "void at::native::reduce_kernel<(int)128, (int)4, at::native::ReduceOp<float, at::native::N",
      "ms": 4.4685310000000005
     }
    ]
   },
   {
    "unit": "vae/up_blocks.3.resnets.0.norm2/WanRMS_norm",
    "ms": 45.368418999999996,
    "share": 0.009925469563348359,
    "count": 408,
    "families": {
     "elementwise_reduce": 45.368418999999996
    },
    "top_kernels": [
     {
      "name": "void at::native::elementwise_kernel<(int)128, (int)2, void at::native::gpu_kernel_impl_noc",
      "ms": 25.284894
     },
     {
      "name": "void at::native::vectorized_elementwise_kernel<(int)4, at::native::CUDAFunctorOnSelf_add<f",
      "ms": 7.779496
     },
     {
      "name": "void at::native::vectorized_elementwise_kernel<(int)4, at::native::AUnaryFunctor<float, fl",
      "ms": 7.731365
     },
     {
      "name": "void at::native::reduce_kernel<(int)128, (int)4, at::native::ReduceOp<float, at::native::N",
      "ms": 4.435991
     }
    ]
   },
   {
    "unit": "vae/up_blocks.3.resnets.1.norm2/WanRMS_norm",
    "ms": 45.341438,
    "share": 0.009919566798822032,
    "count": 408,
    "families": {
     "elementwise_reduce": 45.341438
    },
    "top_kernels": [
     {
      "name": "void at::native::elementwise_kernel<(int)128, (int)2, void at::native::gpu_kernel_impl_noc",
      "ms": 25.249015999999997
     },
     {
      "name": "void at::native::vectorized_elementwise_kernel<(int)4, at::native::CUDAFunctorOnSelf_add<f",
      "ms": 7.771045
     },
     {
      "name": "void at::native::vectorized_elementwise_kernel<(int)4, at::native::AUnaryFunctor<float, fl",
      "ms": 7.743209
     },
     {
      "name": "void at::native::reduce_kernel<(int)128, (int)4, at::native::ReduceOp<float, at::native::N",
      "ms": 4.440888
     }
    ]
   },
   {
    "unit": "vae/up_blocks.3.resnets.2.norm2/WanRMS_norm",
    "ms": 45.314464,
    "share": 0.009913665565719735,
    "count": 408,
    "families": {
     "elementwise_reduce": 45.314464
    },
    "top_kernels": [
     {
      "name": "void at::native::elementwise_kernel<(int)128, (int)2, void at::native::gpu_kernel_impl_noc",
      "ms": 25.230616
     },
     {
      "name": "void at::native::vectorized_elementwise_kernel<(int)4, at::native::CUDAFunctorOnSelf_add<f",
      "ms": 7.776646
     },
     {
      "name": "void at::native::vectorized_elementwise_kernel<(int)4, at::native::AUnaryFunctor<float, fl",
      "ms": 7.7280050000000005
     },
     {
      "name": "void at::native::reduce_kernel<(int)128, (int)4, at::native::ReduceOp<float, at::native::N",
      "ms": 4.441884
     }
    ]
   },
   {
    "unit": "vae/up_blocks.3.resnets.2.norm1/WanRMS_norm",
    "ms": 45.28742000000001,
    "share": 0.009907749018377163,
    "count": 408,
    "families": {
     "elementwise_reduce": 45.28742000000001
    },
    "top_kernels": [
     {
      "name": "void at::native::elementwise_kernel<(int)128, (int)2, void at::native::gpu_kernel_impl_noc",
      "ms": 25.237343999999997
     },
     {
      "name": "void at::native::vectorized_elementwise_kernel<(int)4, at::native::CUDAFunctorOnSelf_add<f",
      "ms": 7.77431
     },
     {
      "name": "void at::native::vectorized_elementwise_kernel<(int)4, at::native::AUnaryFunctor<float, fl",
      "ms": 7.732901
     },
     {
      "name": "void at::native::reduce_kernel<(int)128, (int)4, at::native::ReduceOp<float, at::native::N",
      "ms": 4.399792
     }
    ]
   },
   {
    "unit": "vae/norm_out/WanRMS_norm",
    "ms": 45.275231,
    "share": 0.00990508237159567,
    "count": 408,
    "families": {
     "elementwise_reduce": 45.275231
    },
    "top_kernels": [
     {
      "name": "void at::native::elementwise_kernel<(int)128, (int)2, void at::native::gpu_kernel_impl_noc",
      "ms": 25.216122
     },
     {
      "name": "void at::native::vectorized_elementwise_kernel<(int)4, at::native::CUDAFunctorOnSelf_add<f",
      "ms": 7.77783
     },
     {
      "name": "void at::native::vectorized_elementwise_kernel<(int)4, at::native::AUnaryFunctor<float, fl",
      "ms": 7.745801999999999
     },
     {
      "name": "void at::native::reduce_kernel<(int)128, (int)4, at::native::ReduceOp<float, at::native::N",
      "ms": 4.393492999999999
     }
    ]
   },
   {
    "unit": "vae/up_blocks.3.resnets.1.norm1/WanRMS_norm",
    "ms": 45.177945,
    "share": 0.009883798640462349,
    "count": 408,
    "families": {
     "elementwise_reduce": 45.177945
    },
    "top_kernels": [
     {
      "name": "void at::native::elementwise_kernel<(int)128, (int)2, void at::native::gpu_kernel_impl_noc",
      "ms": 25.261243
     },
     {
      "name": "void at::native::vectorized_elementwise_kernel<(int)4, at::native::CUDAFunctorOnSelf_add<f",
      "ms": 7.752453
     },
     {
      "name": "void at::native::vectorized_elementwise_kernel<(int)4, at::native::AUnaryFunctor<float, fl",
      "ms": 7.737447
     },
     {
      "name": "void at::native::reduce_kernel<(int)128, (int)4, at::native::ReduceOp<float, at::native::N",
      "ms": 4.290578
     }
    ]
   },
   {
    "unit": "vae/up_blocks.1.resnets.0.conv1/WanDistCausalConv3d",
    "ms": 40.09631999999999,
    "share": 0.008772066837115835,
    "count": 1339,
    "families": {
     "conv": 27.460133999999993,
     "elementwise_reduce": 8.215845999999999,
     "nccl": 4.4203399999999995
    },
    "top_kernels": [
     {
      "name": "sm90_xmma_fprop_implicit_gemm_f32f32_tf32f32_f32_nhwckrsc_nhwc_tilesize128x128x32_warpgrou",
      "ms": 24.561335999999997
     },
     {
      "name": "void at::native::elementwise_kernel<(int)128, (int)2, void at::native::gpu_kernel_impl_noc",
      "ms": 7.697475999999999
     },
     {
      "name": "ncclDevKernel_SendRecv(ncclDevKernelArgsStorage<(unsigned long)4096>)",
      "ms": 4.4203399999999995
     },
     {
      "name": "void cudnn::engines_precompiled::nchwToNhwcKernel<float, float, float, (bool)0, (bool)1, (",
      "ms": 1.774794
     }
    ]
   },
   {
    "unit": "vae/up_blocks.2.upsamplers.0.resample.0/WanUpsample",
    "ms": 32.625986999999995,
    "share": 0.007137745772950545,
    "count": 136,
    "families": {
     "gather_index_cat": 26.226373,
     "elementwise_reduce": 6.399614000000001
    },
    "top_kernels": [
     {
      "name": "void at::native::<unnamed>::upsample_nearest2d_out_frame<float, &at::native::nearest_neigh",
      "ms": 26.226373
     },
     {
      "name": "void at::native::elementwise_kernel<(int)128, (int)2, void at::native::gpu_kernel_impl_noc",
      "ms": 6.399614000000001
     }
    ]
   },
   {
    "unit": "ncclAllReduce",
    "ms": 31.890542,
    "share": 0.006976848895256467,
    "count": 100,
    "families": {
     "nccl": 31.890542
    },
    "top_kernels": [
     {
      "name": "ncclDevKernel_AllReduce_Sum_f32_RING_LL(ncclDevKernelArgsStorage<(unsigned long)4096>)",
      "ms": 31.890542
     }
    ]
   },
   {
    "unit": "vae/up_blocks.2.resnets.0.norm1/WanRMS_norm",
    "ms": 23.977753999999997,
    "share": 0.005245729799939785,
    "count": 408,
    "families": {
     "elementwise_reduce": 23.977753999999997
    },
    "top_kernels": [
     {
      "name": "void at::native::elementwise_kernel<(int)128, (int)2, void at::native::gpu_kernel_impl_noc",
      "ms": 13.616547
     },
     {
      "name": "void at::native::vectorized_elementwise_kernel<(int)4, at::native::CUDAFunctorOnSelf_add<f",
      "ms": 3.890616
     },
     {
      "name": "void at::native::vectorized_elementwise_kernel<(int)4, at::native::AUnaryFunctor<float, fl",
      "ms": 3.872564
     },
     {
      "name": "void at::native::reduce_kernel<(int)128, (int)4, at::native::ReduceOp<float, at::native::N",
      "ms": 2.510955
     }
    ]
   },
   {
    "unit": "vae/up_blocks.2.resnets.0.norm2/WanRMS_norm",
    "ms": 23.132047000000004,
    "share": 0.005060710368515239,
    "count": 408,
    "families": {
     "elementwise_reduce": 23.132047000000004
    },
    "top_kernels": [
     {
      "name": "void at::native::elementwise_kernel<(int)128, (int)2, void at::native::gpu_kernel_impl_noc",
      "ms": 12.748768
     },
     {
      "name": "void at::native::vectorized_elementwise_kernel<(int)4, at::native::CUDAFunctorOnSelf_add<f",
      "ms": 3.8840190000000003
     },
     {
      "name": "void at::native::vectorized_elementwise_kernel<(int)4, at::native::AUnaryFunctor<float, fl",
      "ms": 3.8833450000000003
     },
     {
      "name": "void at::native::reduce_kernel<(int)128, (int)4, at::native::ReduceOp<float, at::native::N",
      "ms": 2.524586
     }
    ]
   },
   {
    "unit": "vae/up_blocks.2.resnets.1.norm2/WanRMS_norm",
    "ms": 23.121046999999997,
    "share": 0.00505830384504355,
    "count": 408,
    "families": {
     "elementwise_reduce": 23.121046999999997
    },
    "top_kernels": [
     {
      "name": "void at::native::elementwise_kernel<(int)128, (int)2, void at::native::gpu_kernel_impl_noc",
      "ms": 12.726029
     },
     {
      "name": "void at::native::vectorized_elementwise_kernel<(int)4, at::native::AUnaryFunctor<float, fl",
      "ms": 3.8888439999999997
     },
     {
      "name": "void at::native::vectorized_elementwise_kernel<(int)4, at::native::CUDAFunctorOnSelf_add<f",
      "ms": 3.878003
     },
     {
      "name": "void at::native::reduce_kernel<(int)128, (int)4, at::native::ReduceOp<float, at::native::N",
      "ms": 2.536682
     }
    ]
   },
   {
    "unit": "vae/up_blocks.2.resnets.2.norm2/WanRMS_norm",
    "ms": 23.100143,
    "share": 0.0050537305753479015,
    "count": 408,
    "families": {
     "elementwise_reduce": 23.100143
    },
    "top_kernels": [
     {
      "name": "void at::native::elementwise_kernel<(int)128, (int)2, void at::native::gpu_kernel_impl_noc",
      "ms": 12.716288
     },
     {
      "name": "void at::native::vectorized_elementwise_kernel<(int)4, at::native::CUDAFunctorOnSelf_add<f",
      "ms": 3.8851380000000004
     },
     {
      "name": "void at::native::vectorized_elementwise_kernel<(int)4, at::native::AUnaryFunctor<float, fl",
      "ms": 3.8846260000000004
     },
     {
      "name": "void at::native::reduce_kernel<(int)128, (int)4, at::native::ReduceOp<float, at::native::N",
      "ms": 2.519914
     }
    ]
   },
   {
    "unit": "vae/up_blocks.2.resnets.2.norm1/WanRMS_norm",
    "ms": 23.058387000000003,
    "share": 0.005044595412249378,
    "count": 408,
    "families": {
     "elementwise_reduce": 23.058387000000003
    },
    "top_kernels": [
     {
      "name": "void at::native::elementwise_kernel<(int)128, (int)2, void at::native::gpu_kernel_impl_noc",
      "ms": 12.719867
     },
     {
      "name": "void at::native::vectorized_elementwise_kernel<(int)4, at::native::CUDAFunctorOnSelf_add<f",
      "ms": 3.894552
     },
     {
      "name": "void at::native::vectorized_elementwise_kernel<(int)4, at::native::AUnaryFunctor<float, fl",
      "ms": 3.8886279999999998
     },
     {
      "name": "void at::native::reduce_kernel<(int)128, (int)4, at::native::ReduceOp<float, at::native::N",
      "ms": 2.468044
     }
    ]
   },
   {
    "unit": "vae/up_blocks.2.resnets.1.norm1/WanRMS_norm",
    "ms": 22.96356599999999,
    "share": 0.005023850961148572,
    "count": 408,
    "families": {
     "elementwise_reduce": 22.96356599999999
    },
    "top_kernels": [
     {
      "name": "void at::native::elementwise_kernel<(int)128, (int)2, void at::native::gpu_kernel_impl_noc",
      "ms": 12.736194999999999
     },
     {
      "name": "void at::native::vectorized_elementwise_kernel<(int)4, at::native::AUnaryFunctor<float, fl",
      "ms": 3.888816
     },
     {
      "name": "void at::native::vectorized_elementwise_kernel<(int)4, at::native::CUDAFunctorOnSelf_add<f",
      "ms": 3.882258
     },
     {
      "name": "void at::native::reduce_kernel<(int)128, (int)4, at::native::ReduceOp<float, at::native::N",
      "ms": 2.367625
     }
    ]
   },
   {
    "unit": "vae/up_blocks.0.upsamplers.0.resample.1/WanDistConv2d",
    "ms": 18.56351500000001,
    "share": 0.004061230414955849,
    "count": 1087,
    "families": {
     "conv": 8.990508,
     "nccl": 4.264691,
     "elementwise_reduce": 3.286318999999999,
     "gather_index_cat": 2.021997
    },
    "top_kernels": [
     {
      "name": "sm90_xmma_fprop_implicit_gemm_f32f32_tf32f32_f32_nhwckrsc_nhwc_tilesize128x128x32_warpgrou",
      "ms": 6.80198
     },
     {
      "name": "ncclDevKernel_SendRecv(ncclDevKernelArgsStorage<(unsigned long)4096>)",
      "ms": 4.264691
     },
     {
      "name": "void at::native::elementwise_kernel<(int)128, (int)2, void at::native::gpu_kernel_impl_noc",
      "ms": 2.7549559999999995
     },
     {
      "name": "void at::native::<unnamed>::CatArrayBatchedCopy_alignedK_contig<at::native::<unnamed>::Opa",
      "ms": 2.021997
     }
    ]
   },
   {
    "unit": "vae/up_blocks.1.upsamplers.0.time_conv/WanDistCausalConv3d",
    "ms": 18.459545,
    "share": 0.004038484392651183,
    "count": 468,
    "families": {
     "conv": 12.022941999999999,
     "elementwise_reduce": 6.436603000000001
    },
    "top_kernels": [
     {
      "name": "sm90_xmma_fprop_implicit_gemm_f32f32_tf32f32_f32_nhwckrsc_nhwc_tilesize128x128x32_warpgrou",
      "ms": 6.496448
     },
     {
      "name": "void at::native::elementwise_kernel<(int)128, (int)2, void at::native::gpu_kernel_impl_noc",
      "ms": 6.436603000000001
     },
     {
      "name": "void cudnn::engines_precompiled::nchwToNhwcKernel<float, float, float, (bool)0, (bool)1, (",
      "ms": 3.226222
     },
     {
      "name": "void cudnn::engines_precompiled::nhwcToNchwKernel<float, float, float, (bool)1, (bool)0, (",
      "ms": 2.300272
     }
    ]
   },
   {
    "unit": "vae/mid_block.resnets.1.conv1/WanDistCausalConv3d",
    "ms": 17.036052,
    "share": 0.0037270599093528023,
    "count": 1392,
    "families": {
     "conv": 8.151593000000002,
     "nccl": 6.080829,
     "elementwise_reduce": 2.8036300000000005
    },
    "top_kernels": [
     {
      "name": "sm90_xmma_fprop_implicit_gemm_f32f32_tf32f32_f32_nhwckrsc_nhwc_tilesize128x128x32_warpgrou",
      "ms": 6.9608360000000005
     },
     {
      "name": "ncclDevKernel_SendRecv(ncclDevKernelArgsStorage<(unsigned long)4096>)",
      "ms": 6.080829
     },
     {
      "name": "void at::native::elementwise_kernel<(int)128, (int)2, void at::native::gpu_kernel_impl_noc",
      "ms": 2.5633090000000003
     },
     {
      "name": "void cudnn::engines_precompiled::nchwToNhwcKernel<float, float, float, (bool)0, (bool)1, (",
      "ms": 1.01258
     }
    ]
   },
   {
    "unit": "vae/up_blocks.1.upsamplers.0.resample.0/WanUpsample",
    "ms": 17.01762,
    "share": 0.0037230274511136988,
    "count": 136,
    "families": {
     "gather_index_cat": 13.776485000000001,
     "elementwise_reduce": 3.241135
    },
    "top_kernels": [
     {
      "name": "void at::native::<unnamed>::upsample_nearest2d_out_frame<float, &at::native::nearest_neigh",
      "ms": 13.776485000000001
     },
     {
      "name": "void at::native::elementwise_kernel<(int)128, (int)2, void at::native::gpu_kernel_impl_noc",
      "ms": 3.241135
     }
    ]
   },
   {
    "unit": "vae/mid_block.resnets.0.conv2/WanDistCausalConv3d",
    "ms": 16.128398000000004,
    "share": 0.003528488031610019,
    "count": 1392,
    "families": {
     "conv": 8.039076,
     "nccl": 5.279804,
     "elementwise_reduce": 2.809518
    },
    "top_kernels": [
     {
      "name": "sm90_xmma_fprop_implicit_gemm_f32f32_tf32f32_f32_nhwckrsc_nhwc_tilesize128x128x32_warpgrou",
      "ms": 6.849374
     },
     {
      "name": "ncclDevKernel_SendRecv(ncclDevKernelArgsStorage<(unsigned long)4096>)",
      "ms": 5.279804
     },
     {
      "name": "void at::native::elementwise_kernel<(int)128, (int)2, void at::native::gpu_kernel_impl_noc",
      "ms": 2.5690680000000006
     },
     {
      "name": "void cudnn::engines_precompiled::nchwToNhwcKernel<float, float, float, (bool)0, (bool)1, (",
      "ms": 1.0113969999999999
     }
    ]
   },
   {
    "unit": "vae/up_blocks.3.resnets.1.nonlinearity/SiLU",
    "ms": 15.76571,
    "share": 0.0034491410148009974,
    "count": 136,
    "families": {
     "elementwise_reduce": 15.76571
    },
    "top_kernels": [
     {
      "name": "void at::native::vectorized_elementwise_kernel<(int)4, at::native::<unnamed>::silu_kernel(",
      "ms": 15.76571
     }
    ]
   },
   {
    "unit": "vae/up_blocks.3.resnets.2.nonlinearity/SiLU",
    "ms": 15.723984,
    "share": 0.0034400124149483054,
    "count": 136,
    "families": {
     "elementwise_reduce": 15.723984
    },
    "top_kernels": [
     {
      "name": "void at::native::vectorized_elementwise_kernel<(int)4, at::native::<unnamed>::silu_kernel(",
      "ms": 15.723984
     }
    ]
   },
   {
    "unit": "vae/up_blocks.3.resnets.0.nonlinearity/SiLU",
    "ms": 15.719785,
    "share": 0.003439093779306704,
    "count": 136,
    "families": {
     "elementwise_reduce": 15.719785
    },
    "top_kernels": [
     {
      "name": "void at::native::vectorized_elementwise_kernel<(int)4, at::native::<unnamed>::silu_kernel(",
      "ms": 15.719785
     }
    ]
   },
   {
    "unit": "vae/mid_block.resnets.0.conv1/WanDistCausalConv3d",
    "ms": 15.116842000000002,
    "share": 0.0033071850082531233,
    "count": 1392,
    "families": {
     "conv": 8.037959,
     "nccl": 4.274259,
     "elementwise_reduce": 2.8046239999999996
    },
    "top_kernels": [
     {
      "name": "sm90_xmma_fprop_implicit_gemm_f32f32_tf32f32_f32_nhwckrsc_nhwc_tilesize128x128x32_warpgrou",
      "ms": 6.852642
     },
     {
      "name": "ncclDevKernel_SendRecv(ncclDevKernelArgsStorage<(unsigned long)4096>)",
      "ms": 4.274259
     },
     {
      "name": "void at::native::elementwise_kernel<(int)128, (int)2, void at::native::gpu_kernel_impl_noc",
      "ms": 2.5635660000000002
     },
     {
      "name": "void cudnn::engines_precompiled::nchwToNhwcKernel<float, float, float, (bool)0, (bool)1, (",
      "ms": 1.007493
     }
    ]
   },
   {
    "unit": "vae/up_blocks.0.resnets.2.conv2/WanDistCausalConv3d",
    "ms": 14.762248,
    "share": 0.0032296087551695416,
    "count": 1392,
    "families": {
     "conv": 8.124455,
     "nccl": 3.836112,
     "elementwise_reduce": 2.801681
    },
    "top_kernels": [
     {
      "name": "sm90_xmma_fprop_implicit_gemm_f32f32_tf32f32_f32_nhwckrsc_nhwc_tilesize128x128x32_warpgrou",
      "ms": 6.9424030000000005
     },
     {
      "name": "ncclDevKernel_SendRecv(ncclDevKernelArgsStorage<(unsigned long)4096>)",
      "ms": 3.836112
     },
     {
      "name": "void at::native::elementwise_kernel<(int)128, (int)2, void at::native::gpu_kernel_impl_noc",
      "ms": 2.56056
     },
     {
      "name": "void cudnn::engines_precompiled::nchwToNhwcKernel<float, float, float, (bool)0, (bool)1, (",
      "ms": 1.003684
     }
    ]
   },
   {
    "unit": "vae/up_blocks.0.resnets.2.conv1/WanDistCausalConv3d",
    "ms": 14.679972,
    "share": 0.0032116088347007667,
    "count": 1392,
    "families": {
     "conv": 8.124389,
     "nccl": 3.759247,
     "elementwise_reduce": 2.7963359999999993
    },
    "top_kernels": [
     {
      "name": "sm90_xmma_fprop_implicit_gemm_f32f32_tf32f32_f32_nhwckrsc_nhwc_tilesize128x128x32_warpgrou",
      "ms": 6.945216
     },
     {
      "name": "ncclDevKernel_SendRecv(ncclDevKernelArgsStorage<(unsigned long)4096>)",
      "ms": 3.759247
     },
     {
      "name": "void at::native::elementwise_kernel<(int)128, (int)2, void at::native::gpu_kernel_impl_noc",
      "ms": 2.5543179999999994
     },
     {
      "name": "void cudnn::engines_precompiled::nchwToNhwcKernel<float, float, float, (bool)0, (bool)1, (",
      "ms": 1.001061
     }
    ]
   },
   {
    "unit": "vae/up_blocks.0.resnets.1.conv2/WanDistCausalConv3d",
    "ms": 14.647656000000001,
    "share": 0.003204538906290673,
    "count": 1392,
    "families": {
     "conv": 8.122440000000001,
     "nccl": 3.731344,
     "elementwise_reduce": 2.793872
    },
    "top_kernels": [
     {
      "name": "sm90_xmma_fprop_implicit_gemm_f32f32_tf32f32_f32_nhwckrsc_nhwc_tilesize128x128x32_warpgrou",
      "ms": 6.936929
     },
     {
      "name": "ncclDevKernel_SendRecv(ncclDevKernelArgsStorage<(unsigned long)4096>)",
      "ms": 3.731344
     },
     {
      "name": "void at::native::elementwise_kernel<(int)128, (int)2, void at::native::gpu_kernel_impl_noc",
      "ms": 2.5535200000000002
     },
     {
      "name": "void cudnn::engines_precompiled::nchwToNhwcKernel<float, float, float, (bool)0, (bool)1, (",
      "ms": 1.008647
     }
    ]
   },
   {
    "unit": "ncclAllGather",
    "ms": 14.431937999999999,
    "share": 0.0031573452308120012,
    "count": 144,
    "families": {
     "nccl": 14.431937999999999
    },
    "top_kernels": [
     {
      "name": "ncclDevKernel_AllGather_RING_LL(ncclDevKernelArgsStorage<(unsigned long)4096>)",
      "ms": 14.431937999999999
     }
    ]
   },
   {
    "unit": "vae/mid_block.resnets.1.conv2/WanDistCausalConv3d",
    "ms": 13.235748999999998,
    "share": 0.002895649148532561,
    "count": 1392,
    "families": {
     "conv": 8.185516999999999,
     "elementwise_reduce": 2.810127999999999,
     "nccl": 2.240104
    },
    "top_kernels": [
     {
      "name": "sm90_xmma_fprop_implicit_gemm_f32f32_tf32f32_f32_nhwckrsc_nhwc_tilesize128x128x32_warpgrou",
      "ms": 6.995976
     },
     {
      "name": "void at::native::elementwise_kernel<(int)128, (int)2, void at::native::gpu_kernel_impl_noc",
      "ms": 2.5697419999999993
     },
     {
      "name": "ncclDevKernel_SendRecv(ncclDevKernelArgsStorage<(unsigned long)4096>)",
      "ms": 2.240104
     },
     {
      "name": "void cudnn::engines_precompiled::nchwToNhwcKernel<float, float, float, (bool)0, (bool)1, (",
      "ms": 1.012005
     }
    ]
   },
   {
    "unit": "vae/up_blocks.0.resnets.1.conv1/WanDistCausalConv3d",
    "ms": 12.389021000000003,
    "share": 0.0027104063479748697,
    "count": 1392,
    "families": {
     "conv": 8.142603,
     "elementwise_reduce": 2.804908,
     "nccl": 1.44151
    },
    "top_kernels": [
     {
      "name": "sm90_xmma_fprop_implicit_gemm_f32f32_tf32f32_f32_nhwckrsc_nhwc_tilesize128x128x32_warpgrou",
      "ms": 6.9571570000000005
     },
     {
      "name": "void at::native::elementwise_kernel<(int)128, (int)2, void at::native::gpu_kernel_impl_noc",
      "ms": 2.564587
     },
     {
      "name": "ncclDevKernel_SendRecv(ncclDevKernelArgsStorage<(unsigned long)4096>)",
      "ms": 1.44151
     },
     {
      "name": "void cudnn::engines_precompiled::nchwToNhwcKernel<float, float, float, (bool)0, (bool)1, (",
      "ms": 1.007462
     }
    ]
   },
   {
    "unit": "vae/up_blocks.0.resnets.0.conv1/WanDistCausalConv3d",
    "ms": 11.87369,
    "share": 0.0025976648800486914,
    "count": 1392,
    "families": {
     "conv": 8.147526,
     "elementwise_reduce": 2.8043969999999994,
     "nccl": 0.921767
    },
    "top_kernels": [
     {
      "name": "sm90_xmma_fprop_implicit_gemm_f32f32_tf32f32_f32_nhwckrsc_nhwc_tilesize128x128x32_warpgrou",
      "ms": 6.962462
     },
     {
      "name": "void at::native::elementwise_kernel<(int)128, (int)2, void at::native::gpu_kernel_impl_noc",
      "ms": 2.5632410000000005
     },
     {
      "name": "void cudnn::engines_precompiled::nchwToNhwcKernel<float, float, float, (bool)0, (bool)1, (",
      "ms": 1.008199
     },
     {
      "name": "ncclDevKernel_SendRecv(ncclDevKernelArgsStorage<(unsigned long)4096>)",
      "ms": 0.921767
     }
    ]
   },
   {
    "unit": "vae/up_blocks.0.resnets.0.conv2/WanDistCausalConv3d",
    "ms": 11.793685,
    "share": 0.0025801617972893895,
    "count": 1392,
    "families": {
     "conv": 8.131755,
     "elementwise_reduce": 2.7966779999999996,
     "nccl": 0.865252
    },
    "top_kernels": [
     {
      "name": "sm90_xmma_fprop_implicit_gemm_f32f32_tf32f32_f32_nhwckrsc_nhwc_tilesize128x128x32_warpgrou",
      "ms": 6.9461450000000005
     },
     {
      "name": "void at::native::elementwise_kernel<(int)128, (int)2, void at::native::gpu_kernel_impl_noc",
      "ms": 2.554982
     },
     {
      "name": "void cudnn::engines_precompiled::nchwToNhwcKernel<float, float, float, (bool)0, (bool)1, (",
      "ms": 1.007881
     },
     {
      "name": "ncclDevKernel_SendRecv(ncclDevKernelArgsStorage<(unsigned long)4096>)",
      "ms": 0.865252
     }
    ]
   },
   {
    "unit": "vae/up_blocks.2.resnets.1.nonlinearity/SiLU",
    "ms": 7.971074999999999,
    "share": 0.0017438708256434286,
    "count": 136,
    "families": {
     "elementwise_reduce": 7.971074999999999
    },
    "top_kernels": [
     {
      "name": "void at::native::vectorized_elementwise_kernel<(int)4, at::native::<unnamed>::silu_kernel(",
      "ms": 7.971074999999999
     }
    ]
   },
   {
    "unit": "vae/up_blocks.2.resnets.2.nonlinearity/SiLU",
    "ms": 7.961444,
    "share": 0.0017417638049565362,
    "count": 136,
    "families": {
     "elementwise_reduce": 7.961444
    },
    "top_kernels": [
     {
      "name": "void at::native::vectorized_elementwise_kernel<(int)4, at::native::<unnamed>::silu_kernel(",
      "ms": 7.961444
     }
    ]
   },
   {
    "unit": "vae/up_blocks.2.resnets.0.nonlinearity/SiLU",
    "ms": 7.9552380000000005,
    "share": 0.0017404060881687827,
    "count": 136,
    "families": {
     "elementwise_reduce": 7.9552380000000005
    },
    "top_kernels": [
     {
      "name": "void at::native::vectorized_elementwise_kernel<(int)4, at::native::<unnamed>::silu_kernel(",
      "ms": 7.9552380000000005
     }
    ]
   },
   {
    "unit": "vae/nonlinearity/SiLU",
    "ms": 7.855049,
    "share": 0.001718487253613796,
    "count": 68,
    "families": {
     "elementwise_reduce": 7.855049
    },
    "top_kernels": [
     {
      "name": "void at::native::vectorized_elementwise_kernel<(int)4, at::native::<unnamed>::silu_kernel(",
      "ms": 7.855049
     }
    ]
   },
   {
    "unit": "vae/up_blocks.1.resnets.2.norm1/WanRMS_norm",
    "ms": 6.480516999999999,
    "share": 0.001417774206287894,
    "count": 408,
    "families": {
     "elementwise_reduce": 6.480516999999999
    },
    "top_kernels": [
     {
      "name": "void at::native::elementwise_kernel<(int)128, (int)2, void at::native::gpu_kernel_impl_noc",
      "ms": 3.3264499999999995
     },
     {
      "name": "void at::native::reduce_kernel<(int)128, (int)4, at::native::ReduceOp<float, at::native::N",
      "ms": 1.112263
     },
     {
      "name": "void at::native::vectorized_elementwise_kernel<(int)4, at::native::CUDAFunctorOnSelf_add<f",
      "ms": 0.994024
     },
     {
      "name": "void at::native::vectorized_elementwise_kernel<(int)4, at::native::AUnaryFunctor<float, fl",
      "ms": 0.98666
     }
    ]
   },
   {
    "unit": "vae/up_blocks.1.resnets.1.norm1/WanRMS_norm",
    "ms": 6.469403000000001,
    "share": 0.001415342742482046,
    "count": 408,
    "families": {
     "elementwise_reduce": 6.469403000000001
    },
    "top_kernels": [
     {
      "name": "void at::native::elementwise_kernel<(int)128, (int)2, void at::native::gpu_kernel_impl_noc",
      "ms": 3.3276930000000005
     },
     {
      "name": "void at::native::reduce_kernel<(int)128, (int)4, at::native::ReduceOp<float, at::native::N",
      "ms": 1.1002939999999999
     },
     {
      "name": "void at::native::vectorized_elementwise_kernel<(int)4, at::native::CUDAFunctorOnSelf_add<f",
      "ms": 0.992963
     },
     {
      "name": "void at::native::vectorized_elementwise_kernel<(int)4, at::native::AUnaryFunctor<float, fl",
      "ms": 0.9881639999999999
     }
    ]
   },
   {
    "unit": "vae/up_blocks.1.resnets.1.norm2/WanRMS_norm",
    "ms": 6.2945579999999985,
    "share": 0.0013770910518995803,
    "count": 408,
    "families": {
     "elementwise_reduce": 6.2945579999999985
    },
    "top_kernels": [
     {
      "name": "void at::native::elementwise_kernel<(int)128, (int)2, void at::native::gpu_kernel_impl_noc",
      "ms": 3.350295
     },
     {
      "name": "void at::native::vectorized_elementwise_kernel<(int)4, at::native::AUnaryFunctor<float, fl",
      "ms": 0.99408
     },
     {
      "name": "void at::native::vectorized_elementwise_kernel<(int)4, at::native::CUDAFunctorOnSelf_add<f",
      "ms": 0.993191
     },
     {
      "name": "void at::native::reduce_kernel<(int)128, (int)4, at::native::ReduceOp<float, at::native::N",
      "ms": 0.89216
     }
    ]
   },
   {
    "unit": "vae/up_blocks.1.resnets.2.norm2/WanRMS_norm",
    "ms": 6.294497,
    "share": 0.0013770777066330559,
    "count": 408,
    "families": {
     "elementwise_reduce": 6.294497
    },
    "top_kernels": [
     {
      "name": "void at::native::elementwise_kernel<(int)128, (int)2, void at::native::gpu_kernel_impl_noc",
      "ms": 3.353047
     },
     {
      "name": "void at::native::vectorized_elementwise_kernel<(int)4, at::native::CUDAFunctorOnSelf_add<f",
      "ms": 0.9992030000000001
     },
     {
      "name": "void at::native::vectorized_elementwise_kernel<(int)4, at::native::AUnaryFunctor<float, fl",
      "ms": 0.988995
     },
     {
      "name": "void at::native::reduce_kernel<(int)128, (int)4, at::native::ReduceOp<float, at::native::N",
      "ms": 0.886948
     }
    ]
   },
   {
    "unit": "vae/up_blocks.1.resnets.0.norm2/WanRMS_norm",
    "ms": 6.289634999999998,
    "share": 0.0013760140232585696,
    "count": 408,
    "families": {
     "elementwise_reduce": 6.289634999999998
    },
    "top_kernels": [
     {
      "name": "void at::native::elementwise_kernel<(int)128, (int)2, void at::native::gpu_kernel_impl_noc",
      "ms": 3.363091
     },
     {
      "name": "void at::native::vectorized_elementwise_kernel<(int)4, at::native::CUDAFunctorOnSelf_add<f",
      "ms": 0.992741
     },
     {
      "name": "void at::native::vectorized_elementwise_kernel<(int)4, at::native::AUnaryFunctor<float, fl",
      "ms": 0.982885
     },
     {
      "name": "void at::native::reduce_kernel<(int)128, (int)4, at::native::ReduceOp<float, at::native::N",
      "ms": 0.886373
     }
    ]
   },
   {
    "unit": "vae/mid_block.attentions.0.to_qkv/Conv2d",
    "ms": 6.136767,
    "share": 0.0013425703477976743,
    "count": 144,
    "families": {
     "gemm": 3.819573,
     "elementwise_reduce": 2.3171939999999998
    },
    "top_kernels": [
     {
      "name": "void cutlass::Kernel2<cutlass_80_tensorop_s1688gemm_128x128_32x3_nn_align4>(T1::Params)",
      "ms": 3.819573
     },
     {
      "name": "void at::native::elementwise_kernel<(int)128, (int)2, void at::native::gpu_kernel_impl_noc",
      "ms": 2.3171939999999998
     }
    ]
   },
   {
    "unit": "vae/conv_in/WanDistCausalConv3d",
    "ms": 5.217593000000001,
    "share": 0.0011414781836554513,
    "count": 1320,
    "families": {
     "nccl": 2.674767,
     "conv": 1.636325,
     "elementwise_reduce": 0.906501
    },
    "top_kernels": [
     {
      "name": "ncclDevKernel_SendRecv(ncclDevKernelArgsStorage<(unsigned long)4096>)",
      "ms": 2.674767
     },
     {
      "name": "sm80_xmma_fprop_implicit_gemm_tf32f32_tf32f32_f32_nhwckrsc_nchw_tilesize128x64x32_stage5_w",
      "ms": 1.432645
     },
     {
      "name": "void at::native::elementwise_kernel<(int)128, (int)2, void at::native::gpu_kernel_impl_noc",
      "ms": 0.8365810000000001
     },
     {
      "name": "void cudnn::engines_precompiled::nchwToNhwcKernel<float, float, float, (bool)0, (bool)1, (",
      "ms": 0.20368
     }
    ]
   },
   {
    "unit": "vae/mid_block.attentions.0.proj/Conv2d",
    "ms": 3.984056,
    "share": 0.0008716112978650503,
    "count": 216,
    "families": {
     "elementwise_reduce": 2.567725,
     "gemm": 1.416331
    },
    "top_kernels": [
     {
      "name": "void at::native::elementwise_kernel<(int)128, (int)2, void at::native::gpu_kernel_impl_noc",
      "ms": 2.567725
     },
     {
      "name": "void cutlass::Kernel2<cutlass_80_tensorop_s1688gemm_64x128_32x3_nn_align4>(T1::Params)",
      "ms": 1.416331
     }
    ]
   },
   {
    "unit": "vae/up_blocks.1.resnets.0.conv_shortcut/WanDistCausalConv3d",
    "ms": 3.9184539999999997,
    "share": 0.000857259229429631,
    "count": 204,
    "families": {
     "elementwise_reduce": 2.1675649999999997,
     "gemm": 1.750889
    },
    "top_kernels": [
     {
      "name": "void at::native::elementwise_kernel<(int)128, (int)2, void at::native::gpu_kernel_impl_noc",
      "ms": 2.1675649999999997
     },
     {
      "name": "void cutlass::Kernel2<cutlass_80_tensorop_s1688gemm_256x64_16x4_nn_align4>(T1::Params)",
      "ms": 1.750889
     }
    ]
   },
   {
    "unit": "vae/up_blocks.0.upsamplers.0.time_conv/WanDistCausalConv3d",
    "ms": 3.8964980000000002,
    "share": 0.0008524558085801437,
    "count": 536,
    "families": {
     "conv": 2.8198490000000005,
     "elementwise_reduce": 1.076649
    },
    "top_kernels": [
     {
      "name": "sm90_xmma_fprop_implicit_gemm_f32f32_tf32f32_f32_nhwckrsc_nhwc_tilesize128x128x32_warpgrou",
      "ms": 1.286534
     },
     {
      "name": "void cudnn::engines_precompiled::nchwToNhwcKernel<float, float, float, (bool)0, (bool)1, (",
      "ms": 1.119874
     },
     {
      "name": "void at::native::elementwise_kernel<(int)128, (int)2, void at::native::gpu_kernel_impl_noc",
      "ms": 1.076649
     },
     {
      "name": "void cudnn::engines_precompiled::nhwcToNchwKernel<float, float, float, (bool)1, (bool)0, (",
      "ms": 0.265569
     }
    ]
   },
   {
    "unit": "vae/up_blocks.1.resnets.0.norm1/WanRMS_norm",
    "ms": 3.2981889999999994,
    "share": 0.0007215608402327256,
    "count": 408,
    "families": {
     "elementwise_reduce": 3.2981889999999994
    },
    "top_kernels": [
     {
      "name": "void at::native::elementwise_kernel<(int)128, (int)2, void at::native::gpu_kernel_impl_noc",
      "ms": 1.685288
     },
     {
      "name": "void at::native::reduce_kernel<(int)128, (int)4, at::native::ReduceOp<float, at::native::N",
      "ms": 0.6675850000000001
     },
     {
      "name": "void at::native::vectorized_elementwise_kernel<(int)4, at::native::CUDAFunctorOnSelf_add<f",
      "ms": 0.44234
     },
     {
      "name": "void at::native::vectorized_elementwise_kernel<(int)4, at::native::AUnaryFunctor<float, fl",
      "ms": 0.441952
     }
    ]
   },
   {
    "unit": "vae/mid_block.attentions.0.norm/WanRMS_norm",
    "ms": 3.112175,
    "share": 0.0006808656532270538,
    "count": 432,
    "families": {
     "elementwise_reduce": 3.112175
    },
    "top_kernels": [
     {
      "name": "void at::native::elementwise_kernel<(int)128, (int)2, void at::native::gpu_kernel_impl_noc",
      "ms": 1.518822
     },
     {
      "name": "void at::native::reduce_kernel<(int)128, (int)4, at::native::ReduceOp<float, at::native::N",
      "ms": 0.64429
     },
     {
      "name": "void at::native::vectorized_elementwise_kernel<(int)4, at::native::CUDAFunctorOnSelf_add<f",
      "ms": 0.441988
     },
     {
      "name": "void at::native::vectorized_elementwise_kernel<(int)4, at::native::AUnaryFunctor<float, fl",
      "ms": 0.440866
     }
    ]
   },
   {
    "unit": "vae/up_blocks.1.resnets.1.nonlinearity/SiLU",
    "ms": 2.11973,
    "share": 0.0004637436362399232,
    "count": 136,
    "families": {
     "elementwise_reduce": 2.11973
    },
    "top_kernels": [
     {
      "name": "void at::native::vectorized_elementwise_kernel<(int)4, at::native::<unnamed>::silu_kernel(",
      "ms": 2.11973
     }
    ]
   },
   {
    "unit": "vae/up_blocks.1.resnets.2.nonlinearity/SiLU",
    "ms": 2.1126840000000002,
    "share": 0.00046220214856887713,
    "count": 136,
    "families": {
     "elementwise_reduce": 2.1126840000000002
    },
    "top_kernels": [
     {
      "name": "void at::native::vectorized_elementwise_kernel<(int)4, at::native::<unnamed>::silu_kernel(",
      "ms": 2.1126840000000002
     }
    ]
   },
   {
    "unit": "vae/up_blocks.0.upsamplers.0.resample.0/WanUpsample",
    "ms": 1.9902179999999998,
    "share": 0.0004354096664340021,
    "count": 136,
    "families": {
     "gather_index_cat": 1.606055,
     "elementwise_reduce": 0.384163
    },
    "top_kernels": [
     {
      "name": "void at::native::<unnamed>::upsample_nearest2d_out_frame<float, &at::native::nearest_neigh",
      "ms": 1.606055
     },
     {
      "name": "void at::native::elementwise_kernel<(int)128, (int)2, void at::native::gpu_kernel_impl_noc",
      "ms": 0.384163
     }
    ]
   },
   {
    "unit": "vae/up_blocks.1.resnets.0.nonlinearity/SiLU",
    "ms": 1.512324,
    "share": 0.00033085847298142004,
    "count": 140,
    "families": {
     "elementwise_reduce": 1.512324
    },
    "top_kernels": [
     {
      "name": "void at::native::vectorized_elementwise_kernel<(int)4, at::native::<unnamed>::silu_kernel(",
      "ms": 1.512324
     }
    ]
   },
   {
    "unit": "vae_convert/ep*_chunk*",
    "ms": 1.359174,
    "share": 0.0002973531030097047,
    "count": 25,
    "families": {
     "elementwise_reduce": 1.359174
    },
    "top_kernels": [
     {
      "name": "void at::native::unrolled_elementwise_kernel<at::native::direct_copy_kernel_cuda(at::Tenso",
      "ms": 0.360226
     },
     {
      "name": "void at::native::vectorized_elementwise_kernel<(int)4, at::native::AUnaryFunctor<float, fl",
      "ms": 0.253027
     },
     {
      "name": "void at::native::vectorized_elementwise_kernel<(int)4, at::native::<unnamed>::launch_clamp",
      "ms": 0.251232
     },
     {
      "name": "void at::native::vectorized_elementwise_kernel<(int)4, at::native::round_kernel_cuda(at::T",
      "ms": 0.248096
     }
    ]
   },
   {
    "unit": "vae/mid_block.resnets.1.norm1/WanRMS_norm",
    "ms": 1.3298319999999997,
    "share": 0.00029093381103640996,
    "count": 432,
    "families": {
     "elementwise_reduce": 1.3298319999999997
    },
    "top_kernels": [
     {
      "name": "void at::native::reduce_kernel<(int)128, (int)4, at::native::ReduceOp<float, at::native::N",
      "ms": 0.505124
     },
     {
      "name": "void at::native::elementwise_kernel<(int)128, (int)2, void at::native::gpu_kernel_impl_noc",
      "ms": 0.4971859999999999
     },
     {
      "name": "void at::native::vectorized_elementwise_kernel<(int)4, at::native::CUDAFunctorOnSelf_add<f",
      "ms": 0.134944
     },
     {
      "name": "void at::native::vectorized_elementwise_kernel<(int)4, at::native::AUnaryFunctor<float, fl",
      "ms": 0.134338
     }
    ]
   },
   {
    "unit": "vae/mid_block.resnets.0.norm1/WanRMS_norm",
    "ms": 1.326788,
    "share": 0.00029026786035933594,
    "count": 432,
    "families": {
     "elementwise_reduce": 1.326788
    },
    "top_kernels": [
     {
      "name": "void at::native::reduce_kernel<(int)128, (int)4, at::native::ReduceOp<float, at::native::N",
      "ms": 0.502786
     },
     {
      "name": "void at::native::elementwise_kernel<(int)128, (int)2, void at::native::gpu_kernel_impl_noc",
      "ms": 0.496513
     },
     {
      "name": "void at::native::vectorized_elementwise_kernel<(int)4, at::native::CUDAFunctorOnSelf_add<f",
      "ms": 0.135297
     },
     {
      "name": "void at::native::vectorized_elementwise_kernel<(int)4, at::native::AUnaryFunctor<float, fl",
      "ms": 0.133792
     }
    ]
   },
   {
    "unit": "vae/mid_block.resnets.0.norm2/WanRMS_norm",
    "ms": 1.3265349999999998,
    "share": 0.00029021251031948706,
    "count": 432,
    "families": {
     "elementwise_reduce": 1.3265349999999998
    },
    "top_kernels": [
     {
      "name": "void at::native::reduce_kernel<(int)128, (int)4, at::native::ReduceOp<float, at::native::N",
      "ms": 0.504007
     },
     {
      "name": "void at::native::elementwise_kernel<(int)128, (int)2, void at::native::gpu_kernel_impl_noc",
      "ms": 0.49494399999999994
     },
     {
      "name": "void at::native::vectorized_elementwise_kernel<(int)4, at::native::CUDAFunctorOnSelf_add<f",
      "ms": 0.13436800000000002
     },
     {
      "name": "void at::native::vectorized_elementwise_kernel<(int)4, at::native::AUnaryFunctor<float, fl",
      "ms": 0.13424
     }
    ]
   },
   {
    "unit": "vae/mid_block.resnets.1.norm2/WanRMS_norm",
    "ms": 1.323783,
    "share": 0.00028961044190184323,
    "count": 432,
    "families": {
     "elementwise_reduce": 1.323783
    },
    "top_kernels": [
     {
      "name": "void at::native::reduce_kernel<(int)128, (int)4, at::native::ReduceOp<float, at::native::N",
      "ms": 0.50189
     },
     {
      "name": "void at::native::elementwise_kernel<(int)128, (int)2, void at::native::gpu_kernel_impl_noc",
      "ms": 0.494371
     },
     {
      "name": "void at::native::vectorized_elementwise_kernel<(int)4, at::native::CUDAFunctorOnSelf_add<f",
      "ms": 0.135136
     },
     {
      "name": "void at::native::vectorized_elementwise_kernel<(int)4, at::native::AUnaryFunctor<float, fl",
      "ms": 0.133602
     }
    ]
   },
   {
    "unit": "vae/up_blocks.0.resnets.2.norm2/WanRMS_norm",
    "ms": 1.3232760000000001,
    "share": 0.0002894995230472846,
    "count": 432,
    "families": {
     "elementwise_reduce": 1.3232760000000001
    },
    "top_kernels": [
     {
      "name": "void at::native::reduce_kernel<(int)128, (int)4, at::native::ReduceOp<float, at::native::N",
      "ms": 0.501155
     },
     {
      "name": "void at::native::elementwise_kernel<(int)128, (int)2, void at::native::gpu_kernel_impl_noc",
      "ms": 0.493799
     },
     {
      "name": "void at::native::vectorized_elementwise_kernel<(int)4, at::native::CUDAFunctorOnSelf_add<f",
      "ms": 0.134881
     },
     {
      "name": "void at::native::vectorized_elementwise_kernel<(int)4, at::native::AUnaryFunctor<float, fl",
      "ms": 0.133888
     }
    ]
   },
   {
    "unit": "vae/up_blocks.0.resnets.1.norm2/WanRMS_norm",
    "ms": 1.3231099999999998,
    "share": 0.00028946320642034816,
    "count": 432,
    "families": {
     "elementwise_reduce": 1.3231099999999998
    },
    "top_kernels": [
     {
      "name": "void at::native::reduce_kernel<(int)128, (int)4, at::native::ReduceOp<float, at::native::N",
      "ms": 0.503522
     },
     {
      "name": "void at::native::elementwise_kernel<(int)128, (int)2, void at::native::gpu_kernel_impl_noc",
      "ms": 0.49229
     },
     {
      "name": "void at::native::vectorized_elementwise_kernel<(int)4, at::native::CUDAFunctorOnSelf_add<f",
      "ms": 0.13459300000000002
     },
     {
      "name": "void at::native::vectorized_elementwise_kernel<(int)4, at::native::AUnaryFunctor<float, fl",
      "ms": 0.133857
     }
    ]
   },
   {
    "unit": "vae/up_blocks.0.resnets.0.norm2/WanRMS_norm",
    "ms": 1.3214469999999998,
    "share": 0.0002890993838264013,
    "count": 432,
    "families": {
     "elementwise_reduce": 1.3214469999999998
    },
    "top_kernels": [
     {
      "name": "void at::native::reduce_kernel<(int)128, (int)4, at::native::ReduceOp<float, at::native::N",
      "ms": 0.501699
     },
     {
      "name": "void at::native::elementwise_kernel<(int)128, (int)2, void at::native::gpu_kernel_impl_noc",
      "ms": 0.49142600000000003
     },
     {
      "name": "void at::native::vectorized_elementwise_kernel<(int)4, at::native::CUDAFunctorOnSelf_add<f",
      "ms": 0.135136
     },
     {
      "name": "void at::native::vectorized_elementwise_kernel<(int)4, at::native::AUnaryFunctor<float, fl",
      "ms": 0.134498
     }
    ]
   },
   {
    "unit": "vae/up_blocks.0.resnets.2.norm1/WanRMS_norm",
    "ms": 1.3184069999999999,
    "share": 0.00028843430824877146,
    "count": 432,
    "families": {
     "elementwise_reduce": 1.3184069999999999
    },
    "top_kernels": [
     {
      "name": "void at::native::reduce_kernel<(int)128, (int)4, at::native::ReduceOp<float, at::native::N",
      "ms": 0.500131
     },
     {
      "name": "void at::native::elementwise_kernel<(int)128, (int)2, void at::native::gpu_kernel_impl_noc",
      "ms": 0.493474
     },
     {
      "name": "void at::native::vectorized_elementwise_kernel<(int)4, at::native::CUDAFunctorOnSelf_add<f",
      "ms": 0.134305
     },
     {
      "name": "void at::native::vectorized_elementwise_kernel<(int)4, at::native::AUnaryFunctor<float, fl",
      "ms": 0.131936
     }
    ]
   },
   {
    "unit": "vae/up_blocks.0.resnets.0.norm1/WanRMS_norm",
    "ms": 1.317891,
    "share": 0.0002883214204204632,
    "count": 432,
    "families": {
     "elementwise_reduce": 1.317891
    },
    "top_kernels": [
     {
      "name": "void at::native::reduce_kernel<(int)128, (int)4, at::native::ReduceOp<float, at::native::N",
      "ms": 0.497538
     },
     {
      "name": "void at::native::elementwise_kernel<(int)128, (int)2, void at::native::gpu_kernel_impl_noc",
      "ms": 0.49408100000000005
     },
     {
      "name": "void at::native::vectorized_elementwise_kernel<(int)4, at::native::CUDAFunctorOnSelf_add<f",
      "ms": 0.13504
     },
     {
      "name": "void at::native::vectorized_elementwise_kernel<(int)4, at::native::AUnaryFunctor<float, fl",
      "ms": 0.133408
     }
    ]
   },
   {
    "unit": "vae/up_blocks.0.resnets.1.norm1/WanRMS_norm",
    "ms": 1.317031,
    "share": 0.0002881332740399495,
    "count": 432,
    "families": {
     "elementwise_reduce": 1.317031
    },
    "top_kernels": [
     {
      "name": "void at::native::reduce_kernel<(int)128, (int)4, at::native::ReduceOp<float, at::native::N",
      "ms": 0.49888299999999997
     },
     {
      "name": "void at::native::elementwise_kernel<(int)128, (int)2, void at::native::gpu_kernel_impl_noc",
      "ms": 0.49200099999999997
     },
     {
      "name": "void at::native::vectorized_elementwise_kernel<(int)4, at::native::CUDAFunctorOnSelf_add<f",
      "ms": 0.134882
     },
     {
      "name": "void at::native::vectorized_elementwise_kernel<(int)4, at::native::AUnaryFunctor<float, fl",
      "ms": 0.133184
     }
    ]
   },
   {
    "unit": "no context",
    "ms": 0.7604839999999999,
    "share": 0.00016637478144022194,
    "count": 92,
    "families": {
     "elementwise_reduce": 0.7604839999999999
    },
    "top_kernels": [
     {
      "name": "void at::native::vectorized_elementwise_kernel<(int)4, at::native::AbsFunctor<float>, std:",
      "ms": 0.239682
     },
     {
      "name": "void at::native::vectorized_elementwise_kernel<(int)8, at::native::AUnaryFunctor<float, fl",
      "ms": 0.165216
     },
     {
      "name": "void at::native::vectorized_elementwise_kernel<(int)8, at::native::BinaryFunctor<float, fl",
      "ms": 0.158976
     },
     {
      "name": "void at::native::vectorized_elementwise_kernel<(int)8, at::native::BinaryFunctor<bool, boo",
      "ms": 0.11116799999999999
     }
    ]
   },
   {
    "unit": "vae_decode/ep*_chunk*",
    "ms": 0.6420509999999999,
    "share": 0.00014046461832001192,
    "count": 22,
    "families": {
     "elementwise_reduce": 0.6420509999999999
    },
    "top_kernels": [
     {
      "name": "void at::native::elementwise_kernel<(int)128, (int)2, void at::native::gpu_kernel_impl_noc",
      "ms": 0.39853099999999997
     },
     {
      "name": "void at::native::vectorized_elementwise_kernel<(int)4, at::native::<unnamed>::launch_clamp",
      "ms": 0.24352000000000001
     }
    ]
   },
   {
    "unit": "vae/mid_block.resnets.0.nonlinearity/SiLU",
    "ms": 0.284356,
    "share": 6.220994439227617e-05,
    "count": 144,
    "families": {
     "elementwise_reduce": 0.284356
    },
    "top_kernels": [
     {
      "name": "void at::native::vectorized_elementwise_kernel<(int)4, at::native::<unnamed>::silu_kernel(",
      "ms": 0.284356
     }
    ]
   },
   {
    "unit": "vae/mid_block.resnets.1.nonlinearity/SiLU",
    "ms": 0.28112099999999995,
    "share": 6.150220771673912e-05,
    "count": 144,
    "families": {
     "elementwise_reduce": 0.28112099999999995
    },
    "top_kernels": [
     {
      "name": "void at::native::vectorized_elementwise_kernel<(int)4, at::native::<unnamed>::silu_kernel(",
      "ms": 0.28112099999999995
     }
    ]
   },
   {
    "unit": "vae/up_blocks.0.resnets.1.nonlinearity/SiLU",
    "ms": 0.280897,
    "share": 6.145320214786114e-05,
    "count": 144,
    "families": {
     "elementwise_reduce": 0.280897
    },
    "top_kernels": [
     {
      "name": "void at::native::vectorized_elementwise_kernel<(int)4, at::native::<unnamed>::silu_kernel(",
      "ms": 0.280897
     }
    ]
   },
   {
    "unit": "vae/up_blocks.0.resnets.0.nonlinearity/SiLU",
    "ms": 0.27962,
    "share": 6.117382665028438e-05,
    "count": 144,
    "families": {
     "elementwise_reduce": 0.27962
    },
    "top_kernels": [
     {
      "name": "void at::native::vectorized_elementwise_kernel<(int)4, at::native::<unnamed>::silu_kernel(",
      "ms": 0.27962
     }
    ]
   },
   {
    "unit": "vae/up_blocks.0.resnets.2.nonlinearity/SiLU",
    "ms": 0.279074,
    "share": 6.105437557614427e-05,
    "count": 144,
    "families": {
     "elementwise_reduce": 0.279074
    },
    "top_kernels": [
     {
      "name": "void at::native::vectorized_elementwise_kernel<(int)4, at::native::<unnamed>::silu_kernel(",
      "ms": 0.279074
     }
    ]
   },
   {
    "unit": "vae_pqc/root/WanCausalConv3d",
    "ms": 0.13846599999999998,
    "share": 3.0292879911874243e-05,
    "count": 48,
    "families": {
     "gemm": 0.08541,
     "elementwise_reduce": 0.053056
    },
    "top_kernels": [
     {
      "name": "void cutlass::Kernel2<cutlass_80_tensorop_s1688gemm_128x64_16x6_nn_align4>(T1::Params)",
      "ms": 0.08541
     },
     {
      "name": "void at::native::elementwise_kernel<(int)128, (int)2, void at::native::gpu_kernel_impl_noc",
      "ms": 0.053056
     }
    ]
   },
   {
    "unit": "vae_h2d/ep*_chunk*",
    "ms": 0.12,
    "share": 2.6252983327494905e-05,
    "count": 160,
    "families": {
     "elementwise_reduce": 0.12
    },
    "top_kernels": [
     {
      "name": "void at::native::elementwise_kernel<(int)128, (int)2, void at::native::gpu_kernel_impl_noc",
      "ms": 0.12
     }
    ]
   },
   {
    "unit": "vae_barrier/ep*_chunk*",
    "ms": 0.012704,
    "share": 2.7793158349374608e-06,
    "count": 80,
    "families": {
     "elementwise_reduce": 0.012704
    },
    "top_kernels": [
     {
      "name": "void at::native::vectorized_elementwise_kernel<(int)4, at::native::FillFunctor<float>, std",
      "ms": 0.012704
     }
    ]
   },
   {
    "unit": "vae_barrier2/ep*_chunk*",
    "ms": 0.012224,
    "share": 2.674303901627481e-06,
    "count": 20,
    "families": {
     "elementwise_reduce": 0.012224
    },
    "top_kernels": [
     {
      "name": "void at::native::vectorized_elementwise_kernel<(int)4, at::native::FillFunctor<float>, std",
      "ms": 0.012224
     }
    ]
   }
  ],
  "by_family": {
   "conv": 2206.5006910000047,
   "elementwise_reduce": 1699.5156299999983,
   "attention_core": 430.418209,
   "nccl": 139.33019499999995,
   "gather_index_cat": 88.072157,
   "gemm": 7.072202999999999
  }
 },
 "dit_measured": {
  "gpu_ms": 88018.61646800001,
  "groups": 2119,
  "by_unit": [
   {
    "unit": "vllm_omni.ar_diffusion_paged_write_attn: ar_diffusion_paged_write_attn",
    "ms": 35312.65884499998,
    "share": 0.4011953409633317,
    "count": 142560,
    "families": {
     "attention_core": 32869.459299,
     "gather_index_cat": 2400.3450249999983,
     "elementwise_reduce": 23.575448999999995,
     "attention_prep": 19.279072
    },
    "top_kernels": [
     {
      "name": "void cutlass::device_kernel<flash::enable_sm90_or_later<flash::FlashAttnFwdSm90<flash::Col",
      "ms": 32869.459299
     },
     {
      "name": "void at::native::vectorized_gather_kernel<(int)16, long>(char *, char *, T2 *, int, long, ",
      "ms": 1260.9733910000002
     },
     {
      "name": "void at::native::index_elementwise_kernel<(int)128, (int)4, void at::native::gpu_index_ker",
      "ms": 1127.2445109999999
     },
     {
      "name": "void flash::prepare_varlen_num_blocks_kernel<(int)1, (bool)0>(int, int, int, const int *, ",
      "ms": 19.279072
     }
    ]
   },
   {
    "unit": "non-graph: no context",
    "ms": 15761.801650000003,
    "share": 0.1790734992492225,
    "count": 126296,
    "families": {
     "elementwise_reduce": 8525.503144999993,
     "conv": 6878.432679,
     "attention_core": 261.483392,
     "gemm": 66.08377900000002,
     "gather_index_cat": 25.680902,
     "other": 4.617753
    },
    "top_kernels": [
     {
      "name": "sm80_xmma_fprop_implicit_gemm_bf16bf16_bf16f32_f32_nhwckrsc_nhwc_tilesize256x32x32_stage4_",
      "ms": 2823.608482
     },
     {
      "name": "void at::native::elementwise_kernel<(int)128, (int)2, void at::native::gpu_kernel_impl_noc",
      "ms": 2745.8217240000004
     },
     {
      "name": "void at::native::elementwise_kernel<(int)128, (int)4, void at::native::gpu_kernel_impl<at:",
      "ms": 2225.4552689999996
     },
     {
      "name": "sm90_xmma_fprop_implicit_gemm_bf16bf16_bf16f32_f32_nhwckrsc_nhwc_tilesize256x128x64_warpgr",
      "ms": 1746.023414
     }
    ]
   },
   {
    "unit": "addmm: float_1, modulation, chunk, float_2, layer_norm, scale_msa, add_1, mul",
    "ms": 4541.831867000001,
    "share": 0.051600809570225704,
    "count": 15840,
    "families": {
     "gemm": 4541.831867000001
    },
    "top_kernels": [
     {
      "name": "nvjet_sm90_tst_192x208_64x4_2x1_v_bz_coopB_bias_TNT",
      "ms": 4541.831867000001
     }
    ]
   },
   {
    "unit": "addmm: output_parallel_6, input_1, output_parallel_7",
    "ms": 4163.160148999999,
    "share": 0.047298631994670756,
    "count": 15840,
    "families": {
     "gemm": 4163.160148999999
    },
    "top_kernels": [
     {
      "name": "nvjet_sm90_tst_192x192_64x4_1x2_h_bz_coopB_bias_TNN",
      "ms": 4163.160148999999
     }
    ]
   },
   {
    "unit": "addmm: output_parallel_6",
    "ms": 3930.459572,
    "share": 0.04465486654665784,
    "count": 15840,
    "families": {
     "gemm": 3930.459572
    },
    "top_kernels": [
     {
      "name": "nvjet_sm90_tst_128x240_64x4_2x1_v_bz_coopA_bias_TNT",
      "ms": 3930.459572
     }
    ]
   },
   {
    "unit": "_c10d_functional.all_to_all_single: output_parallel, split, value_7, stack, reshape, transpose, input_t, t",
    "ms": 2354.410602,
    "share": 0.02674900715868407,
    "count": 15840,
    "families": {
     "nccl": 2354.410602
    },
    "top_kernels": [
     {
      "name": "ncclDevKernel_SendRecv(ncclDevKernelArgsStorage<(unsigned long)4096>)",
      "ms": 2354.410602
     }
    ]
   },
   {
    "unit": "_c10d_functional.all_to_all_single: output_5, reshape_3, transpose_3, transpose_4, contiguous_7, tensor_1",
    "ms": 2103.5247690000006,
    "share": 0.023898634782162892,
    "count": 15840,
    "families": {
     "nccl": 1153.4821889999996,
     "gemm": 950.0425799999999
    },
    "top_kernels": [
     {
      "name": "ncclDevKernel_SendRecv(ncclDevKernelArgsStorage<(unsigned long)4096>)",
      "ms": 1153.4821889999996
     },
     {
      "name": "nvjet_sm90_tst_192x192_64x4_1x2_h_bz_coopB_bias_TNN",
      "ms": 950.0425799999999
     }
    ]
   },
   {
    "unit": "addmm: output_parallel_2",
    "ms": 2053.1359249999996,
    "share": 0.023326155390620534,
    "count": 15840,
    "families": {
     "nccl": 1083.2016879999999,
     "gemm": 969.934237
    },
    "top_kernels": [
     {
      "name": "ncclDevKernel_SendRecv(ncclDevKernelArgsStorage<(unsigned long)4096>)",
      "ms": 1083.2016879999999
     },
     {
      "name": "nvjet_sm90_tst_192x192_64x4_1x2_h_bz_coopB_bias_TNN",
      "ms": 969.934237
     }
    ]
   },
   {
    "unit": "addmm: output_parallel_2, input_parallel_1, output_parallel_3",
    "ms": 1942.9516529999999,
    "share": 0.02207432621604974,
    "count": 15840,
    "families": {
     "gemm": 1942.9516529999999
    },
    "top_kernels": [
     {
      "name": "nvjet_sm90_tst_192x192_64x4_1x2_h_bz_coopB_bias_TNN",
      "ms": 1942.9516529999999
     }
    ]
   },
   {
    "unit": "_c10d_functional.all_to_all_single: output_parallel_4, value_float_2, truediv_4, rms_2, truediv_5, float_1",
    "ms": 1685.883012,
    "share": 0.01915370951795088,
    "count": 15840,
    "families": {
     "nccl": 1685.883012
    },
    "top_kernels": [
     {
      "name": "ncclDevKernel_SendRecv(ncclDevKernelArgsStorage<(unsigned long)4096>)",
      "ms": 1685.883012
     }
    ]
   },
   {
    "unit": "addmm: contiguous_12, input_parallel_2, output_parallel_5",
    "ms": 1661.3866299999997,
    "share": 0.018875400417183476,
    "count": 15840,
    "families": {
     "gemm": 1661.3866299999997
    },
    "top_kernels": [
     {
      "name": "nvjet_sm90_tst_192x192_64x4_1x2_h_bz_coopB_bias_TNN",
      "ms": 1661.3866299999997
     }
    ]
   },
   {
    "unit": "addmm: output_parallel_3, camera_features, camera_scale",
    "ms": 1649.913328,
    "share": 0.01874504956118961,
    "count": 15840,
    "families": {
     "gemm": 1649.913328
    },
    "top_kernels": [
     {
      "name": "nvjet_sm90_tst_192x192_64x4_1x2_h_bz_coopB_bias_TNN",
      "ms": 1649.913328
     }
    ]
   },
   {
    "unit": "addmm: contiguous_8, input_parallel, output_parallel_1",
    "ms": 1625.9181130000002,
    "share": 0.018472434335424007,
    "count": 15840,
    "families": {
     "gemm": 1625.9181130000002
    },
    "top_kernels": [
     {
      "name": "nvjet_sm90_tst_192x192_64x4_1x2_h_bz_coopB_bias_TNN",
      "ms": 1625.9181130000002
     }
    ]
   },
   {
    "unit": "addmm: float_8, layer_norm_1, float_9, float_10, to_6, output_parallel_4",
    "ms": 1619.51534,
    "share": 0.01839969094025455,
    "count": 15840,
    "families": {
     "gemm": 1619.51534
    },
    "top_kernels": [
     {
      "name": "nvjet_sm90_tst_192x192_64x4_1x2_h_bz_coopB_bias_TNN",
      "ms": 1619.51534
     }
    ]
   },
   {
    "unit": "addmm: camera_shift",
    "ms": 1597.8859699999998,
    "share": 0.01815395463050622,
    "count": 15840,
    "families": {
     "gemm": 1597.8859699999998
    },
    "top_kernels": [
     {
      "name": "nvjet_sm90_tst_192x192_64x4_1x2_h_bz_coopB_bias_TNN",
      "ms": 1597.8859699999998
     }
    ]
   },
   {
    "unit": "_c10d_functional.all_to_all_single: reshape_10, transpose_8, transpose_9, contiguous_11, tensor_3",
    "ms": 742.303982,
    "share": 0.008433488411736982,
    "count": 15840,
    "families": {
     "nccl": 742.303982
    },
    "top_kernels": [
     {
      "name": "ncclDevKernel_SendRecv(ncclDevKernelArgsStorage<(unsigned long)4096>)",
      "ms": 742.303982
     }
    ]
   },
   {
    "unit": "fa3_fwd._flash_attn_forward: _flash_attn_forward_default, key_4, value_9",
    "ms": 510.03532500000006,
    "share": 0.005794630107432195,
    "count": 15840,
    "families": {
     "attention_core": 510.03532500000006
    },
    "top_kernels": [
     {
      "name": "void cutlass::device_kernel<flash::enable_sm90_or_later<flash::FlashAttnFwdSm90<flash::Col",
      "ms": 510.03532500000006
     }
    ]
   },
   {
    "unit": "triton_red_fused__to_copy_add_mul_native_layer_norm_split_squeeze_view_9: float_1, modulation, chunk, camera_scale, add_11, output_parallel_1, g",
    "ms": 445.67757300000005,
    "share": 0.0050634466989381765,
    "count": 15840,
    "families": {
     "triton_fused": 445.67757300000005
    },
    "top_kernels": [
     {
      "name": "triton_red_fused__to_copy_add_mul_native_layer_norm_split_squeeze_view_9",
      "ms": 445.67757300000005
     }
    ]
   },
   {
    "unit": "triton_poi_fused_gelu_view_13: output_parallel_6, input_1",
    "ms": 416.03577700000005,
    "share": 0.004726679351421682,
    "count": 15840,
    "families": {
     "triton_fused": 416.03577700000005
    },
    "top_kernels": [
     {
      "name": "triton_poi_fused_gelu_view_13",
      "ms": 416.03577700000005
     }
    ]
   },
   {
    "unit": "triton_red_fused__to_copy_add_mul_native_layer_norm_split_squeeze_1: float_1, modulation, chunk, float_2, layer_norm, scale_msa, add_1, mul",
    "ms": 383.8201109999999,
    "share": 0.004360669667416793,
    "count": 15840,
    "families": {
     "triton_fused": 383.8201109999999
    },
    "top_kernels": [
     {
      "name": "triton_red_fused__to_copy_add_mul_native_layer_norm_split_squeeze_1",
      "ms": 383.8201109999999
     }
    ]
   },
   {
    "unit": "triton_red_fused__to_copy_add_mul_native_layer_norm_split_squeeze_view_12: float_1, modulation, chunk, output_parallel_5, hidden_states_2, float_",
    "ms": 368.372073,
    "share": 0.004185160910066397,
    "count": 15840,
    "families": {
     "triton_fused": 368.372073
    },
    "top_kernels": [
     {
      "name": "triton_red_fused__to_copy_add_mul_native_layer_norm_split_squeeze_view_12",
      "ms": 368.372073
     }
    ]
   },
   {
    "unit": "triton_poi_fused_all_to_all_single_clone_split_with_sizes_stack_transpose_view_4: output_parallel, split, value_7, stack, reshape, transpose, input_t, t",
    "ms": 326.6999710000001,
    "share": 0.0037117144543935767,
    "count": 15840,
    "families": {
     "triton_fused": 326.6999710000001
    },
    "top_kernels": [
     {
      "name": "triton_poi_fused_all_to_all_single_clone_split_with_sizes_stack_transpose_view_4",
      "ms": 326.6999710000001
     }
    ]
   },
   {
    "unit": "triton_poi_fused__to_copy_add_mul_split_squeeze_view_14: float_1, modulation, chunk, output_parallel_5, hidden_states_2, output",
    "ms": 256.621918,
    "share": 0.002915541374060308,
    "count": 15840,
    "families": {
     "triton_fused": 256.621918
    },
    "top_kernels": [
     {
      "name": "triton_poi_fused__to_copy_add_mul_split_squeeze_view_14",
      "ms": 256.621918
     }
    ]
   },
   {
    "unit": "triton_poi_fused__to_copy_all_to_all_single_clone_div_mul_sqrt_transpose_view_11: output_parallel_4, value_float_2, truediv_4, rms_2, truediv_5, float_1",
    "ms": 198.71966100000006,
    "share": 0.0022577003476559583,
    "count": 15840,
    "families": {
     "triton_fused": 198.71966100000006
    },
    "top_kernels": [
     {
      "name": "triton_poi_fused__to_copy_all_to_all_single_clone_div_mul_sqrt_transpose_view_11",
      "ms": 198.71966100000006
     }
    ]
   },
   {
    "unit": "triton_poi_fused_silu_view_6: output_parallel_2, input_parallel_1",
    "ms": 189.03850200000002,
    "share": 0.0021477104456501736,
    "count": 15840,
    "families": {
     "triton_fused": 189.03850200000002
    },
    "top_kernels": [
     {
      "name": "triton_poi_fused_silu_view_6",
      "ms": 189.03850200000002
     }
    ]
   },
   {
    "unit": "rotary_kernel_0: output_parallel, split, value_float_1, truediv_2, rms_1, truediv_3, fl",
    "ms": 184.33243500000003,
    "share": 0.0020942437224858654,
    "count": 15840,
    "families": {
     "triton_fused": 184.33243500000003
    },
    "top_kernels": [
     {
      "name": "rotary_kernel",
      "ms": 184.33243500000003
     }
    ]
   },
   {
    "unit": "rotary_kernel_0: output_parallel, split, value_float, truediv, rms, truediv_1, float_5,",
    "ms": 183.05923700000002,
    "share": 0.0020797786234978246,
    "count": 15840,
    "families": {
     "triton_fused": 183.05923700000002
    },
    "top_kernels": [
     {
      "name": "rotary_kernel",
      "ms": 183.05923700000002
     }
    ]
   },
   {
    "unit": "triton_red_fused__to_copy_div_mul_pow_split_with_sizes_sqrt_sum_view_3: output_parallel, split, value_float_1, pow_2, sum_of_squares_1, truedi",
    "ms": 168.13111599999996,
    "share": 0.001910176764265837,
    "count": 15840,
    "families": {
     "triton_fused": 168.13111599999996
    },
    "top_kernels": [
     {
      "name": "triton_red_fused__to_copy_div_mul_pow_split_with_sizes_sqrt_sum_view_3",
      "ms": 168.13111599999996
     }
    ]
   },
   {
    "unit": "triton_poi_fused_add_view_7: output_parallel_3, camera_features",
    "ms": 164.6718170000001,
    "share": 0.0018708748627043925,
    "count": 15840,
    "families": {
     "triton_fused": 164.6718170000001
    },
    "top_kernels": [
     {
      "name": "triton_poi_fused_add_view_7",
      "ms": 164.6718170000001
     }
    ]
   },
   {
    "unit": "non-graph: tfm_sub/head/_LingBotHead",
    "ms": 164.27058700000023,
    "share": 0.0018663163952335282,
    "count": 3600,
    "families": {
     "elementwise_reduce": 157.65740800000017,
     "gemm": 6.613178999999999
    },
    "top_kernels": [
     {
      "name": "void at::native::elementwise_kernel<(int)128, (int)2, void at::native::gpu_kernel_impl_noc",
      "ms": 75.093638
     },
     {
      "name": "void at::native::unrolled_elementwise_kernel<at::native::direct_copy_kernel_cuda(at::Tenso",
      "ms": 34.494684
     },
     {
      "name": "void at::native::<unnamed>::vectorized_layer_norm_kernel<float, float, (bool)0>(int, T2, c",
      "ms": 31.915
     },
     {
      "name": "void at::native::vectorized_elementwise_kernel<(int)8, at::native::bfloat16_copy_kernel_cu",
      "ms": 15.070656
     }
    ]
   },
   {
    "unit": "non-graph: tfm_sub/sp_prepare/_LingBotSPPrepare",
    "ms": 149.246677,
    "share": 0.0016956262548646175,
    "count": 400,
    "families": {
     "elementwise_reduce": 149.246677
    },
    "top_kernels": [
     {
      "name": "void at::native::elementwise_kernel<(int)128, (int)4, void at::native::gpu_kernel_impl_noc",
      "ms": 149.246677
     }
    ]
   },
   {
    "unit": "triton_red_fused__to_copy_div_mul_pow_split_with_sizes_sqrt_sum_view_2: output_parallel, split, value_float, pow_1, sum_of_squares, truediv, r",
    "ms": 148.452847,
    "share": 0.001686607367362692,
    "count": 15840,
    "families": {
     "triton_fused": 148.452847
    },
    "top_kernels": [
     {
      "name": "triton_red_fused__to_copy_div_mul_pow_split_with_sizes_sqrt_sum_view_2",
      "ms": 148.452847
     }
    ]
   },
   {
    "unit": "non-graph: tfm_sub/c2ws_hidden_states_layer1/ColumnParallelLinear",
    "ms": 144.86899999999997,
    "share": 0.0016458904469677555,
    "count": 400,
    "families": {
     "gemm": 144.86899999999997
    },
    "top_kernels": [
     {
      "name": "nvjet_sm90_tst_256x128_64x4_1x2_h_bz_coopA_bias_TNT",
      "ms": 144.86899999999997
     }
    ]
   },
   {
    "unit": "non-graph: tfm_sub/c2ws_hidden_states_layer2/RowParallelLinear",
    "ms": 144.674881,
    "share": 0.0016436850158011505,
    "count": 400,
    "families": {
     "gemm": 144.674881
    },
    "top_kernels": [
     {
      "name": "nvjet_sm90_tst_256x128_64x4_1x2_h_bz_coopA_bias_TNT",
      "ms": 144.674881
     }
    ]
   },
   {
    "unit": "triton_poi_fused_all_to_all_single_clone_transpose_unsqueeze_view_5: output_5, reshape_3, transpose_3, transpose_4, contiguous_7, tensor_1",
    "ms": 121.09606000000001,
    "share": 0.0013758005392418957,
    "count": 15840,
    "families": {
     "triton_fused": 121.09606000000001
    },
    "top_kernels": [
     {
      "name": "triton_poi_fused_all_to_all_single_clone_transpose_unsqueeze_view_5",
      "ms": 121.09606000000001
     }
    ]
   },
   {
    "unit": "triton_poi_fused_all_to_all_single_clone_transpose_unsqueeze_view_5: reshape_10, transpose_8, transpose_9, contiguous_11, tensor_3",
    "ms": 120.718056,
    "share": 0.0013715059477660409,
    "count": 15840,
    "families": {
     "triton_fused": 120.718056
    },
    "top_kernels": [
     {
      "name": "triton_poi_fused_all_to_all_single_clone_transpose_unsqueeze_view_5",
      "ms": 120.718056
     }
    ]
   },
   {
    "unit": "triton_poi_fused_clone_transpose_view_8: contiguous_8",
    "ms": 108.40935300000001,
    "share": 0.0012316639064579396,
    "count": 15840,
    "families": {
     "triton_fused": 108.40935300000001
    },
    "top_kernels": [
     {
      "name": "triton_poi_fused_clone_transpose_view_8",
      "ms": 108.40935300000001
     }
    ]
   },
   {
    "unit": "triton_poi_fused_clone_transpose_view_8: contiguous_12",
    "ms": 100.875095,
    "share": 0.001146065446696428,
    "count": 15840,
    "families": {
     "triton_fused": 100.875095
    },
    "top_kernels": [
     {
      "name": "triton_poi_fused_clone_transpose_view_8",
      "ms": 100.875095
     }
    ]
   },
   {
    "unit": "triton_red_fused__to_copy_pow_sum_view_10: output_parallel_4, value_float_2, pow_3, sum_of_squares_2",
    "ms": 76.41671600000001,
    "share": 0.0008681881068623934,
    "count": 15840,
    "families": {
     "triton_fused": 76.41671600000001
    },
    "top_kernels": [
     {
      "name": "triton_red_fused__to_copy_pow_sum_view_10",
      "ms": 76.41671600000001
     }
    ]
   },
   {
    "unit": "non-graph: ncclAllGather",
    "ms": 63.03204299999999,
    "share": 0.000716121719805899,
    "count": 400,
    "families": {
     "nccl": 63.03204299999999
    },
    "top_kernels": [
     {
      "name": "ncclDevKernel_AllGather_RING_LL(ncclDevKernelArgsStorage<(unsigned long)4096>)",
      "ms": 63.03204299999999
     }
    ]
   },
   {
    "unit": "non-graph: tfm_sub/patch_embedding_wancamctrl/_LingBotCameraPatchEmbedding",
    "ms": 60.803933999999984,
    "share": 0.0006908076545614169,
    "count": 800,
    "families": {
     "gemm": 47.01439699999999,
     "elementwise_reduce": 13.789537
    },
    "top_kernels": [
     {
      "name": "nvjet_sm90_tst_256x128_64x4_1x2_h_bz_coopA_bias_TNT",
      "ms": 47.01439699999999
     },
     {
      "name": "void at::native::elementwise_kernel<(int)128, (int)4, void at::native::gpu_kernel_impl_noc",
      "ms": 13.789537
     }
    ]
   },
   {
    "unit": "non-graph: tfm_sub/_timestep_embeddings",
    "ms": 25.140918000000006,
    "share": 0.00028563182436683977,
    "count": 6400,
    "families": {
     "gemm": 20.494974,
     "elementwise_reduce": 4.249301999999994,
     "gather_index_cat": 0.3966419999999999
    },
    "top_kernels": [
     {
      "name": "nvjet_sm90_tst_64x8_64x16_2x1_v_bz_bias_TNT",
      "ms": 15.625439
     },
     {
      "name": "nvjet_sm90_tst_64x8_64x16_4x1_v_bz_bias_TNT",
      "ms": 4.869535
     },
     {
      "name": "void at::native::elementwise_kernel<(int)128, (int)4, void at::native::gpu_kernel_impl<at:",
      "ms": 1.0717809999999999
     },
     {
      "name": "void at::native::vectorized_elementwise_kernel<(int)8, at::native::<unnamed>::silu_kernel(",
      "ms": 0.6366129999999998
     }
    ]
   },
   {
    "unit": "non-graph: tfm_sub/patch_embedding/Conv3dLayer",
    "ms": 13.456153000000004,
    "share": 0.00015287848798318834,
    "count": 800,
    "families": {
     "gemm": 12.072627,
     "elementwise_reduce": 1.383526
    },
    "top_kernels": [
     {
      "name": "nvjet_sm90_tst_128x192_64x5_2x1_v_bz_coopB_bias_TNN",
      "ms": 12.072627
     },
     {
      "name": "void at::native::elementwise_kernel<(int)128, (int)4, void at::native::gpu_kernel_impl_noc",
      "ms": 1.383526
     }
    ]
   },
   {
    "unit": "non-graph: gen/f=*",
    "ms": 4.2812410000000005,
    "share": 4.864017604226358e-05,
    "count": 3200,
    "families": {
     "elementwise_reduce": 4.2812410000000005
    },
    "top_kernels": [
     {
      "name": "void at::native::<unnamed>::distribution_elementwise_grid_stride_kernel<float, (int)4, voi",
      "ms": 0.908101
     },
     {
      "name": "void at::native::elementwise_kernel<(int)128, (int)4, void at::native::gpu_kernel_impl_noc",
      "ms": 0.788995
     },
     {
      "name": "void at::native::vectorized_elementwise_kernel<(int)4, at::native::AUnaryFunctor<float, fl",
      "ms": 0.764933
     },
     {
      "name": "void at::native::vectorized_elementwise_kernel<(int)4, at::native::CUDAFunctor_add<float>,",
      "ms": 0.678628
     }
    ]
   },
   {
    "unit": "non-graph: tfm_sub/_rotary_embedding",
    "ms": 3.2750279999999994,
    "share": 3.7208355816302414e-05,
    "count": 1600,
    "families": {
     "gather_index_cat": 2.5338400000000005,
     "elementwise_reduce": 0.741188
    },
    "top_kernels": [
     {
      "name": "void at::native::<unnamed>::CatArrayBatchedCopy<at::native::<unnamed>::OpaqueType<(unsigne",
      "ms": 2.5338400000000005
     },
     {
      "name": "void at::native::vectorized_elementwise_kernel<(int)8, at::native::bfloat16_copy_kernel_cu",
      "ms": 0.741188
     }
    ]
   },
   {
    "unit": "non-graph: tfm/f=6/c=0/s=*",
    "ms": 2.147081,
    "share": 2.439348726619205e-05,
    "count": 64,
    "families": {
     "elementwise_reduce": 2.147081
    },
    "top_kernels": [
     {
      "name": "void at::native::vectorized_elementwise_kernel<(int)8, at::native::CUDAFunctor_add<c10::BF",
      "ms": 1.16701
     },
     {
      "name": "void at::native::vectorized_elementwise_kernel<(int)8, at::native::<unnamed>::silu_kernel(",
      "ms": 0.980071
     }
    ]
   },
   {
    "unit": "non-graph: tfm/f=0/c=0/s=*",
    "ms": 2.1095099999999998,
    "share": 2.396663438543063e-05,
    "count": 64,
    "families": {
     "elementwise_reduce": 2.1095099999999998
    },
    "top_kernels": [
     {
      "name": "void at::native::vectorized_elementwise_kernel<(int)8, at::native::CUDAFunctor_add<c10::BF",
      "ms": 1.160289
     },
     {
      "name": "void at::native::vectorized_elementwise_kernel<(int)8, at::native::<unnamed>::silu_kernel(",
      "ms": 0.949221
     }
    ]
   },
   {
    "unit": "non-graph: tfm/f=15/c=0/s=*",
    "ms": 2.103434,
    "share": 2.3897603534415053e-05,
    "count": 64,
    "families": {
     "elementwise_reduce": 2.103434
    },
    "top_kernels": [
     {
      "name": "void at::native::vectorized_elementwise_kernel<(int)8, at::native::CUDAFunctor_add<c10::BF",
      "ms": 1.159013
     },
     {
      "name": "void at::native::vectorized_elementwise_kernel<(int)8, at::native::<unnamed>::silu_kernel(",
      "ms": 0.944421
     }
    ]
   },
   {
    "unit": "non-graph: tfm/f=18/c=0/s=*",
    "ms": 2.102921,
    "share": 2.389177522194451e-05,
    "count": 64,
    "families": {
     "elementwise_reduce": 2.102921
    },
    "top_kernels": [
     {
      "name": "void at::native::vectorized_elementwise_kernel<(int)8, at::native::CUDAFunctor_add<c10::BF",
      "ms": 1.159363
     },
     {
      "name": "void at::native::vectorized_elementwise_kernel<(int)8, at::native::<unnamed>::silu_kernel(",
      "ms": 0.943558
     }
    ]
   },
   {
    "unit": "non-graph: tfm/f=9/c=0/s=*",
    "ms": 2.102699,
    "share": 2.3889253028243812e-05,
    "count": 64,
    "families": {
     "elementwise_reduce": 2.102699
    },
    "top_kernels": [
     {
      "name": "void at::native::vectorized_elementwise_kernel<(int)8, at::native::CUDAFunctor_add<c10::BF",
      "ms": 1.153829
     },
     {
      "name": "void at::native::vectorized_elementwise_kernel<(int)8, at::native::<unnamed>::silu_kernel(",
      "ms": 0.94887
     }
    ]
   },
   {
    "unit": "non-graph: tfm/f=21/c=0/s=*",
    "ms": 2.102378,
    "share": 2.3885606072487393e-05,
    "count": 64,
    "families": {
     "elementwise_reduce": 2.102378
    },
    "top_kernels": [
     {
      "name": "void at::native::vectorized_elementwise_kernel<(int)8, at::native::CUDAFunctor_add<c10::BF",
      "ms": 1.159239
     },
     {
      "name": "void at::native::vectorized_elementwise_kernel<(int)8, at::native::<unnamed>::silu_kernel(",
      "ms": 0.943139
     }
    ]
   },
   {
    "unit": "non-graph: tfm/f=12/c=0/s=*",
    "ms": 2.101482,
    "share": 2.38754264078215e-05,
    "count": 64,
    "families": {
     "elementwise_reduce": 2.101482
    },
    "top_kernels": [
     {
      "name": "void at::native::vectorized_elementwise_kernel<(int)8, at::native::CUDAFunctor_add<c10::BF",
      "ms": 1.158405
     },
     {
      "name": "void at::native::vectorized_elementwise_kernel<(int)8, at::native::<unnamed>::silu_kernel(",
      "ms": 0.943077
     }
    ]
   },
   {
    "unit": "non-graph: tfm/f=27/c=0/s=*",
    "ms": 2.100553,
    "share": 2.3864871822470374e-05,
    "count": 64,
    "families": {
     "elementwise_reduce": 2.100553
    },
    "top_kernels": [
     {
      "name": "void at::native::vectorized_elementwise_kernel<(int)8, at::native::CUDAFunctor_add<c10::BF",
      "ms": 1.158438
     },
     {
      "name": "void at::native::vectorized_elementwise_kernel<(int)8, at::native::<unnamed>::silu_kernel(",
      "ms": 0.942115
     }
    ]
   },
   {
    "unit": "non-graph: tfm/f=24/c=0/s=*",
    "ms": 2.100078,
    "share": 2.38594752368495e-05,
    "count": 64,
    "families": {
     "elementwise_reduce": 2.100078
    },
    "top_kernels": [
     {
      "name": "void at::native::vectorized_elementwise_kernel<(int)8, at::native::CUDAFunctor_add<c10::BF",
      "ms": 1.160614
     },
     {
      "name": "void at::native::vectorized_elementwise_kernel<(int)8, at::native::<unnamed>::silu_kernel(",
      "ms": 0.939464
     }
    ]
   },
   {
    "unit": "non-graph: tfm/f=3/c=0/s=*",
    "ms": 2.099148,
    "share": 2.3848909290265485e-05,
    "count": 64,
    "families": {
     "elementwise_reduce": 2.099148
    },
    "top_kernels": [
     {
      "name": "void at::native::vectorized_elementwise_kernel<(int)8, at::native::CUDAFunctor_add<c10::BF",
      "ms": 1.157319
     },
     {
      "name": "void at::native::vectorized_elementwise_kernel<(int)8, at::native::<unnamed>::silu_kernel(",
      "ms": 0.941829
     }
    ]
   },
   {
    "unit": "non-graph: tfm_sub/_unpatchify",
    "ms": 0.9475230000000002,
    "share": 1.0765029467879459e-05,
    "count": 400,
    "families": {
     "elementwise_reduce": 0.9475230000000002
    },
    "top_kernels": [
     {
      "name": "void at::native::elementwise_kernel<(int)128, (int)4, void at::native::gpu_kernel_impl_noc",
      "ms": 0.9475230000000002
     }
    ]
   },
   {
    "unit": "non-graph: tfm/f=0/c=1/s=None",
    "ms": 0.550787,
    "share": 6.2576193776034164e-06,
    "count": 16,
    "families": {
     "elementwise_reduce": 0.550787
    },
    "top_kernels": [
     {
      "name": "void at::native::vectorized_elementwise_kernel<(int)8, at::native::CUDAFunctor_add<c10::BF",
      "ms": 0.289569
     },
     {
      "name": "void at::native::vectorized_elementwise_kernel<(int)8, at::native::<unnamed>::silu_kernel(",
      "ms": 0.261218
     }
    ]
   },
   {
    "unit": "non-graph: tfm/f=3/c=1/s=None",
    "ms": 0.536515,
    "share": 6.0954718618538505e-06,
    "count": 16,
    "families": {
     "elementwise_reduce": 0.536515
    },
    "top_kernels": [
     {
      "name": "void at::native::vectorized_elementwise_kernel<(int)8, at::native::CUDAFunctor_add<c10::BF",
      "ms": 0.292161
     },
     {
      "name": "void at::native::vectorized_elementwise_kernel<(int)8, at::native::<unnamed>::silu_kernel(",
      "ms": 0.244354
     }
    ]
   },
   {
    "unit": "non-graph: tfm/f=6/c=1/s=None",
    "ms": 0.535553,
    "share": 6.0845423558174795e-06,
    "count": 16,
    "families": {
     "elementwise_reduce": 0.535553
    },
    "top_kernels": [
     {
      "name": "void at::native::vectorized_elementwise_kernel<(int)8, at::native::CUDAFunctor_add<c10::BF",
      "ms": 0.290369
     },
     {
      "name": "void at::native::vectorized_elementwise_kernel<(int)8, at::native::<unnamed>::silu_kernel(",
      "ms": 0.245184
     }
    ]
   },
   {
    "unit": "non-graph: tfm/f=12/c=1/s=None",
    "ms": 0.533185,
    "share": 6.057638956343337e-06,
    "count": 16,
    "families": {
     "elementwise_reduce": 0.533185
    },
    "top_kernels": [
     {
      "name": "void at::native::vectorized_elementwise_kernel<(int)8, at::native::CUDAFunctor_add<c10::BF",
      "ms": 0.290945
     },
     {
      "name": "void at::native::vectorized_elementwise_kernel<(int)8, at::native::<unnamed>::silu_kernel(",
      "ms": 0.24224
     }
    ]
   },
   {
    "unit": "non-graph: tfm/f=15/c=1/s=None",
    "ms": 0.532224,
    "share": 6.0467208115398525e-06,
    "count": 16,
    "families": {
     "elementwise_reduce": 0.532224
    },
    "top_kernels": [
     {
      "name": "void at::native::vectorized_elementwise_kernel<(int)8, at::native::CUDAFunctor_add<c10::BF",
      "ms": 0.292064
     },
     {
      "name": "void at::native::vectorized_elementwise_kernel<(int)8, at::native::<unnamed>::silu_kernel(",
      "ms": 0.24016
     }
    ]
   },
   {
    "unit": "non-graph: tfm/f=21/c=1/s=None",
    "ms": 0.530948,
    "share": 6.032223878377264e-06,
    "count": 16,
    "families": {
     "elementwise_reduce": 0.530948
    },
    "top_kernels": [
     {
      "name": "void at::native::vectorized_elementwise_kernel<(int)8, at::native::CUDAFunctor_add<c10::BF",
      "ms": 0.290114
     },
     {
      "name": "void at::native::vectorized_elementwise_kernel<(int)8, at::native::<unnamed>::silu_kernel(",
      "ms": 0.240834
     }
    ]
   },
   {
    "unit": "non-graph: tfm/f=27/c=1/s=None",
    "ms": 0.530211,
    "share": 6.0238506497402535e-06,
    "count": 16,
    "families": {
     "elementwise_reduce": 0.530211
    },
    "top_kernels": [
     {
      "name": "void at::native::vectorized_elementwise_kernel<(int)8, at::native::CUDAFunctor_add<c10::BF",
      "ms": 0.290082
     },
     {
      "name": "void at::native::vectorized_elementwise_kernel<(int)8, at::native::<unnamed>::silu_kernel(",
      "ms": 0.240129
     }
    ]
   },
   {
    "unit": "non-graph: tfm/f=24/c=1/s=None",
    "ms": 0.5290250000000001,
    "share": 6.0103762275374106e-06,
    "count": 16,
    "families": {
     "elementwise_reduce": 0.5290250000000001
    },
    "top_kernels": [
     {
      "name": "void at::native::vectorized_elementwise_kernel<(int)8, at::native::CUDAFunctor_add<c10::BF",
      "ms": 0.289409
     },
     {
      "name": "void at::native::vectorized_elementwise_kernel<(int)8, at::native::<unnamed>::silu_kernel(",
      "ms": 0.239616
     }
    ]
   },
   {
    "unit": "non-graph: tfm/f=9/c=1/s=None",
    "ms": 0.52877,
    "share": 6.0074791131514686e-06,
    "count": 16,
    "families": {
     "elementwise_reduce": 0.52877
    },
    "top_kernels": [
     {
      "name": "void at::native::vectorized_elementwise_kernel<(int)8, at::native::CUDAFunctor_add<c10::BF",
      "ms": 0.289762
     },
     {
      "name": "void at::native::vectorized_elementwise_kernel<(int)8, at::native::<unnamed>::silu_kernel(",
      "ms": 0.239008
     }
    ]
   },
   {
    "unit": "non-graph: tfm/f=18/c=1/s=None",
    "ms": 0.526466,
    "share": 5.981302832582033e-06,
    "count": 16,
    "families": {
     "elementwise_reduce": 0.526466
    },
    "top_kernels": [
     {
      "name": "void at::native::vectorized_elementwise_kernel<(int)8, at::native::CUDAFunctor_add<c10::BF",
      "ms": 0.287809
     },
     {
      "name": "void at::native::vectorized_elementwise_kernel<(int)8, at::native::<unnamed>::silu_kernel(",
      "ms": 0.238657
     }
    ]
   },
   {
    "unit": "non-graph: tfm_sub/text_embedding/Sequential",
    "ms": 0.29046500000000003,
    "share": 3.300040510243663e-06,
    "count": 12,
    "families": {
     "gemm": 0.275233,
     "elementwise_reduce": 0.015232
    },
    "top_kernels": [
     {
      "name": "nvjet_sm90_tst_168x128_64x5_1x2_h_bz_coopA_bias_TNN",
      "ms": 0.275233
     },
     {
      "name": "void at::native::vectorized_elementwise_kernel<(int)8, at::native::GeluCUDAKernelImpl(at::",
      "ms": 0.015232
     }
    ]
   }
  ],
  "by_family": {
   "attention_core": 33640.978015999994,
   "gemm": 25095.09750900003,
   "elementwise_reduce": 8907.793196000017,
   "nccl": 7082.313516,
   "conv": 6878.432679,
   "triton_fused": 3961.1483180000027,
   "gather_index_cat": 2428.956408999999,
   "attention_prep": 19.279072,
   "other": 4.617753
  }
 },
 "vae_measured": {
  "gpu_ms": 11247.420478,
  "groups": 2445,
  "by_unit": [
   {
    "unit": "vae_frame/*",
    "ms": 1400.8379300000008,
    "share": 0.12454748470905355,
    "count": 14796,
    "families": {
     "attention_core": 1076.8601840000001,
     "elementwise_reduce": 323.977746
    },
    "top_kernels": [
     {
      "name": "fmha_cutlassF_f32_aligned_32x128_gmem_sm80(PyTorchMemEffAttention::AttentionKernel<float, ",
      "ms": 1076.8601840000001
     },
     {
      "name": "void at::native::elementwise_kernel<(int)128, (int)2, void at::native::gpu_kernel_impl_noc",
      "ms": 223.118288
     },
     {
      "name": "void at::native::vectorized_elementwise_kernel<(int)4, at::native::CUDAFunctor_add<float>,",
      "ms": 100.85113799999998
     },
     {
      "name": "void at::native::vectorized_elementwise_kernel<(int)4, at::native::FillFunctor<float>, std",
      "ms": 0.00832
     }
    ]
   },
   {
    "unit": "vae/up_blocks.3.resnets.0.conv1/WanDistCausalConv3d",
    "ms": 600.7959480000003,
    "share": 0.053416332142570785,
    "count": 3552,
    "families": {
     "conv": 405.0775559999999,
     "elementwise_reduce": 186.46075400000004,
     "nccl": 9.257638
    },
    "top_kernels": [
     {
      "name": "sm90_xmma_fprop_implicit_gemm_f32f32_tf32f32_f32_nhwckrsc_nhwc_tilesize256x32x32_warpgroup",
      "ms": 344.09578
     },
     {
      "name": "void at::native::elementwise_kernel<(int)128, (int)2, void at::native::gpu_kernel_impl_noc",
      "ms": 173.22529700000004
     },
     {
      "name": "void cudnn::engines_precompiled::nchwToNhwcKernel<float, float, float, (bool)0, (bool)1, (",
      "ms": 38.660353
     },
     {
      "name": "void cudnn::engines_precompiled::nhwcToNchwKernel<float, float, float, (bool)1, (bool)0, (",
      "ms": 22.321423
     }
    ]
   },
   {
    "unit": "vae/up_blocks.3.resnets.0.conv2/WanDistCausalConv3d",
    "ms": 596.8563780000001,
    "share": 0.0530660678301708,
    "count": 3552,
    "families": {
     "conv": 405.6899100000001,
     "elementwise_reduce": 184.56256500000003,
     "nccl": 6.603903
    },
    "top_kernels": [
     {
      "name": "sm90_xmma_fprop_implicit_gemm_f32f32_tf32f32_f32_nhwckrsc_nhwc_tilesize256x32x32_warpgroup",
      "ms": 344.633258
     },
     {
      "name": "void at::native::elementwise_kernel<(int)128, (int)2, void at::native::gpu_kernel_impl_noc",
      "ms": 171.33434600000007
     },
     {
      "name": "void cudnn::engines_precompiled::nchwToNhwcKernel<float, float, float, (bool)0, (bool)1, (",
      "ms": 38.741282
     },
     {
      "name": "void cudnn::engines_precompiled::nhwcToNchwKernel<float, float, float, (bool)1, (bool)0, (",
      "ms": 22.315369999999998
     }
    ]
   },
   {
    "unit": "vae/up_blocks.3.resnets.2.conv2/WanDistCausalConv3d",
    "ms": 596.1596619999998,
    "share": 0.05300412331574964,
    "count": 3552,
    "families": {
     "conv": 405.481574,
     "elementwise_reduce": 185.16669700000003,
     "nccl": 5.511391000000001
    },
    "top_kernels": [
     {
      "name": "sm90_xmma_fprop_implicit_gemm_f32f32_tf32f32_f32_nhwckrsc_nhwc_tilesize256x32x32_warpgroup",
      "ms": 344.478683
     },
     {
      "name": "void at::native::elementwise_kernel<(int)128, (int)2, void at::native::gpu_kernel_impl_noc",
      "ms": 171.933455
     },
     {
      "name": "void cudnn::engines_precompiled::nchwToNhwcKernel<float, float, float, (bool)0, (bool)1, (",
      "ms": 38.698589000000005
     },
     {
      "name": "void cudnn::engines_precompiled::nhwcToNchwKernel<float, float, float, (bool)1, (bool)0, (",
      "ms": 22.304302
     }
    ]
   },
   {
    "unit": "vae/up_blocks.3.resnets.1.conv2/WanDistCausalConv3d",
    "ms": 595.7084450000003,
    "share": 0.05296400593942482,
    "count": 3552,
    "families": {
     "conv": 405.45342500000004,
     "elementwise_reduce": 184.42865200000003,
     "nccl": 5.826368
    },
    "top_kernels": [
     {
      "name": "sm90_xmma_fprop_implicit_gemm_f32f32_tf32f32_f32_nhwckrsc_nhwc_tilesize256x32x32_warpgroup",
      "ms": 344.43360900000005
     },
     {
      "name": "void at::native::elementwise_kernel<(int)128, (int)2, void at::native::gpu_kernel_impl_noc",
      "ms": 171.20260600000003
     },
     {
      "name": "void cudnn::engines_precompiled::nchwToNhwcKernel<float, float, float, (bool)0, (bool)1, (",
      "ms": 38.724374
     },
     {
      "name": "void cudnn::engines_precompiled::nhwcToNchwKernel<float, float, float, (bool)1, (bool)0, (",
      "ms": 22.295442
     }
    ]
   },
   {
    "unit": "vae/up_blocks.3.resnets.1.conv1/WanDistCausalConv3d",
    "ms": 595.2874740000003,
    "share": 0.052926577713030734,
    "count": 3552,
    "families": {
     "conv": 405.048705,
     "elementwise_reduce": 184.5139789999999,
     "nccl": 5.72479
    },
    "top_kernels": [
     {
      "name": "sm90_xmma_fprop_implicit_gemm_f32f32_tf32f32_f32_nhwckrsc_nhwc_tilesize256x32x32_warpgroup",
      "ms": 344.15997200000004
     },
     {
      "name": "void at::native::elementwise_kernel<(int)128, (int)2, void at::native::gpu_kernel_impl_noc",
      "ms": 171.27887499999994
     },
     {
      "name": "void cudnn::engines_precompiled::nchwToNhwcKernel<float, float, float, (bool)0, (bool)1, (",
      "ms": 38.567377
     },
     {
      "name": "void cudnn::engines_precompiled::nhwcToNchwKernel<float, float, float, (bool)1, (bool)0, (",
      "ms": 22.321356
     }
    ]
   },
   {
    "unit": "vae/up_blocks.3.resnets.2.conv1/WanDistCausalConv3d",
    "ms": 594.7286789999997,
    "share": 0.052876895654722915,
    "count": 3552,
    "families": {
     "conv": 405.005546,
     "elementwise_reduce": 184.3612789999999,
     "nccl": 5.361854
    },
    "top_kernels": [
     {
      "name": "sm90_xmma_fprop_implicit_gemm_f32f32_tf32f32_f32_nhwckrsc_nhwc_tilesize256x32x32_warpgroup",
      "ms": 344.04272499999996
     },
     {
      "name": "void at::native::elementwise_kernel<(int)128, (int)2, void at::native::gpu_kernel_impl_noc",
      "ms": 171.12578299999996
     },
     {
      "name": "void cudnn::engines_precompiled::nchwToNhwcKernel<float, float, float, (bool)0, (bool)1, (",
      "ms": 38.639959999999995
     },
     {
      "name": "void cudnn::engines_precompiled::nhwcToNchwKernel<float, float, float, (bool)1, (bool)0, (",
      "ms": 22.322861
     }
    ]
   },
   {
    "unit": "vae/up_blocks.2.upsamplers.0.resample.1/WanDistConv2d",
    "ms": 432.97591400000005,
    "share": 0.03849557459391713,
    "count": 2620,
    "families": {
     "conv": 244.30311599999996,
     "elementwise_reduce": 111.61460900000002,
     "gather_index_cat": 68.617867,
     "nccl": 8.440322
    },
    "top_kernels": [
     {
      "name": "sm90_xmma_fprop_implicit_gemm_f32f32_tf32f32_f32_nhwckrsc_nhwc_tilesize64x192x32_warpgroup",
      "ms": 173.85182899999998
     },
     {
      "name": "void at::native::elementwise_kernel<(int)128, (int)2, void at::native::gpu_kernel_impl_noc",
      "ms": 94.16105300000002
     },
     {
      "name": "void at::native::<unnamed>::CatArrayBatchedCopy_alignedK_contig<at::native::<unnamed>::Opa",
      "ms": 68.617867
     },
     {
      "name": "void cudnn::engines_precompiled::nchwToNhwcKernel<float, float, float, (bool)0, (bool)1, (",
      "ms": 49.019116000000004
     }
    ]
   },
   {
    "unit": "vae/up_blocks.2.resnets.0.conv1/WanDistCausalConv3d",
    "ms": 342.2512600000001,
    "share": 0.0304293113847255,
    "count": 3552,
    "families": {
     "conv": 233.554484,
     "elementwise_reduce": 98.97733600000005,
     "nccl": 9.71944
    },
    "top_kernels": [
     {
      "name": "sm90_xmma_fprop_implicit_gemm_f32f32_tf32f32_f32_nhwckrsc_nhwc_tilesize128x128x32_warpgrou",
      "ms": 203.334922
     },
     {
      "name": "void at::native::elementwise_kernel<(int)128, (int)2, void at::native::gpu_kernel_impl_noc",
      "ms": 92.22504300000003
     },
     {
      "name": "void cudnn::engines_precompiled::nchwToNhwcKernel<float, float, float, (bool)0, (bool)1, (",
      "ms": 19.611008
     },
     {
      "name": "void cudnn::engines_precompiled::nhwcToNchwKernel<float, float, float, (bool)1, (bool)0, (",
      "ms": 10.608554000000002
     }
    ]
   },
   {
    "unit": "vae/up_blocks.2.resnets.1.conv2/WanDistCausalConv3d",
    "ms": 340.4682479999999,
    "share": 0.03027078508053977,
    "count": 3552,
    "families": {
     "conv": 232.53102400000003,
     "elementwise_reduce": 97.86056899999997,
     "nccl": 10.076655
    },
    "top_kernels": [
     {
      "name": "sm90_xmma_fprop_implicit_gemm_f32f32_tf32f32_f32_nhwckrsc_nhwc_tilesize128x128x32_warpgrou",
      "ms": 202.355398
     },
     {
      "name": "void at::native::elementwise_kernel<(int)128, (int)2, void at::native::gpu_kernel_impl_noc",
      "ms": 91.13423599999997
     },
     {
      "name": "void cudnn::engines_precompiled::nchwToNhwcKernel<float, float, float, (bool)0, (bool)1, (",
      "ms": 19.576247
     },
     {
      "name": "void cudnn::engines_precompiled::nhwcToNchwKernel<float, float, float, (bool)1, (bool)0, (",
      "ms": 10.599379
     }
    ]
   },
   {
    "unit": "vae/up_blocks.2.resnets.0.conv2/WanDistCausalConv3d",
    "ms": 340.1790930000002,
    "share": 0.030245076519135376,
    "count": 3552,
    "families": {
     "conv": 232.42827199999994,
     "elementwise_reduce": 97.920597,
     "nccl": 9.830224
    },
    "top_kernels": [
     {
      "name": "sm90_xmma_fprop_implicit_gemm_f32f32_tf32f32_f32_nhwckrsc_nhwc_tilesize128x128x32_warpgrou",
      "ms": 202.22406199999998
     },
     {
      "name": "void at::native::elementwise_kernel<(int)128, (int)2, void at::native::gpu_kernel_impl_noc",
      "ms": 91.17861599999998
     },
     {
      "name": "void cudnn::engines_precompiled::nchwToNhwcKernel<float, float, float, (bool)0, (bool)1, (",
      "ms": 19.600607999999994
     },
     {
      "name": "void cudnn::engines_precompiled::nhwcToNchwKernel<float, float, float, (bool)1, (bool)0, (",
      "ms": 10.603601999999999
     }
    ]
   },
   {
    "unit": "vae/conv_out/WanDistCausalConv3d",
    "ms": 340.06895299999996,
    "share": 0.03023528405158998,
    "count": 3548,
    "families": {
     "conv": 177.62936000000002,
     "elementwise_reduce": 155.35700100000005,
     "nccl": 7.082592
    },
    "top_kernels": [
     {
      "name": "void at::native::elementwise_kernel<(int)128, (int)2, void at::native::gpu_kernel_impl_noc",
      "ms": 142.11978400000004
     },
     {
      "name": "sm80_xmma_fprop_implicit_gemm_indexed_wo_smem_tf32f32_tf32f32_f32_nhwckrsc_nhwc_tilesize12",
      "ms": 133.95312
     },
     {
      "name": "void cudnn::engines_precompiled::nchwToNhwcKernel<float, float, float, (bool)0, (bool)1, (",
      "ms": 38.956055
     },
     {
      "name": "void at::native::vectorized_elementwise_kernel<(int)4, at::native::FillFunctor<float>, std",
      "ms": 13.237217
     }
    ]
   },
   {
    "unit": "vae/up_blocks.2.resnets.2.conv1/WanDistCausalConv3d",
    "ms": 339.64852300000007,
    "share": 0.03019790392511367,
    "count": 3552,
    "families": {
     "conv": 232.74342199999998,
     "elementwise_reduce": 97.83124100000002,
     "nccl": 9.07386
    },
    "top_kernels": [
     {
      "name": "sm90_xmma_fprop_implicit_gemm_f32f32_tf32f32_f32_nhwckrsc_nhwc_tilesize128x128x32_warpgrou",
      "ms": 202.541158
     },
     {
      "name": "void at::native::elementwise_kernel<(int)128, (int)2, void at::native::gpu_kernel_impl_noc",
      "ms": 91.104807
     },
     {
      "name": "void cudnn::engines_precompiled::nchwToNhwcKernel<float, float, float, (bool)0, (bool)1, (",
      "ms": 19.611454
     },
     {
      "name": "void cudnn::engines_precompiled::nhwcToNchwKernel<float, float, float, (bool)1, (bool)0, (",
      "ms": 10.59081
     }
    ]
   },
   {
    "unit": "vae/up_blocks.2.resnets.1.conv1/WanDistCausalConv3d",
    "ms": 339.5924489999999,
    "share": 0.030192918426428895,
    "count": 3552,
    "families": {
     "conv": 232.382576,
     "elementwise_reduce": 97.86573900000002,
     "nccl": 9.344134
    },
    "top_kernels": [
     {
      "name": "sm90_xmma_fprop_implicit_gemm_f32f32_tf32f32_f32_nhwckrsc_nhwc_tilesize128x128x32_warpgrou",
      "ms": 202.155055
     },
     {
      "name": "void at::native::elementwise_kernel<(int)128, (int)2, void at::native::gpu_kernel_impl_noc",
      "ms": 91.11611000000003
     },
     {
      "name": "void cudnn::engines_precompiled::nchwToNhwcKernel<float, float, float, (bool)0, (bool)1, (",
      "ms": 19.626773
     },
     {
      "name": "void cudnn::engines_precompiled::nhwcToNchwKernel<float, float, float, (bool)1, (bool)0, (",
      "ms": 10.600748000000001
     }
    ]
   },
   {
    "unit": "vae/up_blocks.2.resnets.2.conv2/WanDistCausalConv3d",
    "ms": 339.57105900000005,
    "share": 0.030191016657037266,
    "count": 3552,
    "families": {
     "conv": 232.94651299999995,
     "elementwise_reduce": 97.80770200000005,
     "nccl": 8.816844
    },
    "top_kernels": [
     {
      "name": "sm90_xmma_fprop_implicit_gemm_f32f32_tf32f32_f32_nhwckrsc_nhwc_tilesize128x128x32_warpgrou",
      "ms": 202.74644099999998
     },
     {
      "name": "void at::native::elementwise_kernel<(int)128, (int)2, void at::native::gpu_kernel_impl_noc",
      "ms": 91.06681500000003
     },
     {
      "name": "void cudnn::engines_precompiled::nchwToNhwcKernel<float, float, float, (bool)0, (bool)1, (",
      "ms": 19.59811
     },
     {
      "name": "void cudnn::engines_precompiled::nhwcToNchwKernel<float, float, float, (bool)1, (bool)0, (",
      "ms": 10.601961999999999
     }
    ]
   },
   {
    "unit": "vae/up_blocks.1.upsamplers.0.resample.1/WanDistConv2d",
    "ms": 257.05724599999996,
    "share": 0.022854773368062926,
    "count": 2860,
    "families": {
     "conv": 150.03594600000002,
     "elementwise_reduce": 58.31064,
     "gather_index_cat": 39.736511,
     "nccl": 8.974149
    },
    "top_kernels": [
     {
      "name": "sm90_xmma_fprop_implicit_gemm_f32f32_tf32f32_f32_nhwckrsc_nhwc_tilesize128x128x32_warpgrou",
      "ms": 114.417664
     },
     {
      "name": "void at::native::elementwise_kernel<(int)128, (int)2, void at::native::gpu_kernel_impl_noc",
      "ms": 49.445203
     },
     {
      "name": "void at::native::<unnamed>::CatArrayBatchedCopy_alignedK_contig<at::native::<unnamed>::Opa",
      "ms": 39.736511
     },
     {
      "name": "void cudnn::engines_precompiled::nchwToNhwcKernel<float, float, float, (bool)0, (bool)1, (",
      "ms": 24.847381
     }
    ]
   },
   {
    "unit": "vae/up_blocks.1.resnets.0.conv2/WanDistCausalConv3d",
    "ms": 178.59214699999998,
    "share": 0.015878498305396064,
    "count": 3552,
    "families": {
     "conv": 131.865488,
     "elementwise_reduce": 35.18852900000001,
     "nccl": 11.538129999999999
    },
    "top_kernels": [
     {
      "name": "sm90_xmma_fprop_implicit_gemm_f32f32_tf32f32_f32_nhwckrsc_nhwc_tilesize128x128x32_warpgrou",
      "ms": 121.20339200000001
     },
     {
      "name": "void at::native::elementwise_kernel<(int)128, (int)2, void at::native::gpu_kernel_impl_noc",
      "ms": 32.78640700000001
     },
     {
      "name": "ncclDevKernel_SendRecv(ncclDevKernelArgsStorage<(unsigned long)4096>)",
      "ms": 11.538129999999999
     },
     {
      "name": "void cudnn::engines_precompiled::nchwToNhwcKernel<float, float, float, (bool)0, (bool)1, (",
      "ms": 8.047046
     }
    ]
   },
   {
    "unit": "vae/up_blocks.1.resnets.1.conv1/WanDistCausalConv3d",
    "ms": 176.06921799999995,
    "share": 0.01565418651720117,
    "count": 3552,
    "families": {
     "conv": 131.981293,
     "elementwise_reduce": 35.182548,
     "nccl": 8.905377
    },
    "top_kernels": [
     {
      "name": "sm90_xmma_fprop_implicit_gemm_f32f32_tf32f32_f32_nhwckrsc_nhwc_tilesize128x128x32_warpgrou",
      "ms": 121.318688
     },
     {
      "name": "void at::native::elementwise_kernel<(int)128, (int)2, void at::native::gpu_kernel_impl_noc",
      "ms": 32.778537
     },
     {
      "name": "ncclDevKernel_SendRecv(ncclDevKernelArgsStorage<(unsigned long)4096>)",
      "ms": 8.905377
     },
     {
      "name": "void cudnn::engines_precompiled::nchwToNhwcKernel<float, float, float, (bool)0, (bool)1, (",
      "ms": 8.05008
     }
    ]
   },
   {
    "unit": "vae/up_blocks.1.resnets.2.conv2/WanDistCausalConv3d",
    "ms": 176.03687100000002,
    "share": 0.01565131056888367,
    "count": 3552,
    "families": {
     "conv": 132.711381,
     "elementwise_reduce": 35.159572999999995,
     "nccl": 8.165917
    },
    "top_kernels": [
     {
      "name": "sm90_xmma_fprop_implicit_gemm_f32f32_tf32f32_f32_nhwckrsc_nhwc_tilesize128x128x32_warpgrou",
      "ms": 122.05062700000002
     },
     {
      "name": "void at::native::elementwise_kernel<(int)128, (int)2, void at::native::gpu_kernel_impl_noc",
      "ms": 32.745383
     },
     {
      "name": "ncclDevKernel_SendRecv(ncclDevKernelArgsStorage<(unsigned long)4096>)",
      "ms": 8.165917
     },
     {
      "name": "void cudnn::engines_precompiled::nchwToNhwcKernel<float, float, float, (bool)0, (bool)1, (",
      "ms": 8.046149999999999
     }
    ]
   },
   {
    "unit": "vae/up_blocks.1.resnets.2.conv1/WanDistCausalConv3d",
    "ms": 174.95346599999993,
    "share": 0.015554985815833028,
    "count": 3552,
    "families": {
     "conv": 131.869937,
     "elementwise_reduce": 35.16480600000002,
     "nccl": 7.918723
    },
    "top_kernels": [
     {
      "name": "sm90_xmma_fprop_implicit_gemm_f32f32_tf32f32_f32_nhwckrsc_nhwc_tilesize128x128x32_warpgrou",
      "ms": 121.213833
     },
     {
      "name": "void at::native::elementwise_kernel<(int)128, (int)2, void at::native::gpu_kernel_impl_noc",
      "ms": 32.756800000000005
     },
     {
      "name": "void cudnn::engines_precompiled::nchwToNhwcKernel<float, float, float, (bool)0, (bool)1, (",
      "ms": 8.034846
     },
     {
      "name": "ncclDevKernel_SendRecv(ncclDevKernelArgsStorage<(unsigned long)4096>)",
      "ms": 7.918723
     }
    ]
   },
   {
    "unit": "vae/up_blocks.1.resnets.1.conv2/WanDistCausalConv3d",
    "ms": 174.859075,
    "share": 0.015546593580459187,
    "count": 3552,
    "families": {
     "conv": 131.72852799999998,
     "elementwise_reduce": 35.163171,
     "nccl": 7.967376000000001
    },
    "top_kernels": [
     {
      "name": "sm90_xmma_fprop_implicit_gemm_f32f32_tf32f32_f32_nhwckrsc_nhwc_tilesize128x128x32_warpgrou",
      "ms": 121.06583099999999
     },
     {
      "name": "void at::native::elementwise_kernel<(int)128, (int)2, void at::native::gpu_kernel_impl_noc",
      "ms": 32.753048
     },
     {
      "name": "void cudnn::engines_precompiled::nchwToNhwcKernel<float, float, float, (bool)0, (bool)1, (",
      "ms": 8.043776
     },
     {
      "name": "ncclDevKernel_SendRecv(ncclDevKernelArgsStorage<(unsigned long)4096>)",
      "ms": 7.967376000000001
     }
    ]
   },
   {
    "unit": "vae/up_blocks.3.resnets.0.norm1/WanRMS_norm",
    "ms": 114.66221499999997,
    "share": 0.010194534402290706,
    "count": 1416,
    "families": {
     "elementwise_reduce": 114.66221499999997
    },
    "top_kernels": [
     {
      "name": "void at::native::elementwise_kernel<(int)128, (int)2, void at::native::gpu_kernel_impl_noc",
      "ms": 65.56950499999999
     },
     {
      "name": "void at::native::vectorized_elementwise_kernel<(int)4, at::native::CUDAFunctorOnSelf_add<f",
      "ms": 18.940599
     },
     {
      "name": "void at::native::vectorized_elementwise_kernel<(int)4, at::native::AUnaryFunctor<float, fl",
      "ms": 18.895929000000002
     },
     {
      "name": "void at::native::reduce_kernel<(int)128, (int)4, at::native::ReduceOp<float, at::native::N",
      "ms": 10.916980999999998
     }
    ]
   },
   {
    "unit": "vae/up_blocks.3.resnets.0.norm2/WanRMS_norm",
    "ms": 110.66696399999996,
    "share": 0.00983931953255104,
    "count": 1416,
    "families": {
     "elementwise_reduce": 110.66696399999996
    },
    "top_kernels": [
     {
      "name": "void at::native::elementwise_kernel<(int)128, (int)2, void at::native::gpu_kernel_impl_noc",
      "ms": 61.661190000000005
     },
     {
      "name": "void at::native::vectorized_elementwise_kernel<(int)4, at::native::CUDAFunctorOnSelf_add<f",
      "ms": 18.949757
     },
     {
      "name": "void at::native::vectorized_elementwise_kernel<(int)4, at::native::AUnaryFunctor<float, fl",
      "ms": 18.862072
     },
     {
      "name": "void at::native::reduce_kernel<(int)128, (int)4, at::native::ReduceOp<float, at::native::N",
      "ms": 10.854391
     }
    ]
   },
   {
    "unit": "vae/up_blocks.3.resnets.1.norm2/WanRMS_norm",
    "ms": 110.61112099999997,
    "share": 0.009834354571908802,
    "count": 1416,
    "families": {
     "elementwise_reduce": 110.61112099999997
    },
    "top_kernels": [
     {
      "name": "void at::native::elementwise_kernel<(int)128, (int)2, void at::native::gpu_kernel_impl_noc",
      "ms": 61.620283
     },
     {
      "name": "void at::native::vectorized_elementwise_kernel<(int)4, at::native::CUDAFunctorOnSelf_add<f",
      "ms": 18.940989
     },
     {
      "name": "void at::native::vectorized_elementwise_kernel<(int)4, at::native::AUnaryFunctor<float, fl",
      "ms": 18.856382
     },
     {
      "name": "void at::native::reduce_kernel<(int)128, (int)4, at::native::ReduceOp<float, at::native::N",
      "ms": 10.858234000000001
     }
    ]
   },
   {
    "unit": "vae/up_blocks.3.resnets.2.norm2/WanRMS_norm",
    "ms": 110.53252399999998,
    "share": 0.009827366569623857,
    "count": 1416,
    "families": {
     "elementwise_reduce": 110.53252399999998
    },
    "top_kernels": [
     {
      "name": "void at::native::elementwise_kernel<(int)128, (int)2, void at::native::gpu_kernel_impl_noc",
      "ms": 61.545122
     },
     {
      "name": "void at::native::vectorized_elementwise_kernel<(int)4, at::native::CUDAFunctorOnSelf_add<f",
      "ms": 18.942712999999998
     },
     {
      "name": "void at::native::vectorized_elementwise_kernel<(int)4, at::native::AUnaryFunctor<float, fl",
      "ms": 18.854523
     },
     {
      "name": "void at::native::reduce_kernel<(int)128, (int)4, at::native::ReduceOp<float, at::native::N",
      "ms": 10.854194
     }
    ]
   },
   {
    "unit": "vae/up_blocks.3.resnets.2.norm1/WanRMS_norm",
    "ms": 110.48483699999997,
    "share": 0.009823126753028284,
    "count": 1416,
    "families": {
     "elementwise_reduce": 110.48483699999997
    },
    "top_kernels": [
     {
      "name": "void at::native::elementwise_kernel<(int)128, (int)2, void at::native::gpu_kernel_impl_noc",
      "ms": 61.582281
     },
     {
      "name": "void at::native::vectorized_elementwise_kernel<(int)4, at::native::CUDAFunctorOnSelf_add<f",
      "ms": 18.947157
     },
     {
      "name": "void at::native::vectorized_elementwise_kernel<(int)4, at::native::AUnaryFunctor<float, fl",
      "ms": 18.853079
     },
     {
      "name": "void at::native::reduce_kernel<(int)128, (int)4, at::native::ReduceOp<float, at::native::N",
      "ms": 10.747052
     }
    ]
   },
   {
    "unit": "vae/norm_out/WanRMS_norm",
    "ms": 110.42763099999999,
    "share": 0.00981804060904426,
    "count": 1416,
    "families": {
     "elementwise_reduce": 110.42763099999999
    },
    "top_kernels": [
     {
      "name": "void at::native::elementwise_kernel<(int)128, (int)2, void at::native::gpu_kernel_impl_noc",
      "ms": 61.523691
     },
     {
      "name": "void at::native::vectorized_elementwise_kernel<(int)4, at::native::CUDAFunctorOnSelf_add<f",
      "ms": 18.938425
     },
     {
      "name": "void at::native::vectorized_elementwise_kernel<(int)4, at::native::AUnaryFunctor<float, fl",
      "ms": 18.879196
     },
     {
      "name": "void at::native::reduce_kernel<(int)128, (int)4, at::native::ReduceOp<float, at::native::N",
      "ms": 10.733934999999999
     }
    ]
   },
   {
    "unit": "vae/up_blocks.3.resnets.1.norm1/WanRMS_norm",
    "ms": 110.26822999999997,
    "share": 0.009803868381704505,
    "count": 1416,
    "families": {
     "elementwise_reduce": 110.26822999999997
    },
    "top_kernels": [
     {
      "name": "void at::native::elementwise_kernel<(int)128, (int)2, void at::native::gpu_kernel_impl_noc",
      "ms": 61.640256
     },
     {
      "name": "void at::native::vectorized_elementwise_kernel<(int)4, at::native::CUDAFunctorOnSelf_add<f",
      "ms": 18.930456
     },
     {
      "name": "void at::native::vectorized_elementwise_kernel<(int)4, at::native::AUnaryFunctor<float, fl",
      "ms": 18.852415
     },
     {
      "name": "void at::native::reduce_kernel<(int)128, (int)4, at::native::ReduceOp<float, at::native::N",
      "ms": 10.506446
     }
    ]
   },
   {
    "unit": "vae/up_blocks.1.resnets.0.conv1/WanDistCausalConv3d",
    "ms": 102.10345299999999,
    "share": 0.009077943978329499,
    "count": 3548,
    "families": {
     "conv": 67.582461,
     "elementwise_reduce": 20.088487000000008,
     "nccl": 14.432505000000003
    },
    "top_kernels": [
     {
      "name": "sm90_xmma_fprop_implicit_gemm_f32f32_tf32f32_f32_nhwckrsc_nhwc_tilesize128x128x32_warpgrou",
      "ms": 60.412189
     },
     {
      "name": "void at::native::elementwise_kernel<(int)128, (int)2, void at::native::gpu_kernel_impl_noc",
      "ms": 18.801377000000002
     },
     {
      "name": "ncclDevKernel_SendRecv(ncclDevKernelArgsStorage<(unsigned long)4096>)",
      "ms": 14.432505000000003
     },
     {
      "name": "void cudnn::engines_precompiled::nchwToNhwcKernel<float, float, float, (bool)0, (bool)1, (",
      "ms": 4.402744000000001
     }
    ]
   },
   {
    "unit": "vae/up_blocks.2.upsamplers.0.resample.0/WanUpsample",
    "ms": 79.36550100000001,
    "share": 0.007056329151669865,
    "count": 468,
    "families": {
     "gather_index_cat": 63.91249800000001,
     "elementwise_reduce": 15.453003
    },
    "top_kernels": [
     {
      "name": "void at::native::<unnamed>::upsample_nearest2d_out_frame<float, &at::native::nearest_neigh",
      "ms": 63.91249800000001
     },
     {
      "name": "void at::native::elementwise_kernel<(int)128, (int)2, void at::native::gpu_kernel_impl_noc",
      "ms": 15.453003
     }
    ]
   },
   {
    "unit": "ncclAllReduce",
    "ms": 71.91083900000001,
    "share": 0.006393540558091333,
    "count": 160,
    "families": {
     "nccl": 71.91083900000001
    },
    "top_kernels": [
     {
      "name": "ncclDevKernel_AllReduce_Sum_f32_RING_LL(ncclDevKernelArgsStorage<(unsigned long)4096>)",
      "ms": 71.91083900000001
     }
    ]
   },
   {
    "unit": "vae/up_blocks.2.resnets.0.norm1/WanRMS_norm",
    "ms": 58.407101999999995,
    "share": 0.005192933092013811,
    "count": 1416,
    "families": {
     "elementwise_reduce": 58.407101999999995
    },
    "top_kernels": [
     {
      "name": "void at::native::elementwise_kernel<(int)128, (int)2, void at::native::gpu_kernel_impl_noc",
      "ms": 33.153189
     },
     {
      "name": "void at::native::vectorized_elementwise_kernel<(int)4, at::native::CUDAFunctorOnSelf_add<f",
      "ms": 9.473879
     },
     {
      "name": "void at::native::vectorized_elementwise_kernel<(int)4, at::native::AUnaryFunctor<float, fl",
      "ms": 9.455371999999999
     },
     {
      "name": "void at::native::reduce_kernel<(int)128, (int)4, at::native::ReduceOp<float, at::native::N",
      "ms": 6.109812999999999
     }
    ]
   },
   {
    "unit": "vae/up_blocks.2.resnets.0.norm2/WanRMS_norm",
    "ms": 56.40314100000001,
    "share": 0.005014762372432397,
    "count": 1416,
    "families": {
     "elementwise_reduce": 56.40314100000001
    },
    "top_kernels": [
     {
      "name": "void at::native::elementwise_kernel<(int)128, (int)2, void at::native::gpu_kernel_impl_noc",
      "ms": 31.057309999999998
     },
     {
      "name": "void at::native::vectorized_elementwise_kernel<(int)4, at::native::CUDAFunctorOnSelf_add<f",
      "ms": 9.482281
     },
     {
      "name": "void at::native::vectorized_elementwise_kernel<(int)4, at::native::AUnaryFunctor<float, fl",
      "ms": 9.470088
     },
     {
      "name": "void at::native::reduce_kernel<(int)128, (int)4, at::native::ReduceOp<float, at::native::N",
      "ms": 6.168757
     }
    ]
   },
   {
    "unit": "vae/up_blocks.2.resnets.1.norm2/WanRMS_norm",
    "ms": 56.357393999999985,
    "share": 0.005010695039830268,
    "count": 1416,
    "families": {
     "elementwise_reduce": 56.357393999999985
    },
    "top_kernels": [
     {
      "name": "void at::native::elementwise_kernel<(int)128, (int)2, void at::native::gpu_kernel_impl_noc",
      "ms": 31.032254000000002
     },
     {
      "name": "void at::native::vectorized_elementwise_kernel<(int)4, at::native::AUnaryFunctor<float, fl",
      "ms": 9.470635
     },
     {
      "name": "void at::native::vectorized_elementwise_kernel<(int)4, at::native::CUDAFunctorOnSelf_add<f",
      "ms": 9.460043
     },
     {
      "name": "void at::native::reduce_kernel<(int)128, (int)4, at::native::ReduceOp<float, at::native::N",
      "ms": 6.168762999999999
     }
    ]
   },
   {
    "unit": "vae/up_blocks.2.resnets.2.norm2/WanRMS_norm",
    "ms": 56.34062200000001,
    "share": 0.0050092038534704465,
    "count": 1416,
    "families": {
     "elementwise_reduce": 56.34062200000001
    },
    "top_kernels": [
     {
      "name": "void at::native::elementwise_kernel<(int)128, (int)2, void at::native::gpu_kernel_impl_noc",
      "ms": 31.019403999999998
     },
     {
      "name": "void at::native::vectorized_elementwise_kernel<(int)4, at::native::CUDAFunctorOnSelf_add<f",
      "ms": 9.471923
     },
     {
      "name": "void at::native::vectorized_elementwise_kernel<(int)4, at::native::AUnaryFunctor<float, fl",
      "ms": 9.464652999999998
     },
     {
      "name": "void at::native::reduce_kernel<(int)128, (int)4, at::native::ReduceOp<float, at::native::N",
      "ms": 6.154049
     }
    ]
   },
   {
    "unit": "vae/up_blocks.2.resnets.2.norm1/WanRMS_norm",
    "ms": 56.241997000000005,
    "share": 0.0050004351762263696,
    "count": 1416,
    "families": {
     "elementwise_reduce": 56.241997000000005
    },
    "top_kernels": [
     {
      "name": "void at::native::elementwise_kernel<(int)128, (int)2, void at::native::gpu_kernel_impl_noc",
      "ms": 30.998700999999997
     },
     {
      "name": "void at::native::vectorized_elementwise_kernel<(int)4, at::native::AUnaryFunctor<float, fl",
      "ms": 9.488399000000001
     },
     {
      "name": "void at::native::vectorized_elementwise_kernel<(int)4, at::native::CUDAFunctorOnSelf_add<f",
      "ms": 9.486678000000001
     },
     {
      "name": "void at::native::reduce_kernel<(int)128, (int)4, at::native::ReduceOp<float, at::native::N",
      "ms": 6.0509390000000005
     }
    ]
   },
   {
    "unit": "vae/up_blocks.2.resnets.1.norm1/WanRMS_norm",
    "ms": 56.004808999999995,
    "share": 0.004979346963114398,
    "count": 1416,
    "families": {
     "elementwise_reduce": 56.004808999999995
    },
    "top_kernels": [
     {
      "name": "void at::native::elementwise_kernel<(int)128, (int)2, void at::native::gpu_kernel_impl_noc",
      "ms": 31.040729
     },
     {
      "name": "void at::native::vectorized_elementwise_kernel<(int)4, at::native::AUnaryFunctor<float, fl",
      "ms": 9.475083999999999
     },
     {
      "name": "void at::native::vectorized_elementwise_kernel<(int)4, at::native::CUDAFunctorOnSelf_add<f",
      "ms": 9.474029999999999
     },
     {
      "name": "void at::native::reduce_kernel<(int)128, (int)4, at::native::ReduceOp<float, at::native::N",
      "ms": 5.793652000000001
     }
    ]
   },
   {
    "unit": "vae/up_blocks.0.upsamplers.0.resample.1/WanDistConv2d",
    "ms": 50.20224699999999,
    "share": 0.0044634453827164895,
    "count": 2616,
    "families": {
     "conv": 22.139243999999998,
     "nccl": 15.014906000000002,
     "elementwise_reduce": 8.080747999999998,
     "gather_index_cat": 4.9673490000000005
    },
    "top_kernels": [
     {
      "name": "sm90_xmma_fprop_implicit_gemm_f32f32_tf32f32_f32_nhwckrsc_nhwc_tilesize128x128x32_warpgrou",
      "ms": 16.748653
     },
     {
      "name": "ncclDevKernel_SendRecv(ncclDevKernelArgsStorage<(unsigned long)4096>)",
      "ms": 15.014906000000002
     },
     {
      "name": "void at::native::elementwise_kernel<(int)128, (int)2, void at::native::gpu_kernel_impl_noc",
      "ms": 6.772901999999999
     },
     {
      "name": "void at::native::<unnamed>::CatArrayBatchedCopy_alignedK_contig<at::native::<unnamed>::Opa",
      "ms": 4.9673490000000005
     }
    ]
   },
   {
    "unit": "vae/up_blocks.1.upsamplers.0.time_conv/WanDistCausalConv3d",
    "ms": 44.52632200000001,
    "share": 0.003958803006173164,
    "count": 1624,
    "families": {
     "conv": 29.039908000000004,
     "elementwise_reduce": 15.486414
    },
    "top_kernels": [
     {
      "name": "sm90_xmma_fprop_implicit_gemm_f32f32_tf32f32_f32_nhwckrsc_nhwc_tilesize128x128x32_warpgrou",
      "ms": 15.666604999999999
     },
     {
      "name": "void at::native::elementwise_kernel<(int)128, (int)2, void at::native::gpu_kernel_impl_noc",
      "ms": 15.409358
     },
     {
      "name": "void cudnn::engines_precompiled::nchwToNhwcKernel<float, float, float, (bool)0, (bool)1, (",
      "ms": 7.805692
     },
     {
      "name": "void cudnn::engines_precompiled::nhwcToNchwKernel<float, float, float, (bool)1, (bool)0, (",
      "ms": 5.567610999999999
     }
    ]
   },
   {
    "unit": "vae/mid_block.resnets.1.conv1/WanDistCausalConv3d",
    "ms": 43.417041,
    "share": 0.003860177636723363,
    "count": 3584,
    "families": {
     "conv": 20.319267999999997,
     "nccl": 16.19358,
     "elementwise_reduce": 6.904193
    },
    "top_kernels": [
     {
      "name": "sm90_xmma_fprop_implicit_gemm_f32f32_tf32f32_f32_nhwckrsc_nhwc_tilesize128x128x32_warpgrou",
      "ms": 17.345585999999997
     },
     {
      "name": "ncclDevKernel_SendRecv(ncclDevKernelArgsStorage<(unsigned long)4096>)",
      "ms": 16.19358
     },
     {
      "name": "void at::native::elementwise_kernel<(int)128, (int)2, void at::native::gpu_kernel_impl_noc",
      "ms": 6.302463000000001
     },
     {
      "name": "void cudnn::engines_precompiled::nchwToNhwcKernel<float, float, float, (bool)0, (bool)1, (",
      "ms": 2.5263500000000003
     }
    ]
   },
   {
    "unit": "vae/up_blocks.1.upsamplers.0.resample.0/WanUpsample",
    "ms": 41.397068999999995,
    "share": 0.0036805833907403774,
    "count": 468,
    "families": {
     "gather_index_cat": 33.555336999999994,
     "elementwise_reduce": 7.841732
    },
    "top_kernels": [
     {
      "name": "void at::native::<unnamed>::upsample_nearest2d_out_frame<float, &at::native::nearest_neigh",
      "ms": 33.555336999999994
     },
     {
      "name": "void at::native::elementwise_kernel<(int)128, (int)2, void at::native::gpu_kernel_impl_noc",
      "ms": 7.841732
     }
    ]
   },
   {
    "unit": "vae/mid_block.resnets.0.conv2/WanDistCausalConv3d",
    "ms": 40.68748500000001,
    "share": 0.0036174947917689123,
    "count": 3584,
    "families": {
     "conv": 20.119226,
     "nccl": 13.626018,
     "elementwise_reduce": 6.942241
    },
    "top_kernels": [
     {
      "name": "sm90_xmma_fprop_implicit_gemm_f32f32_tf32f32_f32_nhwckrsc_nhwc_tilesize128x128x32_warpgrou",
      "ms": 17.139024
     },
     {
      "name": "ncclDevKernel_SendRecv(ncclDevKernelArgsStorage<(unsigned long)4096>)",
      "ms": 13.626018
     },
     {
      "name": "void at::native::elementwise_kernel<(int)128, (int)2, void at::native::gpu_kernel_impl_noc",
      "ms": 6.341342
     },
     {
      "name": "void cudnn::engines_precompiled::nchwToNhwcKernel<float, float, float, (bool)0, (bool)1, (",
      "ms": 2.5330649999999997
     }
    ]
   },
   {
    "unit": "vae/up_blocks.0.resnets.2.conv2/WanDistCausalConv3d",
    "ms": 39.562419999999996,
    "share": 0.0035174660783229587,
    "count": 3584,
    "families": {
     "conv": 20.280185999999997,
     "nccl": 12.368631999999998,
     "elementwise_reduce": 6.913602000000002
    },
    "top_kernels": [
     {
      "name": "sm90_xmma_fprop_implicit_gemm_f32f32_tf32f32_f32_nhwckrsc_nhwc_tilesize128x128x32_warpgrou",
      "ms": 17.323634
     },
     {
      "name": "ncclDevKernel_SendRecv(ncclDevKernelArgsStorage<(unsigned long)4096>)",
      "ms": 12.368631999999998
     },
     {
      "name": "void at::native::elementwise_kernel<(int)128, (int)2, void at::native::gpu_kernel_impl_noc",
      "ms": 6.3134070000000015
     },
     {
      "name": "void cudnn::engines_precompiled::nchwToNhwcKernel<float, float, float, (bool)0, (bool)1, (",
      "ms": 2.512582
     }
    ]
   },
   {
    "unit": "vae/up_blocks.0.resnets.2.conv1/WanDistCausalConv3d",
    "ms": 39.30174599999999,
    "share": 0.003494289741979004,
    "count": 3584,
    "families": {
     "conv": 20.273411,
     "nccl": 12.119186000000001,
     "elementwise_reduce": 6.909148999999999
    },
    "top_kernels": [
     {
      "name": "sm90_xmma_fprop_implicit_gemm_f32f32_tf32f32_f32_nhwckrsc_nhwc_tilesize128x128x32_warpgrou",
      "ms": 17.317946
     },
     {
      "name": "ncclDevKernel_SendRecv(ncclDevKernelArgsStorage<(unsigned long)4096>)",
      "ms": 12.119186000000001
     },
     {
      "name": "void at::native::elementwise_kernel<(int)128, (int)2, void at::native::gpu_kernel_impl_noc",
      "ms": 6.306234
     },
     {
      "name": "void cudnn::engines_precompiled::nchwToNhwcKernel<float, float, float, (bool)0, (bool)1, (",
      "ms": 2.5088079999999997
     }
    ]
   },
   {
    "unit": "vae/mid_block.resnets.0.conv1/WanDistCausalConv3d",
    "ms": 39.217802,
    "share": 0.003486826341800787,
    "count": 3584,
    "families": {
     "conv": 20.079129000000002,
     "nccl": 12.197712,
     "elementwise_reduce": 6.940961
    },
    "top_kernels": [
     {
      "name": "sm90_xmma_fprop_implicit_gemm_f32f32_tf32f32_f32_nhwckrsc_nhwc_tilesize128x128x32_warpgrou",
      "ms": 17.111116
     },
     {
      "name": "ncclDevKernel_SendRecv(ncclDevKernelArgsStorage<(unsigned long)4096>)",
      "ms": 12.197712
     },
     {
      "name": "void at::native::elementwise_kernel<(int)128, (int)2, void at::native::gpu_kernel_impl_noc",
      "ms": 6.334557
     },
     {
      "name": "void cudnn::engines_precompiled::nchwToNhwcKernel<float, float, float, (bool)0, (bool)1, (",
      "ms": 2.5236249999999996
     }
    ]
   },
   {
    "unit": "vae/up_blocks.3.resnets.1.nonlinearity/SiLU",
    "ms": 38.404892,
    "share": 0.003414551103083602,
    "count": 472,
    "families": {
     "elementwise_reduce": 38.404892
    },
    "top_kernels": [
     {
      "name": "void at::native::vectorized_elementwise_kernel<(int)4, at::native::<unnamed>::silu_kernel(",
      "ms": 38.404892
     }
    ]
   },
   {
    "unit": "vae/up_blocks.3.resnets.0.nonlinearity/SiLU",
    "ms": 38.371984,
    "share": 0.003411625276662836,
    "count": 472,
    "families": {
     "elementwise_reduce": 38.371984
    },
    "top_kernels": [
     {
      "name": "void at::native::vectorized_elementwise_kernel<(int)4, at::native::<unnamed>::silu_kernel(",
      "ms": 38.371984
     }
    ]
   },
   {
    "unit": "vae/up_blocks.3.resnets.2.nonlinearity/SiLU",
    "ms": 38.350818000000004,
    "share": 0.0034097434229487873,
    "count": 472,
    "families": {
     "elementwise_reduce": 38.350818000000004
    },
    "top_kernels": [
     {
      "name": "void at::native::vectorized_elementwise_kernel<(int)4, at::native::<unnamed>::silu_kernel(",
      "ms": 38.350818000000004
     }
    ]
   },
   {
    "unit": "ncclAllGather",
    "ms": 36.260711,
    "share": 0.0032239135249656665,
    "count": 480,
    "families": {
     "nccl": 36.260711
    },
    "top_kernels": [
     {
      "name": "ncclDevKernel_AllGather_RING_LL(ncclDevKernelArgsStorage<(unsigned long)4096>)",
      "ms": 36.260711
     }
    ]
   },
   {
    "unit": "vae/up_blocks.0.resnets.1.conv2/WanDistCausalConv3d",
    "ms": 36.129166000000005,
    "share": 0.0032122179543895243,
    "count": 3584,
    "families": {
     "conv": 20.307037,
     "nccl": 8.920933999999999,
     "elementwise_reduce": 6.901194999999999
    },
    "top_kernels": [
     {
      "name": "sm90_xmma_fprop_implicit_gemm_f32f32_tf32f32_f32_nhwckrsc_nhwc_tilesize128x128x32_warpgrou",
      "ms": 17.336145
     },
     {
      "name": "ncclDevKernel_SendRecv(ncclDevKernelArgsStorage<(unsigned long)4096>)",
      "ms": 8.920933999999999
     },
     {
      "name": "void at::native::elementwise_kernel<(int)128, (int)2, void at::native::gpu_kernel_impl_noc",
      "ms": 6.301224999999999
     },
     {
      "name": "void cudnn::engines_precompiled::nchwToNhwcKernel<float, float, float, (bool)0, (bool)1, (",
      "ms": 2.5266669999999998
     }
    ]
   },
   {
    "unit": "vae/mid_block.resnets.1.conv2/WanDistCausalConv3d",
    "ms": 33.931905000000015,
    "share": 0.003016861071956095,
    "count": 3584,
    "families": {
     "conv": 20.338051,
     "elementwise_reduce": 6.911745999999998,
     "nccl": 6.6821079999999995
    },
    "top_kernels": [
     {
      "name": "sm90_xmma_fprop_implicit_gemm_f32f32_tf32f32_f32_nhwckrsc_nhwc_tilesize128x128x32_warpgrou",
      "ms": 17.366227000000002
     },
     {
      "name": "ncclDevKernel_SendRecv(ncclDevKernelArgsStorage<(unsigned long)4096>)",
      "ms": 6.6821079999999995
     },
     {
      "name": "void at::native::elementwise_kernel<(int)128, (int)2, void at::native::gpu_kernel_impl_noc",
      "ms": 6.310556999999998
     },
     {
      "name": "void cudnn::engines_precompiled::nchwToNhwcKernel<float, float, float, (bool)0, (bool)1, (",
      "ms": 2.529163
     }
    ]
   },
   {
    "unit": "vae/up_blocks.0.resnets.1.conv1/WanDistCausalConv3d",
    "ms": 30.752403999999995,
    "share": 0.002734173943274533,
    "count": 3584,
    "families": {
     "conv": 20.350469,
     "elementwise_reduce": 6.911968999999997,
     "nccl": 3.489966
    },
    "top_kernels": [
     {
      "name": "sm90_xmma_fprop_implicit_gemm_f32f32_tf32f32_f32_nhwckrsc_nhwc_tilesize128x128x32_warpgrou",
      "ms": 17.393178
     },
     {
      "name": "void at::native::elementwise_kernel<(int)128, (int)2, void at::native::gpu_kernel_impl_noc",
      "ms": 6.311261999999998
     },
     {
      "name": "ncclDevKernel_SendRecv(ncclDevKernelArgsStorage<(unsigned long)4096>)",
      "ms": 3.489966
     },
     {
      "name": "void cudnn::engines_precompiled::nchwToNhwcKernel<float, float, float, (bool)0, (bool)1, (",
      "ms": 2.513288
     }
    ]
   },
   {
    "unit": "vae/up_blocks.0.resnets.0.conv1/WanDistCausalConv3d",
    "ms": 30.087856000000002,
    "share": 0.0026750894624106895,
    "count": 3584,
    "families": {
     "conv": 20.309412000000002,
     "elementwise_reduce": 6.906617999999998,
     "nccl": 2.871826
    },
    "top_kernels": [
     {
      "name": "sm90_xmma_fprop_implicit_gemm_f32f32_tf32f32_f32_nhwckrsc_nhwc_tilesize128x128x32_warpgrou",
      "ms": 17.343065000000003
     },
     {
      "name": "void at::native::elementwise_kernel<(int)128, (int)2, void at::native::gpu_kernel_impl_noc",
      "ms": 6.305650999999999
     },
     {
      "name": "ncclDevKernel_SendRecv(ncclDevKernelArgsStorage<(unsigned long)4096>)",
      "ms": 2.871826
     },
     {
      "name": "void cudnn::engines_precompiled::nchwToNhwcKernel<float, float, float, (bool)0, (bool)1, (",
      "ms": 2.524393
     }
    ]
   },
   {
    "unit": "vae/up_blocks.0.resnets.0.conv2/WanDistCausalConv3d",
    "ms": 29.440513000000017,
    "share": 0.0026175346656227335,
    "count": 3584,
    "families": {
     "conv": 20.353343000000002,
     "elementwise_reduce": 6.894613,
     "nccl": 2.192557
    },
    "top_kernels": [
     {
      "name": "sm90_xmma_fprop_implicit_gemm_f32f32_tf32f32_f32_nhwckrsc_nhwc_tilesize128x128x32_warpgrou",
      "ms": 17.382061999999998
     },
     {
      "name": "void at::native::elementwise_kernel<(int)128, (int)2, void at::native::gpu_kernel_impl_noc",
      "ms": 6.291699999999998
     },
     {
      "name": "void cudnn::engines_precompiled::nchwToNhwcKernel<float, float, float, (bool)0, (bool)1, (",
      "ms": 2.5258390000000004
     },
     {
      "name": "ncclDevKernel_SendRecv(ncclDevKernelArgsStorage<(unsigned long)4096>)",
      "ms": 2.192557
     }
    ]
   },
   {
    "unit": "vae/up_blocks.2.resnets.1.nonlinearity/SiLU",
    "ms": 19.427414,
    "share": 0.001727277293313618,
    "count": 472,
    "families": {
     "elementwise_reduce": 19.427414
    },
    "top_kernels": [
     {
      "name": "void at::native::vectorized_elementwise_kernel<(int)4, at::native::<unnamed>::silu_kernel(",
      "ms": 19.427414
     }
    ]
   },
   {
    "unit": "vae/up_blocks.2.resnets.0.nonlinearity/SiLU",
    "ms": 19.416315,
    "share": 0.001726290489270708,
    "count": 472,
    "families": {
     "elementwise_reduce": 19.416315
    },
    "top_kernels": [
     {
      "name": "void at::native::vectorized_elementwise_kernel<(int)4, at::native::<unnamed>::silu_kernel(",
      "ms": 19.416315
     }
    ]
   },
   {
    "unit": "vae/up_blocks.2.resnets.2.nonlinearity/SiLU",
    "ms": 19.416145,
    "share": 0.0017262753746939627,
    "count": 472,
    "families": {
     "elementwise_reduce": 19.416145
    },
    "top_kernels": [
     {
      "name": "void at::native::vectorized_elementwise_kernel<(int)4, at::native::<unnamed>::silu_kernel(",
      "ms": 19.416145
     }
    ]
   },
   {
    "unit": "vae/nonlinearity/SiLU",
    "ms": 19.184411,
    "share": 0.001705672072767688,
    "count": 236,
    "families": {
     "elementwise_reduce": 19.184411
    },
    "top_kernels": [
     {
      "name": "void at::native::vectorized_elementwise_kernel<(int)4, at::native::<unnamed>::silu_kernel(",
      "ms": 19.184411
     }
    ]
   },
   {
    "unit": "vae/up_blocks.1.resnets.2.norm1/WanRMS_norm",
    "ms": 15.906769000000004,
    "share": 0.0014142592989311378,
    "count": 1416,
    "families": {
     "elementwise_reduce": 15.906769000000004
    },
    "top_kernels": [
     {
      "name": "void at::native::elementwise_kernel<(int)128, (int)2, void at::native::gpu_kernel_impl_noc",
      "ms": 8.149545999999999
     },
     {
      "name": "void at::native::reduce_kernel<(int)128, (int)4, at::native::ReduceOp<float, at::native::N",
      "ms": 2.729518
     },
     {
      "name": "void at::native::vectorized_elementwise_kernel<(int)4, at::native::CUDAFunctorOnSelf_add<f",
      "ms": 2.445674
     },
     {
      "name": "void at::native::vectorized_elementwise_kernel<(int)4, at::native::AUnaryFunctor<float, fl",
      "ms": 2.430095
     }
    ]
   },
   {
    "unit": "vae/up_blocks.1.resnets.1.norm1/WanRMS_norm",
    "ms": 15.889918999999999,
    "share": 0.0014127611776478656,
    "count": 1416,
    "families": {
     "elementwise_reduce": 15.889918999999999
    },
    "top_kernels": [
     {
      "name": "void at::native::elementwise_kernel<(int)128, (int)2, void at::native::gpu_kernel_impl_noc",
      "ms": 8.162783000000001
     },
     {
      "name": "void at::native::reduce_kernel<(int)128, (int)4, at::native::ReduceOp<float, at::native::N",
      "ms": 2.717614
     },
     {
      "name": "void at::native::vectorized_elementwise_kernel<(int)4, at::native::CUDAFunctorOnSelf_add<f",
      "ms": 2.4369989999999997
     },
     {
      "name": "void at::native::vectorized_elementwise_kernel<(int)4, at::native::AUnaryFunctor<float, fl",
      "ms": 2.4215459999999998
     }
    ]
   },
   {
    "unit": "vae/up_blocks.1.resnets.1.norm2/WanRMS_norm",
    "ms": 15.45371,
    "share": 0.0013739781517217676,
    "count": 1416,
    "families": {
     "elementwise_reduce": 15.45371
    },
    "top_kernels": [
     {
      "name": "void at::native::elementwise_kernel<(int)128, (int)2, void at::native::gpu_kernel_impl_noc",
      "ms": 8.219246999999998
     },
     {
      "name": "void at::native::vectorized_elementwise_kernel<(int)4, at::native::CUDAFunctorOnSelf_add<f",
      "ms": 2.440492
     },
     {
      "name": "void at::native::vectorized_elementwise_kernel<(int)4, at::native::AUnaryFunctor<float, fl",
      "ms": 2.434537
     },
     {
      "name": "void at::native::reduce_kernel<(int)128, (int)4, at::native::ReduceOp<float, at::native::N",
      "ms": 2.195817
     }
    ]
   },
   {
    "unit": "vae/up_blocks.1.resnets.0.norm2/WanRMS_norm",
    "ms": 15.443824999999999,
    "share": 0.001373099283538673,
    "count": 1416,
    "families": {
     "elementwise_reduce": 15.443824999999999
    },
    "top_kernels": [
     {
      "name": "void at::native::elementwise_kernel<(int)128, (int)2, void at::native::gpu_kernel_impl_noc",
      "ms": 8.246668999999999
     },
     {
      "name": "void at::native::vectorized_elementwise_kernel<(int)4, at::native::CUDAFunctorOnSelf_add<f",
      "ms": 2.4318839999999997
     },
     {
      "name": "void at::native::vectorized_elementwise_kernel<(int)4, at::native::AUnaryFunctor<float, fl",
      "ms": 2.4138010000000003
     },
     {
      "name": "void at::native::reduce_kernel<(int)128, (int)4, at::native::ReduceOp<float, at::native::N",
      "ms": 2.192174
     }
    ]
   },
   {
    "unit": "vae/up_blocks.1.resnets.2.norm2/WanRMS_norm",
    "ms": 15.438125999999999,
    "share": 0.0013725925895806098,
    "count": 1416,
    "families": {
     "elementwise_reduce": 15.438125999999999
    },
    "top_kernels": [
     {
      "name": "void at::native::elementwise_kernel<(int)128, (int)2, void at::native::gpu_kernel_impl_noc",
      "ms": 8.220016999999999
     },
     {
      "name": "void at::native::vectorized_elementwise_kernel<(int)4, at::native::CUDAFunctorOnSelf_add<f",
      "ms": 2.433164
     },
     {
      "name": "void at::native::vectorized_elementwise_kernel<(int)4, at::native::AUnaryFunctor<float, fl",
      "ms": 2.430695
     },
     {
      "name": "void at::native::reduce_kernel<(int)128, (int)4, at::native::ReduceOp<float, at::native::N",
      "ms": 2.186441
     }
    ]
   },
   {
    "unit": "vae/mid_block.attentions.0.to_qkv/Conv2d",
    "ms": 15.34913,
    "share": 0.001364680019745235,
    "count": 480,
    "families": {
     "gemm": 9.560653,
     "elementwise_reduce": 5.788476999999999
    },
    "top_kernels": [
     {
      "name": "void cutlass::Kernel2<cutlass_80_tensorop_s1688gemm_128x128_32x3_nn_align4>(T1::Params)",
      "ms": 9.560653
     },
     {
      "name": "void at::native::elementwise_kernel<(int)128, (int)2, void at::native::gpu_kernel_impl_noc",
      "ms": 5.788476999999999
     }
    ]
   },
   {
    "unit": "vae/conv_in/WanDistCausalConv3d",
    "ms": 15.132706,
    "share": 0.0013454379188187757,
    "count": 3344,
    "families": {
     "nccl": 8.792038999999999,
     "conv": 4.094416000000001,
     "elementwise_reduce": 2.2462510000000004
    },
    "top_kernels": [
     {
      "name": "ncclDevKernel_SendRecv(ncclDevKernelArgsStorage<(unsigned long)4096>)",
      "ms": 8.792038999999999
     },
     {
      "name": "sm80_xmma_fprop_implicit_gemm_tf32f32_tf32f32_f32_nhwckrsc_nchw_tilesize128x64x32_stage5_w",
      "ms": 3.58651
     },
     {
      "name": "void at::native::elementwise_kernel<(int)128, (int)2, void at::native::gpu_kernel_impl_noc",
      "ms": 2.0717870000000005
     },
     {
      "name": "void cudnn::engines_precompiled::nchwToNhwcKernel<float, float, float, (bool)0, (bool)1, (",
      "ms": 0.507906
     }
    ]
   },
   {
    "unit": "vae/mid_block.attentions.0.proj/Conv2d",
    "ms": 9.945201,
    "share": 0.0008842206103570907,
    "count": 720,
    "families": {
     "elementwise_reduce": 6.4067810000000005,
     "gemm": 3.53842
    },
    "top_kernels": [
     {
      "name": "void at::native::elementwise_kernel<(int)128, (int)2, void at::native::gpu_kernel_impl_noc",
      "ms": 6.4067810000000005
     },
     {
      "name": "void cutlass::Kernel2<cutlass_80_tensorop_s1688gemm_64x128_32x3_nn_align4>(T1::Params)",
      "ms": 3.53842
     }
    ]
   },
   {
    "unit": "vae/up_blocks.1.resnets.0.conv_shortcut/WanDistCausalConv3d",
    "ms": 9.602701999999999,
    "share": 0.0008537692725885835,
    "count": 704,
    "families": {
     "elementwise_reduce": 5.300635000000001,
     "gemm": 4.302067
    },
    "top_kernels": [
     {
      "name": "void at::native::elementwise_kernel<(int)128, (int)2, void at::native::gpu_kernel_impl_noc",
      "ms": 5.300635000000001
     },
     {
      "name": "void cutlass::Kernel2<cutlass_80_tensorop_s1688gemm_256x64_16x4_nn_align4>(T1::Params)",
      "ms": 4.216403
     },
     {
      "name": "void cutlass::Kernel2<cutlass_80_tensorop_s1688gemm_128x64_16x6_nn_align4>(T1::Params)",
      "ms": 0.085664
     }
    ]
   },
   {
    "unit": "vae/up_blocks.0.upsamplers.0.time_conv/WanDistCausalConv3d",
    "ms": 9.397325,
    "share": 0.0008355093524227361,
    "count": 1856,
    "families": {
     "conv": 6.808858999999999,
     "elementwise_reduce": 2.5884660000000004
    },
    "top_kernels": [
     {
      "name": "sm90_xmma_fprop_implicit_gemm_f32f32_tf32f32_f32_nhwckrsc_nhwc_tilesize128x128x32_warpgrou",
      "ms": 3.106637
     },
     {
      "name": "void cudnn::engines_precompiled::nchwToNhwcKernel<float, float, float, (bool)0, (bool)1, (",
      "ms": 2.7052270000000003
     },
     {
      "name": "void at::native::elementwise_kernel<(int)128, (int)2, void at::native::gpu_kernel_impl_noc",
      "ms": 2.570962
     },
     {
      "name": "void cudnn::engines_precompiled::nhwcToNchwKernel<float, float, float, (bool)1, (bool)0, (",
      "ms": 0.64157
     }
    ]
   },
   {
    "unit": "vae/up_blocks.1.resnets.0.norm1/WanRMS_norm",
    "ms": 8.121924,
    "share": 0.0007221143742146491,
    "count": 1416,
    "families": {
     "elementwise_reduce": 8.121924
    },
    "top_kernels": [
     {
      "name": "void at::native::elementwise_kernel<(int)128, (int)2, void at::native::gpu_kernel_impl_noc",
      "ms": 4.138418
     },
     {
      "name": "void at::native::reduce_kernel<(int)128, (int)4, at::native::ReduceOp<float, at::native::N",
      "ms": 1.6562320000000004
     },
     {
      "name": "void at::native::vectorized_elementwise_kernel<(int)4, at::native::CUDAFunctorOnSelf_add<f",
      "ms": 1.08983
     },
     {
      "name": "void at::native::vectorized_elementwise_kernel<(int)4, at::native::AUnaryFunctor<float, fl",
      "ms": 1.085315
     }
    ]
   },
   {
    "unit": "vae/mid_block.attentions.0.norm/WanRMS_norm",
    "ms": 7.779682000000001,
    "share": 0.00069168588612981,
    "count": 1440,
    "families": {
     "elementwise_reduce": 7.779682000000001
    },
    "top_kernels": [
     {
      "name": "void at::native::elementwise_kernel<(int)128, (int)2, void at::native::gpu_kernel_impl_noc",
      "ms": 3.80389
     },
     {
      "name": "void at::native::reduce_kernel<(int)128, (int)4, at::native::ReduceOp<float, at::native::N",
      "ms": 1.599877
     },
     {
      "name": "void at::native::vectorized_elementwise_kernel<(int)4, at::native::AUnaryFunctor<float, fl",
      "ms": 1.107782
     },
     {
      "name": "void at::native::vectorized_elementwise_kernel<(int)4, at::native::CUDAFunctorOnSelf_add<f",
      "ms": 1.10298
     }
    ]
   },
   {
    "unit": "vae/up_blocks.1.resnets.2.nonlinearity/SiLU",
    "ms": 5.204955,
    "share": 0.00046276877530994,
    "count": 472,
    "families": {
     "elementwise_reduce": 5.204955
    },
    "top_kernels": [
     {
      "name": "void at::native::vectorized_elementwise_kernel<(int)4, at::native::<unnamed>::silu_kernel(",
      "ms": 5.204955
     }
    ]
   },
   {
    "unit": "vae/up_blocks.1.resnets.1.nonlinearity/SiLU",
    "ms": 5.197117,
    "share": 0.00046207190441271246,
    "count": 472,
    "families": {
     "elementwise_reduce": 5.197117
    },
    "top_kernels": [
     {
      "name": "void at::native::vectorized_elementwise_kernel<(int)4, at::native::<unnamed>::silu_kernel(",
      "ms": 5.197117
     }
    ]
   },
   {
    "unit": "vae/up_blocks.0.upsamplers.0.resample.0/WanUpsample",
    "ms": 4.874293000000001,
    "share": 0.00043336985662927225,
    "count": 468,
    "families": {
     "gather_index_cat": 3.946928,
     "elementwise_reduce": 0.9273650000000001
    },
    "top_kernels": [
     {
      "name": "void at::native::<unnamed>::upsample_nearest2d_out_frame<float, &at::native::nearest_neigh",
      "ms": 3.946928
     },
     {
      "name": "void at::native::elementwise_kernel<(int)128, (int)2, void at::native::gpu_kernel_impl_noc",
      "ms": 0.9273650000000001
     }
    ]
   },
   {
    "unit": "vae/up_blocks.1.resnets.0.nonlinearity/SiLU",
    "ms": 3.685259,
    "share": 0.00032765370577265973,
    "count": 476,
    "families": {
     "elementwise_reduce": 3.685259
    },
    "top_kernels": [
     {
      "name": "void at::native::vectorized_elementwise_kernel<(int)4, at::native::<unnamed>::silu_kernel(",
      "ms": 3.685259
     }
    ]
   },
   {
    "unit": "vae/mid_block.resnets.1.norm1/WanRMS_norm",
    "ms": 3.3233129999999997,
    "share": 0.00029547334933377955,
    "count": 1440,
    "families": {
     "elementwise_reduce": 3.3233129999999997
    },
    "top_kernels": [
     {
      "name": "void at::native::reduce_kernel<(int)128, (int)4, at::native::ReduceOp<float, at::native::N",
      "ms": 1.2602969999999998
     },
     {
      "name": "void at::native::elementwise_kernel<(int)128, (int)2, void at::native::gpu_kernel_impl_noc",
      "ms": 1.245029
     },
     {
      "name": "void at::native::vectorized_elementwise_kernel<(int)4, at::native::CUDAFunctorOnSelf_add<f",
      "ms": 0.33801699999999996
     },
     {
      "name": "void at::native::vectorized_elementwise_kernel<(int)4, at::native::AUnaryFunctor<float, fl",
      "ms": 0.335426
     }
    ]
   },
   {
    "unit": "vae/mid_block.resnets.0.norm1/WanRMS_norm",
    "ms": 3.3205930000000006,
    "share": 0.0002952315161058568,
    "count": 1440,
    "families": {
     "elementwise_reduce": 3.3205930000000006
    },
    "top_kernels": [
     {
      "name": "void at::native::reduce_kernel<(int)128, (int)4, at::native::ReduceOp<float, at::native::N",
      "ms": 1.257929
     },
     {
      "name": "void at::native::elementwise_kernel<(int)128, (int)2, void at::native::gpu_kernel_impl_noc",
      "ms": 1.243042
     },
     {
      "name": "void at::native::vectorized_elementwise_kernel<(int)4, at::native::CUDAFunctorOnSelf_add<f",
      "ms": 0.338275
     },
     {
      "name": "void at::native::vectorized_elementwise_kernel<(int)4, at::native::AUnaryFunctor<float, fl",
      "ms": 0.33526700000000004
     }
    ]
   },
   {
    "unit": "vae/mid_block.resnets.0.norm2/WanRMS_norm",
    "ms": 3.3191569999999997,
    "share": 0.00029510384238699746,
    "count": 1440,
    "families": {
     "elementwise_reduce": 3.3191569999999997
    },
    "top_kernels": [
     {
      "name": "void at::native::reduce_kernel<(int)128, (int)4, at::native::ReduceOp<float, at::native::N",
      "ms": 1.2589550000000003
     },
     {
      "name": "void at::native::elementwise_kernel<(int)128, (int)2, void at::native::gpu_kernel_impl_noc",
      "ms": 1.239431
     },
     {
      "name": "void at::native::vectorized_elementwise_kernel<(int)4, at::native::CUDAFunctorOnSelf_add<f",
      "ms": 0.33686499999999997
     },
     {
      "name": "void at::native::vectorized_elementwise_kernel<(int)4, at::native::AUnaryFunctor<float, fl",
      "ms": 0.335617
     }
    ]
   },
   {
    "unit": "vae_convert/ep*_chunk*",
    "ms": 3.3170370000000005,
    "share": 0.00029491535472405767,
    "count": 100,
    "families": {
     "elementwise_reduce": 3.3170370000000005
    },
    "top_kernels": [
     {
      "name": "void at::native::unrolled_elementwise_kernel<at::native::direct_copy_kernel_cuda(at::Tenso",
      "ms": 0.881061
     },
     {
      "name": "void at::native::vectorized_elementwise_kernel<(int)4, at::native::round_kernel_cuda(at::T",
      "ms": 0.611521
     },
     {
      "name": "void at::native::vectorized_elementwise_kernel<(int)4, at::native::<unnamed>::launch_clamp",
      "ms": 0.609696
     },
     {
      "name": "void at::native::vectorized_elementwise_kernel<(int)4, at::native::AUnaryFunctor<float, fl",
      "ms": 0.60874
     }
    ]
   },
   {
    "unit": "vae/mid_block.resnets.1.norm2/WanRMS_norm",
    "ms": 3.306925,
    "share": 0.0002940163041355446,
    "count": 1440,
    "families": {
     "elementwise_reduce": 3.306925
    },
    "top_kernels": [
     {
      "name": "void at::native::reduce_kernel<(int)128, (int)4, at::native::ReduceOp<float, at::native::N",
      "ms": 1.252675
     },
     {
      "name": "void at::native::elementwise_kernel<(int)128, (int)2, void at::native::gpu_kernel_impl_noc",
      "ms": 1.2366460000000001
     },
     {
      "name": "void at::native::vectorized_elementwise_kernel<(int)4, at::native::CUDAFunctorOnSelf_add<f",
      "ms": 0.33692900000000003
     },
     {
      "name": "void at::native::vectorized_elementwise_kernel<(int)4, at::native::AUnaryFunctor<float, fl",
      "ms": 0.332994
     }
    ]
   },
   {
    "unit": "vae/up_blocks.0.resnets.2.norm2/WanRMS_norm",
    "ms": 3.3059050000000005,
    "share": 0.00029392561667507353,
    "count": 1440,
    "families": {
     "elementwise_reduce": 3.3059050000000005
    },
    "top_kernels": [
     {
      "name": "void at::native::reduce_kernel<(int)128, (int)4, at::native::ReduceOp<float, at::native::N",
      "ms": 1.2522929999999999
     },
     {
      "name": "void at::native::elementwise_kernel<(int)128, (int)2, void at::native::gpu_kernel_impl_noc",
      "ms": 1.234473
     },
     {
      "name": "void at::native::vectorized_elementwise_kernel<(int)4, at::native::CUDAFunctorOnSelf_add<f",
      "ms": 0.337058
     },
     {
      "name": "void at::native::vectorized_elementwise_kernel<(int)4, at::native::AUnaryFunctor<float, fl",
      "ms": 0.33456
     }
    ]
   },
   {
    "unit": "vae/up_blocks.0.resnets.1.norm2/WanRMS_norm",
    "ms": 3.3045889999999996,
    "share": 0.0002938086120692108,
    "count": 1440,
    "families": {
     "elementwise_reduce": 3.3045889999999996
    },
    "top_kernels": [
     {
      "name": "void at::native::reduce_kernel<(int)128, (int)4, at::native::ReduceOp<float, at::native::N",
      "ms": 1.255141
     },
     {
      "name": "void at::native::elementwise_kernel<(int)128, (int)2, void at::native::gpu_kernel_impl_noc",
      "ms": 1.2308510000000001
     },
     {
      "name": "void at::native::vectorized_elementwise_kernel<(int)4, at::native::CUDAFunctorOnSelf_add<f",
      "ms": 0.337508
     },
     {
      "name": "void at::native::vectorized_elementwise_kernel<(int)4, at::native::AUnaryFunctor<float, fl",
      "ms": 0.33430499999999996
     }
    ]
   },
   {
    "unit": "vae/up_blocks.0.resnets.0.norm2/WanRMS_norm",
    "ms": 3.304043,
    "share": 0.00029376006760507633,
    "count": 1440,
    "families": {
     "elementwise_reduce": 3.304043
    },
    "top_kernels": [
     {
      "name": "void at::native::reduce_kernel<(int)128, (int)4, at::native::ReduceOp<float, at::native::N",
      "ms": 1.252675
     },
     {
      "name": "void at::native::elementwise_kernel<(int)128, (int)2, void at::native::gpu_kernel_impl_noc",
      "ms": 1.231779
     },
     {
      "name": "void at::native::vectorized_elementwise_kernel<(int)4, at::native::CUDAFunctorOnSelf_add<f",
      "ms": 0.33757000000000004
     },
     {
      "name": "void at::native::vectorized_elementwise_kernel<(int)4, at::native::AUnaryFunctor<float, fl",
      "ms": 0.33593799999999996
     }
    ]
   },
   {
    "unit": "vae/up_blocks.0.resnets.2.norm1/WanRMS_norm",
    "ms": 3.2988009999999997,
    "share": 0.00029329400518567506,
    "count": 1440,
    "families": {
     "elementwise_reduce": 3.2988009999999997
    },
    "top_kernels": [
     {
      "name": "void at::native::reduce_kernel<(int)128, (int)4, at::native::ReduceOp<float, at::native::N",
      "ms": 1.2503760000000002
     },
     {
      "name": "void at::native::elementwise_kernel<(int)128, (int)2, void at::native::gpu_kernel_impl_noc",
      "ms": 1.2338909999999998
     },
     {
      "name": "void at::native::vectorized_elementwise_kernel<(int)4, at::native::CUDAFunctorOnSelf_add<f",
      "ms": 0.33692900000000003
     },
     {
      "name": "void at::native::vectorized_elementwise_kernel<(int)4, at::native::AUnaryFunctor<float, fl",
      "ms": 0.331298
     }
    ]
   },
   {
    "unit": "vae/up_blocks.0.resnets.0.norm1/WanRMS_norm",
    "ms": 3.2952799999999995,
    "share": 0.0002929809556284999,
    "count": 1440,
    "families": {
     "elementwise_reduce": 3.2952799999999995
    },
    "top_kernels": [
     {
      "name": "void at::native::reduce_kernel<(int)128, (int)4, at::native::ReduceOp<float, at::native::N",
      "ms": 1.246505
     },
     {
      "name": "void at::native::elementwise_kernel<(int)128, (int)2, void at::native::gpu_kernel_impl_noc",
      "ms": 1.2341159999999998
     },
     {
      "name": "void at::native::vectorized_elementwise_kernel<(int)4, at::native::CUDAFunctorOnSelf_add<f",
      "ms": 0.336097
     },
     {
      "name": "void at::native::vectorized_elementwise_kernel<(int)4, at::native::AUnaryFunctor<float, fl",
      "ms": 0.333473
     }
    ]
   },
   {
    "unit": "vae/up_blocks.0.resnets.1.norm1/WanRMS_norm",
    "ms": 3.293650999999999,
    "share": 0.0002928361224195711,
    "count": 1440,
    "families": {
     "elementwise_reduce": 3.293650999999999
    },
    "top_kernels": [
     {
      "name": "void at::native::reduce_kernel<(int)128, (int)4, at::native::ReduceOp<float, at::native::N",
      "ms": 1.24775
     },
     {
      "name": "void at::native::elementwise_kernel<(int)128, (int)2, void at::native::gpu_kernel_impl_noc",
      "ms": 1.230631
     },
     {
      "name": "void at::native::vectorized_elementwise_kernel<(int)4, at::native::CUDAFunctorOnSelf_add<f",
      "ms": 0.337283
     },
     {
      "name": "void at::native::vectorized_elementwise_kernel<(int)4, at::native::AUnaryFunctor<float, fl",
      "ms": 0.3328
     }
    ]
   },
   {
    "unit": "no context",
    "ms": 1.8557220000000003,
    "share": 0.00016499089756889592,
    "count": 100,
    "families": {
     "elementwise_reduce": 1.8557220000000003
    },
    "top_kernels": [
     {
      "name": "void at::native::vectorized_elementwise_kernel<(int)4, at::native::AbsFunctor<float>, std:",
      "ms": 0.582756
     },
     {
      "name": "void at::native::vectorized_elementwise_kernel<(int)8, at::native::AUnaryFunctor<float, fl",
      "ms": 0.406563
     },
     {
      "name": "void at::native::vectorized_elementwise_kernel<(int)8, at::native::BinaryFunctor<float, fl",
      "ms": 0.382656
     },
     {
      "name": "void at::native::vectorized_elementwise_kernel<(int)8, at::native::BinaryFunctor<bool, boo",
      "ms": 0.27273600000000003
     }
    ]
   },
   {
    "unit": "vae_decode/ep*_chunk*",
    "ms": 1.570883,
    "share": 0.0001396660685952529,
    "count": 78,
    "families": {
     "elementwise_reduce": 1.570883
    },
    "top_kernels": [
     {
      "name": "void at::native::elementwise_kernel<(int)128, (int)2, void at::native::gpu_kernel_impl_noc",
      "ms": 0.975715
     },
     {
      "name": "void at::native::vectorized_elementwise_kernel<(int)4, at::native::<unnamed>::launch_clamp",
      "ms": 0.595168
     }
    ]
   },
   {
    "unit": "vae/mid_block.resnets.0.nonlinearity/SiLU",
    "ms": 0.706374,
    "share": 6.280320019880739e-05,
    "count": 480,
    "families": {
     "elementwise_reduce": 0.706374
    },
    "top_kernels": [
     {
      "name": "void at::native::vectorized_elementwise_kernel<(int)4, at::native::<unnamed>::silu_kernel(",
      "ms": 0.706374
     }
    ]
   },
   {
    "unit": "vae/up_blocks.0.resnets.2.nonlinearity/SiLU",
    "ms": 0.702051,
    "share": 6.241884540310506e-05,
    "count": 480,
    "families": {
     "elementwise_reduce": 0.702051
    },
    "top_kernels": [
     {
      "name": "void at::native::vectorized_elementwise_kernel<(int)4, at::native::<unnamed>::silu_kernel(",
      "ms": 0.702051
     }
    ]
   },
   {
    "unit": "vae/mid_block.resnets.1.nonlinearity/SiLU",
    "ms": 0.699363,
    "share": 6.217985727198132e-05,
    "count": 480,
    "families": {
     "elementwise_reduce": 0.699363
    },
    "top_kernels": [
     {
      "name": "void at::native::vectorized_elementwise_kernel<(int)4, at::native::<unnamed>::silu_kernel(",
      "ms": 0.699363
     }
    ]
   },
   {
    "unit": "vae/up_blocks.0.resnets.1.nonlinearity/SiLU",
    "ms": 0.6984970000000001,
    "share": 6.21028618398559e-05,
    "count": 480,
    "families": {
     "elementwise_reduce": 0.6984970000000001
    },
    "top_kernels": [
     {
      "name": "void at::native::vectorized_elementwise_kernel<(int)4, at::native::<unnamed>::silu_kernel(",
      "ms": 0.6984970000000001
     }
    ]
   },
   {
    "unit": "vae/up_blocks.0.resnets.0.nonlinearity/SiLU",
    "ms": 0.698276,
    "share": 6.208321289008717e-05,
    "count": 480,
    "families": {
     "elementwise_reduce": 0.698276
    },
    "top_kernels": [
     {
      "name": "void at::native::vectorized_elementwise_kernel<(int)4, at::native::<unnamed>::silu_kernel(",
      "ms": 0.698276
     }
    ]
   },
   {
    "unit": "vae_pqc/root/WanCausalConv3d",
    "ms": 0.344259,
    "share": 3.060781809245702e-05,
    "count": 160,
    "families": {
     "gemm": 0.212002,
     "elementwise_reduce": 0.13225699999999999
    },
    "top_kernels": [
     {
      "name": "void cutlass::Kernel2<cutlass_80_tensorop_s1688gemm_128x64_16x6_nn_align4>(T1::Params)",
      "ms": 0.212002
     },
     {
      "name": "void at::native::elementwise_kernel<(int)128, (int)2, void at::native::gpu_kernel_impl_noc",
      "ms": 0.13225699999999999
     }
    ]
   },
   {
    "unit": "vae_h2d/ep*_chunk*",
    "ms": 0.29859199999999997,
    "share": 2.6547598232327772e-05,
    "count": 160,
    "families": {
     "elementwise_reduce": 0.29859199999999997
    },
    "top_kernels": [
     {
      "name": "void at::native::elementwise_kernel<(int)128, (int)2, void at::native::gpu_kernel_impl_noc",
      "ms": 0.29859199999999997
     }
    ]
   },
   {
    "unit": "vae_barrier/ep*_chunk*",
    "ms": 0.031392,
    "share": 2.791039959909286e-06,
    "count": 80,
    "families": {
     "elementwise_reduce": 0.031392
    },
    "top_kernels": [
     {
      "name": "void at::native::vectorized_elementwise_kernel<(int)4, at::native::FillFunctor<float>, std",
      "ms": 0.031392
     }
    ]
   },
   {
    "unit": "vae_barrier2/ep*_chunk*",
    "ms": 0.030048,
    "share": 2.6715458943474203e-06,
    "count": 80,
    "families": {
     "elementwise_reduce": 0.030048
    },
    "top_kernels": [
     {
      "name": "void at::native::vectorized_elementwise_kernel<(int)4, at::native::FillFunctor<float>, std",
      "ms": 0.030048
     }
    ]
   }
  ],
  "by_family": {
   "conv": 5392.86247599999,
   "elementwise_reduce": 4144.134979999985,
   "attention_core": 1076.8601840000001,
   "nccl": 401.21320600000007,
   "gather_index_cat": 214.73648999999997,
   "gemm": 17.613142
  }
 },
 "self_attention_by_frame": [
  {
   "frame": 15,
   "commit": false,
   "family": "attention_core",
   "ms": 3490.106461,
   "count": 1280
  },
  {
   "frame": 15,
   "commit": false,
   "family": "attention_prep",
   "ms": 1.537285,
   "count": 1280
  },
  {
   "frame": 15,
   "commit": false,
   "family": "elementwise_reduce",
   "ms": 1.8872079999999998,
   "count": 2560
  },
  {
   "frame": 15,
   "commit": false,
   "family": "gather_index_cat",
   "ms": 201.224194,
   "count": 6400
  },
  {
   "frame": 15,
   "commit": true,
   "family": "attention_core",
   "ms": 902.513822,
   "count": 320
  },
  {
   "frame": 15,
   "commit": true,
   "family": "attention_prep",
   "ms": 0.394979,
   "count": 320
  },
  {
   "frame": 15,
   "commit": true,
   "family": "elementwise_reduce",
   "ms": 0.476892,
   "count": 640
  },
  {
   "frame": 15,
   "commit": true,
   "family": "gather_index_cat",
   "ms": 51.060176000000006,
   "count": 1600
  },
  {
   "frame": 18,
   "commit": false,
   "family": "attention_core",
   "ms": 3479.80508,
   "count": 1280
  },
  {
   "frame": 18,
   "commit": false,
   "family": "attention_prep",
   "ms": 1.538087,
   "count": 1280
  },
  {
   "frame": 18,
   "commit": false,
   "family": "elementwise_reduce",
   "ms": 1.881091,
   "count": 2560
  },
  {
   "frame": 18,
   "commit": false,
   "family": "gather_index_cat",
   "ms": 201.20278499999998,
   "count": 6400
  },
  {
   "frame": 18,
   "commit": true,
   "family": "attention_core",
   "ms": 905.882167,
   "count": 320
  },
  {
   "frame": 18,
   "commit": true,
   "family": "attention_prep",
   "ms": 0.393696,
   "count": 320
  },
  {
   "frame": 18,
   "commit": true,
   "family": "elementwise_reduce",
   "ms": 0.48204499999999995,
   "count": 640
  },
  {
   "frame": 18,
   "commit": true,
   "family": "gather_index_cat",
   "ms": 51.06804999999999,
   "count": 1600
  },
  {
   "frame": 21,
   "commit": false,
   "family": "attention_core",
   "ms": 3480.667976,
   "count": 1280
  },
  {
   "frame": 21,
   "commit": false,
   "family": "attention_prep",
   "ms": 1.534023,
   "count": 1280
  },
  {
   "frame": 21,
   "commit": false,
   "family": "elementwise_reduce",
   "ms": 1.883232,
   "count": 2560
  },
  {
   "frame": 21,
   "commit": false,
   "family": "gather_index_cat",
   "ms": 201.061959,
   "count": 6400
  },
  {
   "frame": 21,
   "commit": true,
   "family": "attention_core",
   "ms": 906.626515,
   "count": 320
  },
  {
   "frame": 21,
   "commit": true,
   "family": "attention_prep",
   "ms": 0.393697,
   "count": 320
  },
  {
   "frame": 21,
   "commit": true,
   "family": "elementwise_reduce",
   "ms": 0.47897599999999996,
   "count": 640
  },
  {
   "frame": 21,
   "commit": true,
   "family": "gather_index_cat",
   "ms": 51.086868,
   "count": 1600
  },
  {
   "frame": 24,
   "commit": false,
   "family": "attention_core",
   "ms": 3482.110793,
   "count": 1280
  },
  {
   "frame": 24,
   "commit": false,
   "family": "attention_prep",
   "ms": 1.54372,
   "count": 1280
  },
  {
   "frame": 24,
   "commit": false,
   "family": "elementwise_reduce",
   "ms": 1.879991,
   "count": 2560
  },
  {
   "frame": 24,
   "commit": false,
   "family": "gather_index_cat",
   "ms": 201.23355,
   "count": 6400
  },
  {
   "frame": 24,
   "commit": true,
   "family": "attention_core",
   "ms": 908.099737,
   "count": 320
  },
  {
   "frame": 24,
   "commit": true,
   "family": "attention_prep",
   "ms": 0.393507,
   "count": 320
  },
  {
   "frame": 24,
   "commit": true,
   "family": "elementwise_reduce",
   "ms": 0.477401,
   "count": 640
  },
  {
   "frame": 24,
   "commit": true,
   "family": "gather_index_cat",
   "ms": 51.153942,
   "count": 1600
  }
 ]
}