{
 "generated_at": "2026-09-10T15:57:22+0800",
 "label": "ledger_capture_480x832_v1",
 "steady_gpu_ms_total": 12224.388317,
 "windows_ns": {
  "measured": [
   [
    253743914870,
    263832309519
   ]
  ],
  "steady": [
   [
    259419853196,
    260366664851
   ],
   [
    260366664851,
    261218193875
   ],
   [
    261218193875,
    262070595850
   ],
   [
    262070595850,
    262924720589
   ]
  ]
 },
 "dit_steady": {
  "gpu_ms": 9943.362067,
  "groups": 726,
  "by_unit": [
   {
    "unit": "vllm_omni.ar_diffusion_paged_write_attn: ar_diffusion_paged_write_attn",
    "ms": 3603.860268000002,
    "share": 0.362438805276988,
    "count": 57600,
    "families": {
     "attention_core": 3144.670284,
     "gather_index_cat": 443.6999179999999,
     "elementwise_reduce": 8.314440000000001,
     "attention_prep": 7.175626
    },
    "top_kernels": [
     {
      "name": "void cutlass::device_kernel<flash::enable_sm90_or_later<flash::FlashAttnFwdSm90<flash::Col",
      "ms": 3144.670284
     },
     {
      "name": "void at::native::vectorized_gather_kernel<(int)16, long>(char *, char *, T2 *, int, long, ",
      "ms": 248.05808600000003
     },
     {
      "name": "void at::native::index_elementwise_kernel<(int)128, (int)4, void at::native::gpu_index_ker",
      "ms": 191.032256
     },
     {
      "name": "void flash::prepare_varlen_num_blocks_kernel<(int)1, (bool)0>(int, int, int, const int *, ",
      "ms": 7.175626
     }
    ]
   },
   {
    "unit": "addmm: float_1, modulation, chunk, float_2, layer_norm, scale_msa, add_1, mul",
    "ms": 851.9677899999999,
    "share": 0.08568206450286146,
    "count": 6400,
    "families": {
     "gemm": 851.9677899999999
    },
    "top_kernels": [
     {
      "name": "nvjet_sm90_tst_256x152_64x4_1x2_h_bz_coopA_bias_TNT",
      "ms": 851.9677899999999
     }
    ]
   },
   {
    "unit": "addmm: output_parallel_6, input_1, output_parallel_7",
    "ms": 721.026024,
    "share": 0.07251330275832346,
    "count": 6400,
    "families": {
     "gemm": 721.026024
    },
    "top_kernels": [
     {
      "name": "nvjet_sm90_tst_128x200_64x5_2x1_v_bz_coopA_bias_TNT",
      "ms": 721.026024
     }
    ]
   },
   {
    "unit": "addmm: output_parallel_6",
    "ms": 698.005771,
    "share": 0.07019816499657992,
    "count": 6400,
    "families": {
     "gemm": 698.005771
    },
    "top_kernels": [
     {
      "name": "nvjet_sm90_tst_128x200_64x5_2x1_v_bz_coopA_bias_TNT",
      "ms": 698.005771
     }
    ]
   },
   {
    "unit": "_c10d_functional.all_to_all_single: output_parallel, split, value_7, stack, reshape, transpose, input_t, t",
    "ms": 414.76803400000006,
    "share": 0.04171305753579375,
    "count": 6400,
    "families": {
     "nccl": 414.76803400000006
    },
    "top_kernels": [
     {
      "name": "ncclDevKernel_SendRecv(ncclDevKernelArgsStorage<(unsigned long)4096>)",
      "ms": 414.76803400000006
     }
    ]
   },
   {
    "unit": "addmm: output_parallel_2, input_parallel_1, output_parallel_3",
    "ms": 317.40445900000003,
    "share": 0.03192124121210481,
    "count": 6400,
    "families": {
     "gemm": 317.40445900000003
    },
    "top_kernels": [
     {
      "name": "nvjet_sm90_tst_128x200_64x5_2x1_v_bz_coopA_bias_TNT",
      "ms": 317.40445900000003
     }
    ]
   },
   {
    "unit": "_c10d_functional.all_to_all_single: output_5, reshape_3, transpose_3, transpose_4, contiguous_7, tensor_1",
    "ms": 295.17225799999994,
    "share": 0.02968535752908131,
    "count": 6400,
    "families": {
     "nccl": 295.17225799999994
    },
    "top_kernels": [
     {
      "name": "ncclDevKernel_SendRecv(ncclDevKernelArgsStorage<(unsigned long)4096>)",
      "ms": 295.17225799999994
     }
    ]
   },
   {
    "unit": "addmm: contiguous_12, input_parallel_2, output_parallel_5",
    "ms": 293.444968,
    "share": 0.029511644655270505,
    "count": 6400,
    "families": {
     "gemm": 293.444968
    },
    "top_kernels": [
     {
      "name": "nvjet_sm90_tst_128x200_64x5_2x1_v_bz_coopA_bias_TNT",
      "ms": 293.444968
     }
    ]
   },
   {
    "unit": "addmm: float_8, layer_norm_1, float_9, float_10, to_6, output_parallel_4",
    "ms": 279.64515800000004,
    "share": 0.028123803208181017,
    "count": 6400,
    "families": {
     "gemm": 279.64515800000004
    },
    "top_kernels": [
     {
      "name": "nvjet_sm90_tst_128x200_64x5_2x1_v_bz_coopA_bias_TNT",
      "ms": 279.64515800000004
     }
    ]
   },
   {
    "unit": "addmm: contiguous_8, input_parallel, output_parallel_1",
    "ms": 278.834526,
    "share": 0.028042278267769727,
    "count": 6400,
    "families": {
     "gemm": 278.834526
    },
    "top_kernels": [
     {
      "name": "nvjet_sm90_tst_128x200_64x5_2x1_v_bz_coopA_bias_TNT",
      "ms": 278.834526
     }
    ]
   },
   {
    "unit": "addmm: output_parallel_2",
    "ms": 278.062379,
    "share": 0.027964623748624483,
    "count": 6400,
    "families": {
     "gemm": 278.062379
    },
    "top_kernels": [
     {
      "name": "nvjet_sm90_tst_128x200_64x5_2x1_v_bz_coopA_bias_TNT",
      "ms": 278.062379
     }
    ]
   },
   {
    "unit": "addmm: output_parallel_3, camera_features, camera_scale",
    "ms": 277.86213399999997,
    "share": 0.027944485187979626,
    "count": 6400,
    "families": {
     "gemm": 277.86213399999997
    },
    "top_kernels": [
     {
      "name": "nvjet_sm90_tst_128x200_64x5_2x1_v_bz_coopA_bias_TNT",
      "ms": 277.86213399999997
     }
    ]
   },
   {
    "unit": "addmm: camera_shift",
    "ms": 274.97936500000003,
    "share": 0.027654566247024306,
    "count": 6400,
    "families": {
     "gemm": 274.97936500000003
    },
    "top_kernels": [
     {
      "name": "nvjet_sm90_tst_128x200_64x5_2x1_v_bz_coopA_bias_TNT",
      "ms": 274.97936500000003
     }
    ]
   },
   {
    "unit": "_c10d_functional.all_to_all_single: output_parallel_4, value_float_2, truediv_4, rms_2, truediv_5, float_1",
    "ms": 198.12153699999996,
    "share": 0.019925004808738194,
    "count": 6400,
    "families": {
     "nccl": 198.12153699999996
    },
    "top_kernels": [
     {
      "name": "ncclDevKernel_SendRecv(ncclDevKernelArgsStorage<(unsigned long)4096>)",
      "ms": 198.12153699999996
     }
    ]
   },
   {
    "unit": "_c10d_functional.all_to_all_single: reshape_10, transpose_8, transpose_9, contiguous_11, tensor_3",
    "ms": 165.533301,
    "share": 0.01664761877165988,
    "count": 6400,
    "families": {
     "nccl": 165.533301
    },
    "top_kernels": [
     {
      "name": "ncclDevKernel_SendRecv(ncclDevKernelArgsStorage<(unsigned long)4096>)",
      "ms": 165.533301
     }
    ]
   },
   {
    "unit": "fa3_fwd._flash_attn_forward: _flash_attn_forward_default, key_4, value_9",
    "ms": 92.77299099999999,
    "share": 0.009330143101989086,
    "count": 6400,
    "families": {
     "attention_core": 92.77299099999999
    },
    "top_kernels": [
     {
      "name": "void cutlass::device_kernel<flash::enable_sm90_or_later<flash::FlashAttnFwdSm90<flash::Col",
      "ms": 92.77299099999999
     }
    ]
   },
   {
    "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": 77.775204,
    "share": 0.007821821580662351,
    "count": 6400,
    "families": {
     "triton_fused": 77.775204
    },
    "top_kernels": [
     {
      "name": "triton_red_fused__to_copy_add_mul_native_layer_norm_split_squeeze_view_9",
      "ms": 77.775204
     }
    ]
   },
   {
    "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": 73.54639600000002,
    "share": 0.007396532028546518,
    "count": 6400,
    "families": {
     "triton_fused": 73.54639600000002
    },
    "top_kernels": [
     {
      "name": "triton_red_fused__to_copy_add_mul_native_layer_norm_split_squeeze_view_12",
      "ms": 73.54639600000002
     }
    ]
   },
   {
    "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": 71.26247900000001,
    "share": 0.007166839396958672,
    "count": 6400,
    "families": {
     "triton_fused": 71.26247900000001
    },
    "top_kernels": [
     {
      "name": "triton_red_fused__to_copy_add_mul_native_layer_norm_split_squeeze_1",
      "ms": 71.26247900000001
     }
    ]
   },
   {
    "unit": "triton_poi_fused_gelu_view_13: output_parallel_6, input_1",
    "ms": 71.048524,
    "share": 0.007145322027022995,
    "count": 6400,
    "families": {
     "triton_fused": 71.048524
    },
    "top_kernels": [
     {
      "name": "triton_poi_fused_gelu_view_13",
      "ms": 71.048524
     }
    ]
   },
   {
    "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": 58.59643500000001,
    "share": 0.005893020349170395,
    "count": 6400,
    "families": {
     "triton_fused": 58.59643500000001
    },
    "top_kernels": [
     {
      "name": "triton_poi_fused_all_to_all_single_clone_split_with_sizes_stack_transpose_view_4",
      "ms": 58.59643500000001
     }
    ]
   },
   {
    "unit": "triton_poi_fused__to_copy_add_mul_split_squeeze_view_14: float_1, modulation, chunk, output_parallel_5, hidden_states_2, output",
    "ms": 49.27491000000001,
    "share": 0.004955558257657481,
    "count": 6400,
    "families": {
     "triton_fused": 49.27491000000001
    },
    "top_kernels": [
     {
      "name": "triton_poi_fused__to_copy_add_mul_split_squeeze_view_14",
      "ms": 49.27491000000001
     }
    ]
   },
   {
    "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": 34.926145,
    "share": 0.0035125086228040294,
    "count": 6400,
    "families": {
     "triton_fused": 34.926145
    },
    "top_kernels": [
     {
      "name": "triton_poi_fused__to_copy_all_to_all_single_clone_div_mul_sqrt_transpose_view_11",
      "ms": 34.926145
     }
    ]
   },
   {
    "unit": "triton_poi_fused_clone_transpose_view_8: contiguous_8",
    "ms": 34.891143,
    "share": 0.0035089884854738037,
    "count": 6400,
    "families": {
     "triton_fused": 34.891143
    },
    "top_kernels": [
     {
      "name": "triton_poi_fused_clone_transpose_view_8",
      "ms": 34.891143
     }
    ]
   },
   {
    "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": 34.19756999999999,
    "share": 0.0034392361225077765,
    "count": 6400,
    "families": {
     "triton_fused": 34.19756999999999
    },
    "top_kernels": [
     {
      "name": "triton_red_fused__to_copy_div_mul_pow_split_with_sizes_sqrt_sum_view_3",
      "ms": 34.19756999999999
     }
    ]
   },
   {
    "unit": "triton_poi_fused_silu_view_6: output_parallel_2, input_parallel_1",
    "ms": 33.03871,
    "share": 0.0033226900295272132,
    "count": 6400,
    "families": {
     "triton_fused": 33.03871
    },
    "top_kernels": [
     {
      "name": "triton_poi_fused_silu_view_6",
      "ms": 33.03871
     }
    ]
   },
   {
    "unit": "rotary_kernel_0: output_parallel, split, value_float, truediv, rms, truediv_1, float_5,",
    "ms": 31.093873,
    "share": 0.0031270985397579207,
    "count": 6400,
    "families": {
     "triton_fused": 31.093873
    },
    "top_kernels": [
     {
      "name": "rotary_kernel",
      "ms": 31.093873
     }
    ]
   },
   {
    "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": 31.027860999999994,
    "share": 0.003120459738962455,
    "count": 6400,
    "families": {
     "triton_fused": 31.027860999999994
    },
    "top_kernels": [
     {
      "name": "triton_red_fused__to_copy_div_mul_pow_split_with_sizes_sqrt_sum_view_2",
      "ms": 31.027860999999994
     }
    ]
   },
   {
    "unit": "rotary_kernel_0: output_parallel, split, value_float_1, truediv_2, rms_1, truediv_3, fl",
    "ms": 30.98676,
    "share": 0.0031163262276085435,
    "count": 6400,
    "families": {
     "triton_fused": 30.98676
    },
    "top_kernels": [
     {
      "name": "rotary_kernel",
      "ms": 30.98676
     }
    ]
   },
   {
    "unit": "non-graph: tfm_sub/head/_LingBotHead",
    "ms": 29.921179999999993,
    "share": 0.0030091612674250606,
    "count": 1440,
    "families": {
     "elementwise_reduce": 28.592601999999996,
     "gemm": 1.3285780000000003
    },
    "top_kernels": [
     {
      "name": "void at::native::elementwise_kernel<(int)128, (int)2, void at::native::gpu_kernel_impl_noc",
      "ms": 13.215067000000005
     },
     {
      "name": "void at::native::unrolled_elementwise_kernel<at::native::direct_copy_kernel_cuda(at::Tenso",
      "ms": 6.214824999999999
     },
     {
      "name": "void at::native::<unnamed>::vectorized_layer_norm_kernel<float, float, (bool)0>(int, T2, c",
      "ms": 6.005807
     },
     {
      "name": "void at::native::vectorized_elementwise_kernel<(int)8, at::native::bfloat16_copy_kernel_cu",
      "ms": 2.741095
     }
    ]
   },
   {
    "unit": "triton_poi_fused_add_view_7: output_parallel_3, camera_features",
    "ms": 29.508433,
    "share": 0.002967651464481264,
    "count": 6400,
    "families": {
     "triton_fused": 29.508433
    },
    "top_kernels": [
     {
      "name": "triton_poi_fused_add_view_7",
      "ms": 29.508433
     }
    ]
   },
   {
    "unit": "triton_poi_fused_clone_transpose_view_8: contiguous_12",
    "ms": 29.271797999999997,
    "share": 0.002943853175893811,
    "count": 6400,
    "families": {
     "triton_fused": 29.271797999999997
    },
    "top_kernels": [
     {
      "name": "triton_poi_fused_clone_transpose_view_8",
      "ms": 29.271797999999997
     }
    ]
   },
   {
    "unit": "non-graph: tfm_sub/sp_prepare/_LingBotSPPrepare",
    "ms": 25.412559,
    "share": 0.002555731032297328,
    "count": 160,
    "families": {
     "elementwise_reduce": 25.412559
    },
    "top_kernels": [
     {
      "name": "void at::native::elementwise_kernel<(int)128, (int)4, void at::native::gpu_kernel_impl_noc",
      "ms": 25.412559
     }
    ]
   },
   {
    "unit": "non-graph: tfm_sub/c2ws_hidden_states_layer1/ColumnParallelLinear",
    "ms": 24.255664000000003,
    "share": 0.0024393825585914877,
    "count": 160,
    "families": {
     "gemm": 24.255664000000003
    },
    "top_kernels": [
     {
      "name": "nvjet_sm90_tst_256x144_64x4_2x1_v_bz_coopA_bias_TNT",
      "ms": 24.255664000000003
     }
    ]
   },
   {
    "unit": "non-graph: tfm_sub/c2ws_hidden_states_layer2/RowParallelLinear",
    "ms": 24.234256,
    "share": 0.0024372295644778516,
    "count": 160,
    "families": {
     "gemm": 24.234256
    },
    "top_kernels": [
     {
      "name": "nvjet_sm90_tst_256x144_64x4_2x1_v_bz_coopA_bias_TNT",
      "ms": 24.234256
     }
    ]
   },
   {
    "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": 22.349157,
    "share": 0.002247645901799384,
    "count": 6400,
    "families": {
     "triton_fused": 22.349157
    },
    "top_kernels": [
     {
      "name": "triton_poi_fused_all_to_all_single_clone_transpose_unsqueeze_view_5",
      "ms": 22.349157
     }
    ]
   },
   {
    "unit": "triton_poi_fused_all_to_all_single_clone_transpose_unsqueeze_view_5: reshape_10, transpose_8, transpose_9, contiguous_11, tensor_3",
    "ms": 20.475788999999995,
    "share": 0.0020592420211625385,
    "count": 6400,
    "families": {
     "triton_fused": 20.475788999999995
    },
    "top_kernels": [
     {
      "name": "triton_poi_fused_all_to_all_single_clone_transpose_unsqueeze_view_5",
      "ms": 20.475788999999995
     }
    ]
   },
   {
    "unit": "triton_red_fused__to_copy_pow_sum_view_10: output_parallel_4, value_float_2, pow_3, sum_of_squares_2",
    "ms": 16.388157,
    "share": 0.0016481504836667836,
    "count": 6400,
    "families": {
     "triton_fused": 16.388157
    },
    "top_kernels": [
     {
      "name": "triton_red_fused__to_copy_pow_sum_view_10",
      "ms": 16.388157
     }
    ]
   },
   {
    "unit": "non-graph: ncclAllGather",
    "ms": 11.746584,
    "share": 0.0011813493183542544,
    "count": 160,
    "families": {
     "nccl": 11.746584
    },
    "top_kernels": [
     {
      "name": "ncclDevKernel_AllGather_RING_LL(ncclDevKernelArgsStorage<(unsigned long)4096>)",
      "ms": 11.746584
     }
    ]
   },
   {
    "unit": "non-graph: tfm_sub/patch_embedding_wancamctrl/_LingBotCameraPatchEmbedding",
    "ms": 10.573814000000002,
    "share": 0.0010634043021617752,
    "count": 320,
    "families": {
     "gemm": 8.050131,
     "elementwise_reduce": 2.523683
    },
    "top_kernels": [
     {
      "name": "nvjet_sm90_tst_256x144_64x4_2x1_v_bz_coopA_bias_TNT",
      "ms": 8.050131
     },
     {
      "name": "void at::native::elementwise_kernel<(int)128, (int)4, void at::native::gpu_kernel_impl_noc",
      "ms": 2.523683
     }
    ]
   },
   {
    "unit": "non-graph: tfm_sub/_timestep_embeddings",
    "ms": 9.905904000000005,
    "share": 0.000996232856980607,
    "count": 2560,
    "families": {
     "gemm": 8.113582000000001,
     "elementwise_reduce": 1.638274,
     "gather_index_cat": 0.154048
    },
    "top_kernels": [
     {
      "name": "nvjet_sm90_tst_64x8_64x16_2x1_v_bz_bias_TNT",
      "ms": 6.19022
     },
     {
      "name": "nvjet_sm90_tst_64x8_64x16_4x1_v_bz_bias_TNT",
      "ms": 1.923362
     },
     {
      "name": "void at::native::elementwise_kernel<(int)128, (int)4, void at::native::gpu_kernel_impl<at:",
      "ms": 0.41145699999999996
     },
     {
      "name": "void at::native::vectorized_elementwise_kernel<(int)8, at::native::<unnamed>::silu_kernel(",
      "ms": 0.244384
     }
    ]
   },
   {
    "unit": "non-graph: no context",
    "ms": 6.3996900000000005,
    "share": 0.0006436142983507835,
    "count": 5864,
    "families": {
     "elementwise_reduce": 3.4380210000000004,
     "gather_index_cat": 1.6456970000000002,
     "conv": 1.264196,
     "gemm": 0.051776
    },
    "top_kernels": [
     {
      "name": "void at::native::elementwise_kernel<(int)128, (int)2, void at::native::gpu_kernel_impl_noc",
      "ms": 1.4932819999999996
     },
     {
      "name": "sm80_xmma_gemm_f32f32_f32f32_f32_tn_n_tilesize128x32x8_stage3_warpsize2x2x1_ffma_aligna4_a",
      "ms": 1.185059
     },
     {
      "name": "void at::native::<unnamed>::CatArrayBatchedCopy<at::native::<unnamed>::OpaqueType<(unsigne",
      "ms": 0.813537
     },
     {
      "name": "void at::native::reduce_kernel<(int)512, (int)1, at::native::ReduceOp<float, at::native::N",
      "ms": 0.784801
     }
    ]
   },
   {
    "unit": "non-graph: tfm_sub/patch_embedding/Conv3dLayer",
    "ms": 2.4152669999999996,
    "share": 0.00024290244926469896,
    "count": 320,
    "families": {
     "gemm": 2.1002259999999997,
     "elementwise_reduce": 0.31504099999999996
    },
    "top_kernels": [
     {
      "name": "nvjet_sm90_tst_256x144_64x4_2x1_v_bz_coopA_bias_TNT",
      "ms": 2.1002259999999997
     },
     {
      "name": "void at::native::elementwise_kernel<(int)128, (int)4, void at::native::gpu_kernel_impl_noc",
      "ms": 0.31504099999999996
     }
    ]
   },
   {
    "unit": "non-graph: gen/f=*",
    "ms": 1.4516499999999999,
    "share": 0.00014599186776248764,
    "count": 3200,
    "families": {
     "elementwise_reduce": 1.4516499999999999
    },
    "top_kernels": [
     {
      "name": "void at::native::<unnamed>::distribution_elementwise_grid_stride_kernel<float, (int)4, voi",
      "ms": 0.342112
     },
     {
      "name": "void at::native::elementwise_kernel<(int)128, (int)4, void at::native::gpu_kernel_impl_noc",
      "ms": 0.291232
     },
     {
      "name": "void at::native::vectorized_elementwise_kernel<(int)4, at::native::AUnaryFunctor<float, fl",
      "ms": 0.228224
     },
     {
      "name": "void at::native::vectorized_elementwise_kernel<(int)4, at::native::CUDAFunctor_add<float>,",
      "ms": 0.210401
     }
    ]
   },
   {
    "unit": "non-graph: tfm/f=15/c=0/s=*",
    "ms": 0.951778,
    "share": 9.571993794319911e-05,
    "count": 64,
    "families": {
     "elementwise_reduce": 0.951778
    },
    "top_kernels": [
     {
      "name": "void at::native::vectorized_elementwise_kernel<(int)8, at::native::CUDAFunctor_add<c10::BF",
      "ms": 0.523201
     },
     {
      "name": "void at::native::vectorized_elementwise_kernel<(int)8, at::native::<unnamed>::silu_kernel(",
      "ms": 0.428577
     }
    ]
   },
   {
    "unit": "non-graph: tfm/f=24/c=0/s=*",
    "ms": 0.949441,
    "share": 9.548490677524476e-05,
    "count": 64,
    "families": {
     "elementwise_reduce": 0.949441
    },
    "top_kernels": [
     {
      "name": "void at::native::vectorized_elementwise_kernel<(int)8, at::native::CUDAFunctor_add<c10::BF",
      "ms": 0.523329
     },
     {
      "name": "void at::native::vectorized_elementwise_kernel<(int)8, at::native::<unnamed>::silu_kernel(",
      "ms": 0.426112
     }
    ]
   },
   {
    "unit": "non-graph: tfm/f=21/c=0/s=*",
    "ms": 0.94909,
    "share": 9.544960684372914e-05,
    "count": 64,
    "families": {
     "elementwise_reduce": 0.94909
    },
    "top_kernels": [
     {
      "name": "void at::native::vectorized_elementwise_kernel<(int)8, at::native::CUDAFunctor_add<c10::BF",
      "ms": 0.523873
     },
     {
      "name": "void at::native::vectorized_elementwise_kernel<(int)8, at::native::<unnamed>::silu_kernel(",
      "ms": 0.425217
     }
    ]
   },
   {
    "unit": "non-graph: tfm/f=18/c=0/s=*",
    "ms": 0.945697,
    "share": 9.51083741724116e-05,
    "count": 64,
    "families": {
     "elementwise_reduce": 0.945697
    },
    "top_kernels": [
     {
      "name": "void at::native::vectorized_elementwise_kernel<(int)8, at::native::CUDAFunctor_add<c10::BF",
      "ms": 0.520385
     },
     {
      "name": "void at::native::vectorized_elementwise_kernel<(int)8, at::native::<unnamed>::silu_kernel(",
      "ms": 0.425312
     }
    ]
   },
   {
    "unit": "non-graph: tfm_sub/_rotary_embedding",
    "ms": 0.894465,
    "share": 8.995599214560915e-05,
    "count": 640,
    "families": {
     "gather_index_cat": 0.676449,
     "elementwise_reduce": 0.218016
    },
    "top_kernels": [
     {
      "name": "void at::native::<unnamed>::CatArrayBatchedCopy<at::native::<unnamed>::OpaqueType<(unsigne",
      "ms": 0.676449
     },
     {
      "name": "void at::native::vectorized_elementwise_kernel<(int)8, at::native::bfloat16_copy_kernel_cu",
      "ms": 0.218016
     }
    ]
   },
   {
    "unit": "non-graph: tfm_sub/_unpatchify",
    "ms": 0.281633,
    "share": 2.8323719693833012e-05,
    "count": 160,
    "families": {
     "elementwise_reduce": 0.281633
    },
    "top_kernels": [
     {
      "name": "void at::native::elementwise_kernel<(int)128, (int)4, void at::native::gpu_kernel_impl_noc",
      "ms": 0.281633
     }
    ]
   },
   {
    "unit": "non-graph: tfm/f=24/c=1/s=None",
    "ms": 0.23852800000000002,
    "share": 2.398866685058427e-05,
    "count": 16,
    "families": {
     "elementwise_reduce": 0.23852800000000002
    },
    "top_kernels": [
     {
      "name": "void at::native::vectorized_elementwise_kernel<(int)8, at::native::CUDAFunctor_add<c10::BF",
      "ms": 0.131136
     },
     {
      "name": "void at::native::vectorized_elementwise_kernel<(int)8, at::native::<unnamed>::silu_kernel(",
      "ms": 0.107392
     }
    ]
   },
   {
    "unit": "non-graph: tfm/f=15/c=1/s=None",
    "ms": 0.238528,
    "share": 2.398866685058427e-05,
    "count": 16,
    "families": {
     "elementwise_reduce": 0.238528
    },
    "top_kernels": [
     {
      "name": "void at::native::vectorized_elementwise_kernel<(int)8, at::native::CUDAFunctor_add<c10::BF",
      "ms": 0.131872
     },
     {
      "name": "void at::native::vectorized_elementwise_kernel<(int)8, at::native::<unnamed>::silu_kernel(",
      "ms": 0.106656
     }
    ]
   },
   {
    "unit": "non-graph: tfm/f=18/c=1/s=None",
    "ms": 0.238368,
    "share": 2.397257571371106e-05,
    "count": 16,
    "families": {
     "elementwise_reduce": 0.238368
    },
    "top_kernels": [
     {
      "name": "void at::native::vectorized_elementwise_kernel<(int)8, at::native::CUDAFunctor_add<c10::BF",
      "ms": 0.130688
     },
     {
      "name": "void at::native::vectorized_elementwise_kernel<(int)8, at::native::<unnamed>::silu_kernel(",
      "ms": 0.10768
     }
    ]
   },
   {
    "unit": "non-graph: tfm/f=21/c=1/s=None",
    "ms": 0.237664,
    "share": 2.3901774711468927e-05,
    "count": 16,
    "families": {
     "elementwise_reduce": 0.237664
    },
    "top_kernels": [
     {
      "name": "void at::native::vectorized_elementwise_kernel<(int)8, at::native::CUDAFunctor_add<c10::BF",
      "ms": 0.131136
     },
     {
      "name": "void at::native::vectorized_elementwise_kernel<(int)8, at::native::<unnamed>::silu_kernel(",
      "ms": 0.106528
     }
    ]
   }
  ],
  "by_family": {
   "gemm": 4339.366786999997,
   "attention_core": 3237.443275,
   "nccl": 1085.341714,
   "triton_fused": 749.659344,
   "gather_index_cat": 446.176112,
   "elementwise_reduce": 76.93501299999997,
   "attention_prep": 7.175626,
   "conv": 1.264196
  }
 },
 "vae_steady": {
  "gpu_ms": 2281.02625,
  "groups": 1555,
  "by_unit": [
   {
    "unit": "vae_frame/*",
    "ms": 154.195943,
    "share": 0.0675993724315974,
    "count": 4364,
    "families": {
     "attention_core": 92.169048,
     "elementwise_reduce": 62.026894999999975
    },
    "top_kernels": [
     {
      "name": "fmha_cutlassF_f32_aligned_32x128_gmem_sm80(PyTorchMemEffAttention::AttentionKernel<float, ",
      "ms": 92.169048
     },
     {
      "name": "void at::native::elementwise_kernel<(int)128, (int)2, void at::native::gpu_kernel_impl_noc",
      "ms": 43.53613899999998
     },
     {
      "name": "void at::native::vectorized_elementwise_kernel<(int)4, at::native::CUDAFunctor_add<float>,",
      "ms": 18.490756
     }
    ]
   },
   {
    "unit": "vae/up_blocks.3.resnets.0.conv1/WanDistCausalConv3d",
    "ms": 112.787742,
    "share": 0.04944605174973326,
    "count": 1347,
    "families": {
     "conv": 75.006289,
     "elementwise_reduce": 35.246725999999995,
     "nccl": 2.534727
    },
    "top_kernels": [
     {
      "name": "sm90_xmma_fprop_implicit_gemm_f32f32_tf32f32_f32_nhwckrsc_nhwc_tilesize256x32x32_warpgroup",
      "ms": 63.901886000000005
     },
     {
      "name": "void at::native::elementwise_kernel<(int)128, (int)2, void at::native::gpu_kernel_impl_noc",
      "ms": 32.809633
     },
     {
      "name": "void cudnn::engines_precompiled::nchwToNhwcKernel<float, float, float, (bool)0, (bool)1, (",
      "ms": 7.177483
     },
     {
      "name": "void cudnn::engines_precompiled::nhwcToNchwKernel<float, float, float, (bool)1, (bool)0, (",
      "ms": 3.92692
     }
    ]
   },
   {
    "unit": "vae/up_blocks.3.resnets.0.conv2/WanDistCausalConv3d",
    "ms": 111.56572400000003,
    "share": 0.0489103200807093,
    "count": 1347,
    "families": {
     "conv": 75.01999100000002,
     "elementwise_reduce": 34.86441799999999,
     "nccl": 1.6813150000000001
    },
    "top_kernels": [
     {
      "name": "sm90_xmma_fprop_implicit_gemm_f32f32_tf32f32_f32_nhwckrsc_nhwc_tilesize256x32x32_warpgroup",
      "ms": 63.936736
     },
     {
      "name": "void at::native::elementwise_kernel<(int)128, (int)2, void at::native::gpu_kernel_impl_noc",
      "ms": 32.430462999999996
     },
     {
      "name": "void cudnn::engines_precompiled::nchwToNhwcKernel<float, float, float, (bool)0, (bool)1, (",
      "ms": 7.165198
     },
     {
      "name": "void cudnn::engines_precompiled::nhwcToNchwKernel<float, float, float, (bool)1, (bool)0, (",
      "ms": 3.918057
     }
    ]
   },
   {
    "unit": "vae/up_blocks.3.resnets.1.conv1/WanDistCausalConv3d",
    "ms": 111.53347000000007,
    "share": 0.04889617995408868,
    "count": 1346,
    "families": {
     "conv": 75.00264300000002,
     "elementwise_reduce": 34.854215,
     "nccl": 1.676612
    },
    "top_kernels": [
     {
      "name": "sm90_xmma_fprop_implicit_gemm_f32f32_tf32f32_f32_nhwckrsc_nhwc_tilesize256x32x32_warpgroup",
      "ms": 63.903551
     },
     {
      "name": "void at::native::elementwise_kernel<(int)128, (int)2, void at::native::gpu_kernel_impl_noc",
      "ms": 32.421701
     },
     {
      "name": "void cudnn::engines_precompiled::nchwToNhwcKernel<float, float, float, (bool)0, (bool)1, (",
      "ms": 7.177647
     },
     {
      "name": "void cudnn::engines_precompiled::nhwcToNchwKernel<float, float, float, (bool)1, (bool)0, (",
      "ms": 3.921445
     }
    ]
   },
   {
    "unit": "vae/up_blocks.3.resnets.1.conv2/WanDistCausalConv3d",
    "ms": 111.46332199999999,
    "share": 0.048865427129565035,
    "count": 1347,
    "families": {
     "conv": 75.14046599999999,
     "elementwise_reduce": 34.834917999999995,
     "nccl": 1.487938
    },
    "top_kernels": [
     {
      "name": "sm90_xmma_fprop_implicit_gemm_f32f32_tf32f32_f32_nhwckrsc_nhwc_tilesize256x32x32_warpgroup",
      "ms": 64.076123
     },
     {
      "name": "void at::native::elementwise_kernel<(int)128, (int)2, void at::native::gpu_kernel_impl_noc",
      "ms": 32.395168
     },
     {
      "name": "void cudnn::engines_precompiled::nchwToNhwcKernel<float, float, float, (bool)0, (bool)1, (",
      "ms": 7.1455519999999995
     },
     {
      "name": "void cudnn::engines_precompiled::nhwcToNchwKernel<float, float, float, (bool)1, (bool)0, (",
      "ms": 3.918791
     }
    ]
   },
   {
    "unit": "vae/up_blocks.3.resnets.2.conv1/WanDistCausalConv3d",
    "ms": 111.37666699999998,
    "share": 0.04882743765004896,
    "count": 1346,
    "families": {
     "conv": 74.877844,
     "elementwise_reduce": 34.83254800000001,
     "nccl": 1.666275
    },
    "top_kernels": [
     {
      "name": "sm90_xmma_fprop_implicit_gemm_f32f32_tf32f32_f32_nhwckrsc_nhwc_tilesize256x32x32_warpgroup",
      "ms": 63.799006000000006
     },
     {
      "name": "void at::native::elementwise_kernel<(int)128, (int)2, void at::native::gpu_kernel_impl_noc",
      "ms": 32.40233800000001
     },
     {
      "name": "void cudnn::engines_precompiled::nchwToNhwcKernel<float, float, float, (bool)0, (bool)1, (",
      "ms": 7.161872
     },
     {
      "name": "void cudnn::engines_precompiled::nhwcToNchwKernel<float, float, float, (bool)1, (bool)0, (",
      "ms": 3.916966
     }
    ]
   },
   {
    "unit": "vae/up_blocks.3.resnets.2.conv2/WanDistCausalConv3d",
    "ms": 111.24233600000002,
    "share": 0.04876854705201224,
    "count": 1347,
    "families": {
     "conv": 74.92248599999999,
     "elementwise_reduce": 34.798083999999996,
     "nccl": 1.521766
    },
    "top_kernels": [
     {
      "name": "sm90_xmma_fprop_implicit_gemm_f32f32_tf32f32_f32_nhwckrsc_nhwc_tilesize256x32x32_warpgroup",
      "ms": 63.847456
     },
     {
      "name": "void at::native::elementwise_kernel<(int)128, (int)2, void at::native::gpu_kernel_impl_noc",
      "ms": 32.372189
     },
     {
      "name": "void cudnn::engines_precompiled::nchwToNhwcKernel<float, float, float, (bool)0, (bool)1, (",
      "ms": 7.155119000000001
     },
     {
      "name": "void cudnn::engines_precompiled::nhwcToNchwKernel<float, float, float, (bool)1, (bool)0, (",
      "ms": 3.919911
     }
    ]
   },
   {
    "unit": "vae/up_blocks.2.upsamplers.0.resample.1/WanDistConv2d",
    "ms": 84.59645700000002,
    "share": 0.037087015986773504,
    "count": 1083,
    "families": {
     "conv": 45.12537400000001,
     "elementwise_reduce": 20.951236,
     "gather_index_cat": 13.05228,
     "nccl": 5.467567
    },
    "top_kernels": [
     {
      "name": "sm90_xmma_fprop_implicit_gemm_f32f32_tf32f32_f32_nhwckrsc_nhwc_tilesize64x192x32_warpgroup",
      "ms": 32.085662
     },
     {
      "name": "void at::native::elementwise_kernel<(int)128, (int)2, void at::native::gpu_kernel_impl_noc",
      "ms": 17.734942000000004
     },
     {
      "name": "void at::native::<unnamed>::CatArrayBatchedCopy_alignedK_contig<at::native::<unnamed>::Opa",
      "ms": 13.05228
     },
     {
      "name": "void cudnn::engines_precompiled::nchwToNhwcKernel<float, float, float, (bool)0, (bool)1, (",
      "ms": 9.21375
     }
    ]
   },
   {
    "unit": "vae/up_blocks.2.resnets.1.conv1/WanDistCausalConv3d",
    "ms": 68.379845,
    "share": 0.02997766685061165,
    "count": 1347,
    "families": {
     "conv": 42.46629299999999,
     "elementwise_reduce": 18.893702000000005,
     "nccl": 7.01985
    },
    "top_kernels": [
     {
      "name": "sm90_xmma_fprop_implicit_gemm_f32f32_tf32f32_f32_nhwckrsc_nhwc_tilesize128x128x32_warpgrou",
      "ms": 36.798315
     },
     {
      "name": "void at::native::elementwise_kernel<(int)128, (int)2, void at::native::gpu_kernel_impl_noc",
      "ms": 17.640101
     },
     {
      "name": "ncclDevKernel_SendRecv(ncclDevKernelArgsStorage<(unsigned long)4096>)",
      "ms": 7.01985
     },
     {
      "name": "void cudnn::engines_precompiled::nchwToNhwcKernel<float, float, float, (bool)0, (bool)1, (",
      "ms": 3.77799
     }
    ]
   },
   {
    "unit": "vae/up_blocks.2.resnets.1.conv2/WanDistCausalConv3d",
    "ms": 67.85712099999999,
    "share": 0.029748505086252295,
    "count": 1347,
    "families": {
     "conv": 42.09041700000001,
     "elementwise_reduce": 18.887392000000002,
     "nccl": 6.879312
    },
    "top_kernels": [
     {
      "name": "sm90_xmma_fprop_implicit_gemm_f32f32_tf32f32_f32_nhwckrsc_nhwc_tilesize128x128x32_warpgrou",
      "ms": 36.433001000000004
     },
     {
      "name": "void at::native::elementwise_kernel<(int)128, (int)2, void at::native::gpu_kernel_impl_noc",
      "ms": 17.625979
     },
     {
      "name": "ncclDevKernel_SendRecv(ncclDevKernelArgsStorage<(unsigned long)4096>)",
      "ms": 6.879312
     },
     {
      "name": "void cudnn::engines_precompiled::nchwToNhwcKernel<float, float, float, (bool)0, (bool)1, (",
      "ms": 3.7702769999999997
     }
    ]
   },
   {
    "unit": "vae/up_blocks.2.resnets.0.conv1/WanDistCausalConv3d",
    "ms": 67.43639099999999,
    "share": 0.029564057406178464,
    "count": 1348,
    "families": {
     "conv": 41.542902,
     "elementwise_reduce": 19.06861,
     "nccl": 6.824879
    },
    "top_kernels": [
     {
      "name": "sm90_xmma_fprop_implicit_gemm_f32f32_tf32f32_f32_nhwckrsc_nhwc_tilesize128x128x32_warpgrou",
      "ms": 35.862406
     },
     {
      "name": "void at::native::elementwise_kernel<(int)128, (int)2, void at::native::gpu_kernel_impl_noc",
      "ms": 17.807679999999998
     },
     {
      "name": "ncclDevKernel_SendRecv(ncclDevKernelArgsStorage<(unsigned long)4096>)",
      "ms": 6.824879
     },
     {
      "name": "void cudnn::engines_precompiled::nchwToNhwcKernel<float, float, float, (bool)0, (bool)1, (",
      "ms": 3.784233
     }
    ]
   },
   {
    "unit": "vae/up_blocks.2.resnets.2.conv2/WanDistCausalConv3d",
    "ms": 67.33974399999998,
    "share": 0.029521687442220355,
    "count": 1346,
    "families": {
     "conv": 42.480209,
     "elementwise_reduce": 18.878399999999996,
     "nccl": 5.981135
    },
    "top_kernels": [
     {
      "name": "sm90_xmma_fprop_implicit_gemm_f32f32_tf32f32_f32_nhwckrsc_nhwc_tilesize128x128x32_warpgrou",
      "ms": 36.811783
     },
     {
      "name": "void at::native::elementwise_kernel<(int)128, (int)2, void at::native::gpu_kernel_impl_noc",
      "ms": 17.610559
     },
     {
      "name": "ncclDevKernel_SendRecv(ncclDevKernelArgsStorage<(unsigned long)4096>)",
      "ms": 5.981135
     },
     {
      "name": "void cudnn::engines_precompiled::nchwToNhwcKernel<float, float, float, (bool)0, (bool)1, (",
      "ms": 3.772454
     }
    ]
   },
   {
    "unit": "vae/up_blocks.2.resnets.2.conv1/WanDistCausalConv3d",
    "ms": 66.94579799999998,
    "share": 0.029348981845342632,
    "count": 1345,
    "families": {
     "conv": 42.41608500000001,
     "elementwise_reduce": 18.871879999999997,
     "nccl": 5.657833
    },
    "top_kernels": [
     {
      "name": "sm90_xmma_fprop_implicit_gemm_f32f32_tf32f32_f32_nhwckrsc_nhwc_tilesize128x128x32_warpgrou",
      "ms": 36.750342
     },
     {
      "name": "void at::native::elementwise_kernel<(int)128, (int)2, void at::native::gpu_kernel_impl_noc",
      "ms": 17.607908999999996
     },
     {
      "name": "ncclDevKernel_SendRecv(ncclDevKernelArgsStorage<(unsigned long)4096>)",
      "ms": 5.657833
     },
     {
      "name": "void cudnn::engines_precompiled::nchwToNhwcKernel<float, float, float, (bool)0, (bool)1, (",
      "ms": 3.776429
     }
    ]
   },
   {
    "unit": "vae/up_blocks.2.resnets.0.conv2/WanDistCausalConv3d",
    "ms": 65.38717399999999,
    "share": 0.028665682387478,
    "count": 1347,
    "families": {
     "conv": 41.12779499999999,
     "elementwise_reduce": 18.882181,
     "nccl": 5.377198
    },
    "top_kernels": [
     {
      "name": "sm90_xmma_fprop_implicit_gemm_f32f32_tf32f32_f32_nhwckrsc_nhwc_tilesize128x128x32_warpgrou",
      "ms": 35.449445999999995
     },
     {
      "name": "void at::native::elementwise_kernel<(int)128, (int)2, void at::native::gpu_kernel_impl_noc",
      "ms": 17.618879
     },
     {
      "name": "ncclDevKernel_SendRecv(ncclDevKernelArgsStorage<(unsigned long)4096>)",
      "ms": 5.377198
     },
     {
      "name": "void cudnn::engines_precompiled::nchwToNhwcKernel<float, float, float, (bool)0, (bool)1, (",
      "ms": 3.7799430000000003
     }
    ]
   },
   {
    "unit": "vae/conv_out/WanDistCausalConv3d",
    "ms": 65.30358400000003,
    "share": 0.028629036601398178,
    "count": 1343,
    "families": {
     "conv": 34.107075,
     "elementwise_reduce": 29.531965000000003,
     "nccl": 1.664544
    },
    "top_kernels": [
     {
      "name": "void at::native::elementwise_kernel<(int)128, (int)2, void at::native::gpu_kernel_impl_noc",
      "ms": 27.101589000000004
     },
     {
      "name": "sm80_xmma_fprop_implicit_gemm_indexed_wo_smem_tf32f32_tf32f32_f32_nhwckrsc_nhwc_tilesize12",
      "ms": 26.015925
     },
     {
      "name": "void cudnn::engines_precompiled::nchwToNhwcKernel<float, float, float, (bool)0, (bool)1, (",
      "ms": 7.185516
     },
     {
      "name": "void at::native::vectorized_elementwise_kernel<(int)4, at::native::FillFunctor<float>, std",
      "ms": 2.4303760000000003
     }
    ]
   },
   {
    "unit": "vae/up_blocks.1.upsamplers.0.resample.1/WanDistConv2d",
    "ms": 56.69377,
    "share": 0.024854501345611433,
    "count": 1159,
    "families": {
     "conv": 27.565616000000002,
     "elementwise_reduce": 11.080886,
     "nccl": 10.12786,
     "gather_index_cat": 7.919408
    },
    "top_kernels": [
     {
      "name": "sm90_xmma_fprop_implicit_gemm_f32f32_tf32f32_f32_nhwckrsc_nhwc_tilesize128x128x32_warpgrou",
      "ms": 20.90039
     },
     {
      "name": "ncclDevKernel_SendRecv(ncclDevKernelArgsStorage<(unsigned long)4096>)",
      "ms": 10.12786
     },
     {
      "name": "void at::native::elementwise_kernel<(int)128, (int)2, void at::native::gpu_kernel_impl_noc",
      "ms": 9.420053
     },
     {
      "name": "void at::native::<unnamed>::CatArrayBatchedCopy_alignedK_contig<at::native::<unnamed>::Opa",
      "ms": 7.919408
     }
    ]
   },
   {
    "unit": "vae/up_blocks.1.resnets.0.conv2/WanDistCausalConv3d",
    "ms": 46.90018500000001,
    "share": 0.02056100187360843,
    "count": 1343,
    "families": {
     "conv": 23.220370000000003,
     "nccl": 16.794816,
     "elementwise_reduce": 6.8849990000000005
    },
    "top_kernels": [
     {
      "name": "sm90_xmma_fprop_implicit_gemm_f32f32_tf32f32_f32_nhwckrsc_nhwc_tilesize128x128x32_warpgrou",
      "ms": 20.913224
     },
     {
      "name": "ncclDevKernel_SendRecv(ncclDevKernelArgsStorage<(unsigned long)4096>)",
      "ms": 16.794816
     },
     {
      "name": "void at::native::elementwise_kernel<(int)128, (int)2, void at::native::gpu_kernel_impl_noc",
      "ms": 6.416295
     },
     {
      "name": "void cudnn::engines_precompiled::nchwToNhwcKernel<float, float, float, (bool)0, (bool)1, (",
      "ms": 1.825128
     }
    ]
   },
   {
    "unit": "vae/up_blocks.1.resnets.1.conv1/WanDistCausalConv3d",
    "ms": 42.92661100000001,
    "share": 0.018818990355766407,
    "count": 1343,
    "families": {
     "conv": 23.030024999999995,
     "nccl": 13.003803000000001,
     "elementwise_reduce": 6.8927830000000005
    },
    "top_kernels": [
     {
      "name": "sm90_xmma_fprop_implicit_gemm_f32f32_tf32f32_f32_nhwckrsc_nhwc_tilesize128x128x32_warpgrou",
      "ms": 20.724867
     },
     {
      "name": "ncclDevKernel_SendRecv(ncclDevKernelArgsStorage<(unsigned long)4096>)",
      "ms": 13.003803000000001
     },
     {
      "name": "void at::native::elementwise_kernel<(int)128, (int)2, void at::native::gpu_kernel_impl_noc",
      "ms": 6.419951
     },
     {
      "name": "void cudnn::engines_precompiled::nchwToNhwcKernel<float, float, float, (bool)0, (bool)1, (",
      "ms": 1.8229480000000002
     }
    ]
   },
   {
    "unit": "vae/up_blocks.1.resnets.1.conv2/WanDistCausalConv3d",
    "ms": 40.659829999999985,
    "share": 0.017825235461450734,
    "count": 1343,
    "families": {
     "conv": 23.060211999999996,
     "nccl": 10.714454,
     "elementwise_reduce": 6.885164000000002
    },
    "top_kernels": [
     {
      "name": "sm90_xmma_fprop_implicit_gemm_f32f32_tf32f32_f32_nhwckrsc_nhwc_tilesize128x128x32_warpgrou",
      "ms": 20.756306
     },
     {
      "name": "ncclDevKernel_SendRecv(ncclDevKernelArgsStorage<(unsigned long)4096>)",
      "ms": 10.714454
     },
     {
      "name": "void at::native::elementwise_kernel<(int)128, (int)2, void at::native::gpu_kernel_impl_noc",
      "ms": 6.414284000000001
     },
     {
      "name": "void cudnn::engines_precompiled::nchwToNhwcKernel<float, float, float, (bool)0, (bool)1, (",
      "ms": 1.8215379999999999
     }
    ]
   },
   {
    "unit": "vae/up_blocks.1.resnets.2.conv1/WanDistCausalConv3d",
    "ms": 39.66871600000002,
    "share": 0.01739073191288352,
    "count": 1343,
    "families": {
     "conv": 22.841165999999998,
     "nccl": 9.938961,
     "elementwise_reduce": 6.888589
    },
    "top_kernels": [
     {
      "name": "sm90_xmma_fprop_implicit_gemm_f32f32_tf32f32_f32_nhwckrsc_nhwc_tilesize128x128x32_warpgrou",
      "ms": 20.534091
     },
     {
      "name": "ncclDevKernel_SendRecv(ncclDevKernelArgsStorage<(unsigned long)4096>)",
      "ms": 9.938961
     },
     {
      "name": "void at::native::elementwise_kernel<(int)128, (int)2, void at::native::gpu_kernel_impl_noc",
      "ms": 6.417901
     },
     {
      "name": "void cudnn::engines_precompiled::nchwToNhwcKernel<float, float, float, (bool)0, (bool)1, (",
      "ms": 1.8243539999999998
     }
    ]
   },
   {
    "unit": "vae/up_blocks.1.resnets.2.conv2/WanDistCausalConv3d",
    "ms": 39.56529199999999,
    "share": 0.017345390917794128,
    "count": 1341,
    "families": {
     "conv": 23.016617,
     "nccl": 9.658064,
     "elementwise_reduce": 6.890611000000002
    },
    "top_kernels": [
     {
      "name": "sm90_xmma_fprop_implicit_gemm_f32f32_tf32f32_f32_nhwckrsc_nhwc_tilesize128x128x32_warpgrou",
      "ms": 20.71991
     },
     {
      "name": "ncclDevKernel_SendRecv(ncclDevKernelArgsStorage<(unsigned long)4096>)",
      "ms": 9.658064
     },
     {
      "name": "void at::native::elementwise_kernel<(int)128, (int)2, void at::native::gpu_kernel_impl_noc",
      "ms": 6.41781
     },
     {
      "name": "void cudnn::engines_precompiled::nchwToNhwcKernel<float, float, float, (bool)0, (bool)1, (",
      "ms": 1.8155849999999998
     }
    ]
   },
   {
    "unit": "vae/up_blocks.1.resnets.0.conv1/WanDistCausalConv3d",
    "ms": 36.052712,
    "share": 0.015805478783946482,
    "count": 1339,
    "families": {
     "nccl": 20.826054999999997,
     "conv": 11.623223,
     "elementwise_reduce": 3.6034339999999996
    },
    "top_kernels": [
     {
      "name": "ncclDevKernel_SendRecv(ncclDevKernelArgsStorage<(unsigned long)4096>)",
      "ms": 20.826054999999997
     },
     {
      "name": "sm90_xmma_fprop_implicit_gemm_f32f32_tf32f32_f32_nhwckrsc_nhwc_tilesize128x128x32_warpgrou",
      "ms": 10.291826
     },
     {
      "name": "void at::native::elementwise_kernel<(int)128, (int)2, void at::native::gpu_kernel_impl_noc",
      "ms": 3.3291609999999996
     },
     {
      "name": "void cudnn::engines_precompiled::nchwToNhwcKernel<float, float, float, (bool)0, (bool)1, (",
      "ms": 0.8551699999999999
     }
    ]
   },
   {
    "unit": "ncclAllReduce",
    "ms": 29.287644,
    "share": 0.012839678631493172,
    "count": 100,
    "families": {
     "nccl": 29.287644
    },
    "top_kernels": [
     {
      "name": "ncclDevKernel_AllReduce_Sum_f32_RING_LL(ncclDevKernelArgsStorage<(unsigned long)4096>)",
      "ms": 29.287644
     }
    ]
   },
   {
    "unit": "vae/up_blocks.0.upsamplers.0.resample.1/WanDistConv2d",
    "ms": 29.284567,
    "share": 0.012838329677266976,
    "count": 1083,
    "families": {
     "nccl": 22.366062,
     "conv": 4.562085000000001,
     "elementwise_reduce": 1.498307,
     "gather_index_cat": 0.858113
    },
    "top_kernels": [
     {
      "name": "ncclDevKernel_SendRecv(ncclDevKernelArgsStorage<(unsigned long)4096>)",
      "ms": 22.366062
     },
     {
      "name": "sm90_xmma_fprop_implicit_gemm_f32f32_tf32f32_f32_nhwckrsc_nhwc_tilesize128x128x32_warpgrou",
      "ms": 3.521092
     },
     {
      "name": "void at::native::elementwise_kernel<(int)128, (int)2, void at::native::gpu_kernel_impl_noc",
      "ms": 1.2273630000000002
     },
     {
      "name": "void at::native::<unnamed>::CatArrayBatchedCopy_alignedK_contig<at::native::<unnamed>::Opa",
      "ms": 0.858113
     }
    ]
   },
   {
    "unit": "vae/up_blocks.0.resnets.2.conv2/WanDistCausalConv3d",
    "ms": 26.481267,
    "share": 0.011609365302131004,
    "count": 1391,
    "families": {
     "nccl": 19.351714,
     "conv": 5.525835,
     "elementwise_reduce": 1.6037179999999998
    },
    "top_kernels": [
     {
      "name": "ncclDevKernel_SendRecv(ncclDevKernelArgsStorage<(unsigned long)4096>)",
      "ms": 19.351714
     },
     {
      "name": "sm90_xmma_fprop_implicit_gemm_f32f32_tf32f32_f32_nhwckrsc_nhwc_tilesize64x64x32_warpgroups",
      "ms": 4.608778
     },
     {
      "name": "void at::native::elementwise_kernel<(int)128, (int)2, void at::native::gpu_kernel_impl_noc",
      "ms": 1.45767
     },
     {
      "name": "void cudnn::engines_precompiled::nchwToNhwcKernel<float, float, float, (bool)0, (bool)1, (",
      "ms": 0.792928
     }
    ]
   },
   {
    "unit": "vae/up_blocks.0.resnets.2.conv1/WanDistCausalConv3d",
    "ms": 25.964140000000004,
    "share": 0.011382657257889955,
    "count": 1392,
    "families": {
     "nccl": 18.880965,
     "conv": 5.481895,
     "elementwise_reduce": 1.6012799999999996
    },
    "top_kernels": [
     {
      "name": "ncclDevKernel_SendRecv(ncclDevKernelArgsStorage<(unsigned long)4096>)",
      "ms": 18.880965
     },
     {
      "name": "sm90_xmma_fprop_implicit_gemm_f32f32_tf32f32_f32_nhwckrsc_nhwc_tilesize64x64x32_warpgroups",
      "ms": 4.564004
     },
     {
      "name": "void at::native::elementwise_kernel<(int)128, (int)2, void at::native::gpu_kernel_impl_noc",
      "ms": 1.4542079999999997
     },
     {
      "name": "void cudnn::engines_precompiled::nchwToNhwcKernel<float, float, float, (bool)0, (bool)1, (",
      "ms": 0.793763
     }
    ]
   },
   {
    "unit": "vae/up_blocks.0.resnets.1.conv2/WanDistCausalConv3d",
    "ms": 25.88594100000001,
    "share": 0.011348374881700731,
    "count": 1391,
    "families": {
     "nccl": 18.757221,
     "conv": 5.527181000000001,
     "elementwise_reduce": 1.6015390000000003
    },
    "top_kernels": [
     {
      "name": "ncclDevKernel_SendRecv(ncclDevKernelArgsStorage<(unsigned long)4096>)",
      "ms": 18.757221
     },
     {
      "name": "sm90_xmma_fprop_implicit_gemm_f32f32_tf32f32_f32_nhwckrsc_nhwc_tilesize64x64x32_warpgroups",
      "ms": 4.604523
     },
     {
      "name": "void at::native::elementwise_kernel<(int)128, (int)2, void at::native::gpu_kernel_impl_noc",
      "ms": 1.4555550000000004
     },
     {
      "name": "void cudnn::engines_precompiled::nchwToNhwcKernel<float, float, float, (bool)0, (bool)1, (",
      "ms": 0.798017
     }
    ]
   },
   {
    "unit": "vae/up_blocks.0.resnets.1.conv1/WanDistCausalConv3d",
    "ms": 25.37557499999999,
    "share": 0.011124630854204327,
    "count": 1392,
    "families": {
     "nccl": 18.266342,
     "conv": 5.511180999999999,
     "elementwise_reduce": 1.598052
    },
    "top_kernels": [
     {
      "name": "ncclDevKernel_SendRecv(ncclDevKernelArgsStorage<(unsigned long)4096>)",
      "ms": 18.266342
     },
     {
      "name": "sm90_xmma_fprop_implicit_gemm_f32f32_tf32f32_f32_nhwckrsc_nhwc_tilesize64x64x32_warpgroups",
      "ms": 4.59345
     },
     {
      "name": "void at::native::elementwise_kernel<(int)128, (int)2, void at::native::gpu_kernel_impl_noc",
      "ms": 1.451236
     },
     {
      "name": "void cudnn::engines_precompiled::nchwToNhwcKernel<float, float, float, (bool)0, (bool)1, (",
      "ms": 0.791843
     }
    ]
   },
   {
    "unit": "vae/up_blocks.0.resnets.0.conv2/WanDistCausalConv3d",
    "ms": 24.513357999999997,
    "share": 0.010746635642619193,
    "count": 1390,
    "families": {
     "nccl": 17.422337,
     "conv": 5.487308000000001,
     "elementwise_reduce": 1.6037130000000002
    },
    "top_kernels": [
     {
      "name": "ncclDevKernel_SendRecv(ncclDevKernelArgsStorage<(unsigned long)4096>)",
      "ms": 17.422337
     },
     {
      "name": "sm90_xmma_fprop_implicit_gemm_f32f32_tf32f32_f32_nhwckrsc_nhwc_tilesize64x64x32_warpgroups",
      "ms": 4.561543
     },
     {
      "name": "void at::native::elementwise_kernel<(int)128, (int)2, void at::native::gpu_kernel_impl_noc",
      "ms": 1.457185
     },
     {
      "name": "void cudnn::engines_precompiled::nchwToNhwcKernel<float, float, float, (bool)0, (bool)1, (",
      "ms": 0.801253
     }
    ]
   },
   {
    "unit": "vae/up_blocks.0.resnets.0.conv1/WanDistCausalConv3d",
    "ms": 23.61159899999999,
    "share": 0.010351305251309577,
    "count": 1392,
    "families": {
     "nccl": 16.491007,
     "conv": 5.517386000000001,
     "elementwise_reduce": 1.603206
    },
    "top_kernels": [
     {
      "name": "ncclDevKernel_SendRecv(ncclDevKernelArgsStorage<(unsigned long)4096>)",
      "ms": 16.491007
     },
     {
      "name": "sm90_xmma_fprop_implicit_gemm_f32f32_tf32f32_f32_nhwckrsc_nhwc_tilesize64x64x32_warpgroups",
      "ms": 4.594602
     },
     {
      "name": "void at::native::elementwise_kernel<(int)128, (int)2, void at::native::gpu_kernel_impl_noc",
      "ms": 1.456358
     },
     {
      "name": "void cudnn::engines_precompiled::nchwToNhwcKernel<float, float, float, (bool)0, (bool)1, (",
      "ms": 0.798176
     }
    ]
   },
   {
    "unit": "vae/up_blocks.3.resnets.0.norm1/WanRMS_norm",
    "ms": 21.336487,
    "share": 0.009353898053562515,
    "count": 408,
    "families": {
     "elementwise_reduce": 21.336487
    },
    "top_kernels": [
     {
      "name": "void at::native::elementwise_kernel<(int)128, (int)2, void at::native::gpu_kernel_impl_noc",
      "ms": 12.046649000000002
     },
     {
      "name": "void at::native::vectorized_elementwise_kernel<(int)4, at::native::CUDAFunctorOnSelf_add<f",
      "ms": 3.429255
     },
     {
      "name": "void at::native::vectorized_elementwise_kernel<(int)4, at::native::AUnaryFunctor<float, fl",
      "ms": 3.4278429999999998
     },
     {
      "name": "void at::native::reduce_kernel<(int)128, (int)4, at::native::ReduceOp<float, at::native::N",
      "ms": 2.333635
     }
    ]
   },
   {
    "unit": "vae/up_blocks.3.resnets.0.norm2/WanRMS_norm",
    "ms": 20.551913,
    "share": 0.009009941468231678,
    "count": 407,
    "families": {
     "elementwise_reduce": 20.551913
    },
    "top_kernels": [
     {
      "name": "void at::native::elementwise_kernel<(int)128, (int)2, void at::native::gpu_kernel_impl_noc",
      "ms": 11.293562
     },
     {
      "name": "void at::native::vectorized_elementwise_kernel<(int)4, at::native::AUnaryFunctor<float, fl",
      "ms": 3.43991
     },
     {
      "name": "void at::native::vectorized_elementwise_kernel<(int)4, at::native::CUDAFunctorOnSelf_add<f",
      "ms": 3.4329
     },
     {
      "name": "void at::native::reduce_kernel<(int)128, (int)4, at::native::ReduceOp<float, at::native::N",
      "ms": 2.289285
     }
    ]
   },
   {
    "unit": "vae/up_blocks.3.resnets.2.norm2/WanRMS_norm",
    "ms": 20.53706600000001,
    "share": 0.009003432555850689,
    "count": 403,
    "families": {
     "elementwise_reduce": 20.53706600000001
    },
    "top_kernels": [
     {
      "name": "void at::native::elementwise_kernel<(int)128, (int)2, void at::native::gpu_kernel_impl_noc",
      "ms": 11.260791000000001
     },
     {
      "name": "void at::native::vectorized_elementwise_kernel<(int)4, at::native::AUnaryFunctor<float, fl",
      "ms": 3.445348
     },
     {
      "name": "void at::native::vectorized_elementwise_kernel<(int)4, at::native::CUDAFunctorOnSelf_add<f",
      "ms": 3.438056
     },
     {
      "name": "void at::native::reduce_kernel<(int)128, (int)4, at::native::ReduceOp<float, at::native::N",
      "ms": 2.292967
     }
    ]
   },
   {
    "unit": "vae/up_blocks.3.resnets.1.norm2/WanRMS_norm",
    "ms": 20.536649,
    "share": 0.009003249743399491,
    "count": 408,
    "families": {
     "elementwise_reduce": 20.536649
    },
    "top_kernels": [
     {
      "name": "void at::native::elementwise_kernel<(int)128, (int)2, void at::native::gpu_kernel_impl_noc",
      "ms": 11.27689
     },
     {
      "name": "void at::native::vectorized_elementwise_kernel<(int)4, at::native::AUnaryFunctor<float, fl",
      "ms": 3.438979
     },
     {
      "name": "void at::native::vectorized_elementwise_kernel<(int)4, at::native::CUDAFunctorOnSelf_add<f",
      "ms": 3.436871
     },
     {
      "name": "void at::native::reduce_kernel<(int)128, (int)4, at::native::ReduceOp<float, at::native::N",
      "ms": 2.285988
     }
    ]
   },
   {
    "unit": "vae/norm_out/WanRMS_norm",
    "ms": 20.510213,
    "share": 0.008991660223112295,
    "count": 408,
    "families": {
     "elementwise_reduce": 20.510213
    },
    "top_kernels": [
     {
      "name": "void at::native::elementwise_kernel<(int)128, (int)2, void at::native::gpu_kernel_impl_noc",
      "ms": 11.260917
     },
     {
      "name": "void at::native::vectorized_elementwise_kernel<(int)4, at::native::AUnaryFunctor<float, fl",
      "ms": 3.4526149999999998
     },
     {
      "name": "void at::native::vectorized_elementwise_kernel<(int)4, at::native::CUDAFunctorOnSelf_add<f",
      "ms": 3.442821
     },
     {
      "name": "void at::native::reduce_kernel<(int)128, (int)4, at::native::ReduceOp<float, at::native::N",
      "ms": 2.255843
     }
    ]
   },
   {
    "unit": "vae/up_blocks.3.resnets.2.norm1/WanRMS_norm",
    "ms": 20.499211999999996,
    "share": 0.008986837393914251,
    "count": 408,
    "families": {
     "elementwise_reduce": 20.499211999999996
    },
    "top_kernels": [
     {
      "name": "void at::native::elementwise_kernel<(int)128, (int)2, void at::native::gpu_kernel_impl_noc",
      "ms": 11.274168000000001
     },
     {
      "name": "void at::native::vectorized_elementwise_kernel<(int)4, at::native::AUnaryFunctor<float, fl",
      "ms": 3.4404869999999996
     },
     {
      "name": "void at::native::vectorized_elementwise_kernel<(int)4, at::native::CUDAFunctorOnSelf_add<f",
      "ms": 3.4363590000000004
     },
     {
      "name": "void at::native::reduce_kernel<(int)128, (int)4, at::native::ReduceOp<float, at::native::N",
      "ms": 2.24999
     }
    ]
   },
   {
    "unit": "vae/up_blocks.3.resnets.1.norm1/WanRMS_norm",
    "ms": 20.428428000000004,
    "share": 0.008955805747522637,
    "count": 408,
    "families": {
     "elementwise_reduce": 20.428428000000004
    },
    "top_kernels": [
     {
      "name": "void at::native::elementwise_kernel<(int)128, (int)2, void at::native::gpu_kernel_impl_noc",
      "ms": 11.275670999999999
     },
     {
      "name": "void at::native::vectorized_elementwise_kernel<(int)4, at::native::AUnaryFunctor<float, fl",
      "ms": 3.442502
     },
     {
      "name": "void at::native::vectorized_elementwise_kernel<(int)4, at::native::CUDAFunctorOnSelf_add<f",
      "ms": 3.440105
     },
     {
      "name": "void at::native::reduce_kernel<(int)128, (int)4, at::native::ReduceOp<float, at::native::N",
      "ms": 2.172228
     }
    ]
   },
   {
    "unit": "vae/mid_block.resnets.1.conv2/WanDistCausalConv3d",
    "ms": 19.30996,
    "share": 0.008465470311882646,
    "count": 1392,
    "families": {
     "nccl": 12.213689,
     "conv": 5.491468,
     "elementwise_reduce": 1.604803
    },
    "top_kernels": [
     {
      "name": "ncclDevKernel_SendRecv(ncclDevKernelArgsStorage<(unsigned long)4096>)",
      "ms": 12.213689
     },
     {
      "name": "sm90_xmma_fprop_implicit_gemm_f32f32_tf32f32_f32_nhwckrsc_nhwc_tilesize64x64x32_warpgroups",
      "ms": 4.565612
     },
     {
      "name": "void at::native::elementwise_kernel<(int)128, (int)2, void at::native::gpu_kernel_impl_noc",
      "ms": 1.4586910000000002
     },
     {
      "name": "void cudnn::engines_precompiled::nchwToNhwcKernel<float, float, float, (bool)0, (bool)1, (",
      "ms": 0.8011840000000001
     }
    ]
   },
   {
    "unit": "vae/up_blocks.2.upsamplers.0.resample.0/WanUpsample",
    "ms": 14.360158000000002,
    "share": 0.006295481255421766,
    "count": 136,
    "families": {
     "gather_index_cat": 11.468984,
     "elementwise_reduce": 2.891174
    },
    "top_kernels": [
     {
      "name": "void at::native::<unnamed>::upsample_nearest2d_out_frame<float, &at::native::nearest_neigh",
      "ms": 11.468984
     },
     {
      "name": "void at::native::elementwise_kernel<(int)128, (int)2, void at::native::gpu_kernel_impl_noc",
      "ms": 2.891174
     }
    ]
   },
   {
    "unit": "vae/mid_block.resnets.1.conv1/WanDistCausalConv3d",
    "ms": 13.749022999999996,
    "share": 0.00602756018261517,
    "count": 1392,
    "families": {
     "nccl": 6.636844,
     "conv": 5.512013999999999,
     "elementwise_reduce": 1.6001650000000005
    },
    "top_kernels": [
     {
      "name": "ncclDevKernel_SendRecv(ncclDevKernelArgsStorage<(unsigned long)4096>)",
      "ms": 6.636844
     },
     {
      "name": "sm90_xmma_fprop_implicit_gemm_f32f32_tf32f32_f32_nhwckrsc_nhwc_tilesize64x64x32_warpgroups",
      "ms": 4.5849709999999995
     },
     {
      "name": "void at::native::elementwise_kernel<(int)128, (int)2, void at::native::gpu_kernel_impl_noc",
      "ms": 1.4543390000000005
     },
     {
      "name": "void cudnn::engines_precompiled::nchwToNhwcKernel<float, float, float, (bool)0, (bool)1, (",
      "ms": 0.8032670000000001
     }
    ]
   },
   {
    "unit": "ncclAllGather",
    "ms": 12.348596999999998,
    "share": 0.005413614595623351,
    "count": 140,
    "families": {
     "nccl": 12.348596999999998
    },
    "top_kernels": [
     {
      "name": "ncclDevKernel_AllGather_RING_LL(ncclDevKernelArgsStorage<(unsigned long)4096>)",
      "ms": 12.348596999999998
     }
    ]
   },
   {
    "unit": "vae/mid_block.resnets.0.conv2/WanDistCausalConv3d",
    "ms": 11.929462999999997,
    "share": 0.005229866600614525,
    "count": 1391,
    "families": {
     "conv": 5.529996000000001,
     "nccl": 4.790024,
     "elementwise_reduce": 1.6094430000000002
    },
    "top_kernels": [
     {
      "name": "ncclDevKernel_SendRecv(ncclDevKernelArgsStorage<(unsigned long)4096>)",
      "ms": 4.790024
     },
     {
      "name": "sm90_xmma_fprop_implicit_gemm_f32f32_tf32f32_f32_nhwckrsc_nhwc_tilesize64x64x32_warpgroups",
      "ms": 4.611148
     },
     {
      "name": "void at::native::elementwise_kernel<(int)128, (int)2, void at::native::gpu_kernel_impl_noc",
      "ms": 1.462274
     },
     {
      "name": "void cudnn::engines_precompiled::nchwToNhwcKernel<float, float, float, (bool)0, (bool)1, (",
      "ms": 0.794656
     }
    ]
   },
   {
    "unit": "vae/mid_block.resnets.0.conv1/WanDistCausalConv3d",
    "ms": 11.369877,
    "share": 0.0049845445662889684,
    "count": 1390,
    "families": {
     "conv": 5.54225,
     "nccl": 4.224199,
     "elementwise_reduce": 1.6034279999999999
    },
    "top_kernels": [
     {
      "name": "sm90_xmma_fprop_implicit_gemm_f32f32_tf32f32_f32_nhwckrsc_nhwc_tilesize64x64x32_warpgroups",
      "ms": 4.614502
     },
     {
      "name": "ncclDevKernel_SendRecv(ncclDevKernelArgsStorage<(unsigned long)4096>)",
      "ms": 4.224199
     },
     {
      "name": "void at::native::elementwise_kernel<(int)128, (int)2, void at::native::gpu_kernel_impl_noc",
      "ms": 1.45818
     },
     {
      "name": "void cudnn::engines_precompiled::nchwToNhwcKernel<float, float, float, (bool)0, (bool)1, (",
      "ms": 0.802947
     }
    ]
   },
   {
    "unit": "vae/up_blocks.2.resnets.0.norm1/WanRMS_norm",
    "ms": 10.764088000000001,
    "share": 0.00471896717541063,
    "count": 408,
    "families": {
     "elementwise_reduce": 10.764088000000001
    },
    "top_kernels": [
     {
      "name": "void at::native::elementwise_kernel<(int)128, (int)2, void at::native::gpu_kernel_impl_noc",
      "ms": 6.1349920000000004
     },
     {
      "name": "void at::native::vectorized_elementwise_kernel<(int)4, at::native::AUnaryFunctor<float, fl",
      "ms": 1.738181
     },
     {
      "name": "void at::native::vectorized_elementwise_kernel<(int)4, at::native::CUDAFunctorOnSelf_add<f",
      "ms": 1.735459
     },
     {
      "name": "void at::native::reduce_kernel<(int)128, (int)4, at::native::ReduceOp<float, at::native::N",
      "ms": 1.0893760000000001
     }
    ]
   },
   {
    "unit": "vae/up_blocks.2.resnets.0.norm2/WanRMS_norm",
    "ms": 10.549773,
    "share": 0.004625011658677756,
    "count": 408,
    "families": {
     "elementwise_reduce": 10.549773
    },
    "top_kernels": [
     {
      "name": "void at::native::elementwise_kernel<(int)128, (int)2, void at::native::gpu_kernel_impl_noc",
      "ms": 5.7442020000000005
     },
     {
      "name": "void at::native::vectorized_elementwise_kernel<(int)4, at::native::CUDAFunctorOnSelf_add<f",
      "ms": 1.743393
     },
     {
      "name": "void at::native::vectorized_elementwise_kernel<(int)4, at::native::AUnaryFunctor<float, fl",
      "ms": 1.7414720000000001
     },
     {
      "name": "void at::native::reduce_kernel<(int)128, (int)4, at::native::ReduceOp<float, at::native::N",
      "ms": 1.25549
     }
    ]
   },
   {
    "unit": "vae/up_blocks.2.resnets.2.norm2/WanRMS_norm",
    "ms": 10.542966999999999,
    "share": 0.00462202791397074,
    "count": 408,
    "families": {
     "elementwise_reduce": 10.542966999999999
    },
    "top_kernels": [
     {
      "name": "void at::native::elementwise_kernel<(int)128, (int)2, void at::native::gpu_kernel_impl_noc",
      "ms": 5.751723999999999
     },
     {
      "name": "void at::native::vectorized_elementwise_kernel<(int)4, at::native::CUDAFunctorOnSelf_add<f",
      "ms": 1.736388
     },
     {
      "name": "void at::native::vectorized_elementwise_kernel<(int)4, at::native::AUnaryFunctor<float, fl",
      "ms": 1.735621
     },
     {
      "name": "void at::native::reduce_kernel<(int)128, (int)4, at::native::ReduceOp<float, at::native::N",
      "ms": 1.254178
     }
    ]
   },
   {
    "unit": "vae/up_blocks.2.resnets.2.norm1/WanRMS_norm",
    "ms": 10.539796999999998,
    "share": 0.0046206381886223355,
    "count": 408,
    "families": {
     "elementwise_reduce": 10.539796999999998
    },
    "top_kernels": [
     {
      "name": "void at::native::elementwise_kernel<(int)128, (int)2, void at::native::gpu_kernel_impl_noc",
      "ms": 5.731623999999999
     },
     {
      "name": "void at::native::vectorized_elementwise_kernel<(int)4, at::native::CUDAFunctorOnSelf_add<f",
      "ms": 1.7375720000000001
     },
     {
      "name": "void at::native::vectorized_elementwise_kernel<(int)4, at::native::AUnaryFunctor<float, fl",
      "ms": 1.7352379999999998
     },
     {
      "name": "void at::native::reduce_kernel<(int)128, (int)4, at::native::ReduceOp<float, at::native::N",
      "ms": 1.268419
     }
    ]
   },
   {
    "unit": "vae/up_blocks.2.resnets.1.norm2/WanRMS_norm",
    "ms": 10.536338999999998,
    "share": 0.004619122204314834,
    "count": 408,
    "families": {
     "elementwise_reduce": 10.536338999999998
    },
    "top_kernels": [
     {
      "name": "void at::native::elementwise_kernel<(int)128, (int)2, void at::native::gpu_kernel_impl_noc",
      "ms": 5.761832
     },
     {
      "name": "void at::native::vectorized_elementwise_kernel<(int)4, at::native::CUDAFunctorOnSelf_add<f",
      "ms": 1.7325810000000001
     },
     {
      "name": "void at::native::vectorized_elementwise_kernel<(int)4, at::native::AUnaryFunctor<float, fl",
      "ms": 1.729955
     },
     {
      "name": "void at::native::reduce_kernel<(int)128, (int)4, at::native::ReduceOp<float, at::native::N",
      "ms": 1.246786
     }
    ]
   },
   {
    "unit": "vae/up_blocks.2.resnets.1.norm1/WanRMS_norm",
    "ms": 10.473332999999997,
    "share": 0.004591500426617184,
    "count": 408,
    "families": {
     "elementwise_reduce": 10.473332999999997
    },
    "top_kernels": [
     {
      "name": "void at::native::elementwise_kernel<(int)128, (int)2, void at::native::gpu_kernel_impl_noc",
      "ms": 5.747437
     },
     {
      "name": "void at::native::vectorized_elementwise_kernel<(int)4, at::native::CUDAFunctorOnSelf_add<f",
      "ms": 1.7268489999999999
     },
     {
      "name": "void at::native::vectorized_elementwise_kernel<(int)4, at::native::AUnaryFunctor<float, fl",
      "ms": 1.7254770000000001
     },
     {
      "name": "void at::native::reduce_kernel<(int)128, (int)4, at::native::ReduceOp<float, at::native::N",
      "ms": 1.2069779999999999
     }
    ]
   },
   {
    "unit": "vae/up_blocks.1.upsamplers.0.time_conv/WanDistCausalConv3d",
    "ms": 9.828278000000003,
    "share": 0.004308708854183507,
    "count": 536,
    "families": {
     "conv": 6.862255000000001,
     "elementwise_reduce": 2.966023
    },
    "top_kernels": [
     {
      "name": "sm90_xmma_fprop_implicit_gemm_f32f32_tf32f32_f32_nhwckrsc_nhwc_tilesize128x128x32_warpgrou",
      "ms": 3.866282
     },
     {
      "name": "void at::native::elementwise_kernel<(int)128, (int)2, void at::native::gpu_kernel_impl_noc",
      "ms": 2.966023
     },
     {
      "name": "void cudnn::engines_precompiled::nchwToNhwcKernel<float, float, float, (bool)0, (bool)1, (",
      "ms": 1.871394
     },
     {
      "name": "void cudnn::engines_precompiled::nhwcToNchwKernel<float, float, float, (bool)1, (bool)0, (",
      "ms": 1.018658
     }
    ]
   },
   {
    "unit": "vae/up_blocks.1.upsamplers.0.resample.0/WanUpsample",
    "ms": 7.221773,
    "share": 0.0031660192424352853,
    "count": 136,
    "families": {
     "gather_index_cat": 5.755147,
     "elementwise_reduce": 1.466626
    },
    "top_kernels": [
     {
      "name": "void at::native::<unnamed>::upsample_nearest2d_out_frame<float, &at::native::nearest_neigh",
      "ms": 5.755147
     },
     {
      "name": "void at::native::elementwise_kernel<(int)128, (int)2, void at::native::gpu_kernel_impl_noc",
      "ms": 1.466626
     }
    ]
   },
   {
    "unit": "vae/up_blocks.3.resnets.2.nonlinearity/SiLU",
    "ms": 7.074217,
    "share": 0.003101330815460804,
    "count": 136,
    "families": {
     "elementwise_reduce": 7.074217
    },
    "top_kernels": [
     {
      "name": "void at::native::vectorized_elementwise_kernel<(int)4, at::native::<unnamed>::silu_kernel(",
      "ms": 7.074217
     }
    ]
   },
   {
    "unit": "vae/up_blocks.3.resnets.1.nonlinearity/SiLU",
    "ms": 7.06961,
    "share": 0.0030993111105144014,
    "count": 136,
    "families": {
     "elementwise_reduce": 7.06961
    },
    "top_kernels": [
     {
      "name": "void at::native::vectorized_elementwise_kernel<(int)4, at::native::<unnamed>::silu_kernel(",
      "ms": 7.06961
     }
    ]
   },
   {
    "unit": "vae/up_blocks.3.resnets.0.nonlinearity/SiLU",
    "ms": 7.068083,
    "share": 0.003098641674991684,
    "count": 136,
    "families": {
     "elementwise_reduce": 7.068083
    },
    "top_kernels": [
     {
      "name": "void at::native::vectorized_elementwise_kernel<(int)4, at::native::<unnamed>::silu_kernel(",
      "ms": 7.068083
     }
    ]
   },
   {
    "unit": "vae/conv_in/WanDistCausalConv3d",
    "ms": 4.687975,
    "share": 0.00205520431867016,
    "count": 1320,
    "families": {
     "nccl": 2.9603249999999997,
     "conv": 0.9608979999999998,
     "elementwise_reduce": 0.7667519999999998
    },
    "top_kernels": [
     {
      "name": "ncclDevKernel_SendRecv(ncclDevKernelArgsStorage<(unsigned long)4096>)",
      "ms": 2.9603249999999997
     },
     {
      "name": "sm80_xmma_fprop_implicit_gemm_tf32f32_tf32f32_f32_nhwckrsc_nchw_tilesize128x64x32_stage5_w",
      "ms": 0.762562
     },
     {
      "name": "void at::native::elementwise_kernel<(int)128, (int)2, void at::native::gpu_kernel_impl_noc",
      "ms": 0.701216
     },
     {
      "name": "void cudnn::engines_precompiled::nchwToNhwcKernel<float, float, float, (bool)0, (bool)1, (",
      "ms": 0.198336
     }
    ]
   },
   {
    "unit": "vae/up_blocks.2.resnets.1.nonlinearity/SiLU",
    "ms": 3.654343,
    "share": 0.0016020609144677751,
    "count": 136,
    "families": {
     "elementwise_reduce": 3.654343
    },
    "top_kernels": [
     {
      "name": "void at::native::vectorized_elementwise_kernel<(int)4, at::native::<unnamed>::silu_kernel(",
      "ms": 3.654343
     }
    ]
   },
   {
    "unit": "vae/up_blocks.2.resnets.0.nonlinearity/SiLU",
    "ms": 3.637352,
    "share": 0.0015946120742801623,
    "count": 136,
    "families": {
     "elementwise_reduce": 3.637352
    },
    "top_kernels": [
     {
      "name": "void at::native::vectorized_elementwise_kernel<(int)4, at::native::<unnamed>::silu_kernel(",
      "ms": 3.637352
     }
    ]
   },
   {
    "unit": "vae/up_blocks.2.resnets.2.nonlinearity/SiLU",
    "ms": 3.635718,
    "share": 0.001593895730046947,
    "count": 136,
    "families": {
     "elementwise_reduce": 3.635718
    },
    "top_kernels": [
     {
      "name": "void at::native::vectorized_elementwise_kernel<(int)4, at::native::<unnamed>::silu_kernel(",
      "ms": 3.635718
     }
    ]
   },
   {
    "unit": "vae/nonlinearity/SiLU",
    "ms": 3.5425649999999997,
    "share": 0.0015530575327662275,
    "count": 68,
    "families": {
     "elementwise_reduce": 3.5425649999999997
    },
    "top_kernels": [
     {
      "name": "void at::native::vectorized_elementwise_kernel<(int)4, at::native::<unnamed>::silu_kernel(",
      "ms": 3.5425649999999997
     }
    ]
   },
   {
    "unit": "vae/mid_block.attentions.0.to_qkv/Conv2d",
    "ms": 2.866434,
    "share": 0.0012566422679265528,
    "count": 144,
    "families": {
     "gemm": 1.93824,
     "elementwise_reduce": 0.928194
    },
    "top_kernels": [
     {
      "name": "void cutlass::Kernel2<cutlass_80_tensorop_s1688gemm_128x128_32x3_nn_align4>(T1::Params)",
      "ms": 1.93824
     },
     {
      "name": "void at::native::elementwise_kernel<(int)128, (int)2, void at::native::gpu_kernel_impl_noc",
      "ms": 0.928194
     }
    ]
   },
   {
    "unit": "vae/up_blocks.1.resnets.2.norm1/WanRMS_norm",
    "ms": 2.696513,
    "share": 0.0011821490436596247,
    "count": 408,
    "families": {
     "elementwise_reduce": 2.696513
    },
    "top_kernels": [
     {
      "name": "void at::native::elementwise_kernel<(int)128, (int)2, void at::native::gpu_kernel_impl_noc",
      "ms": 1.3391359999999999
     },
     {
      "name": "void at::native::reduce_kernel<(int)128, (int)4, at::native::ReduceOp<float, at::native::N",
      "ms": 0.545632
     },
     {
      "name": "void at::native::vectorized_elementwise_kernel<(int)4, at::native::AUnaryFunctor<float, fl",
      "ms": 0.378784
     },
     {
      "name": "void at::native::vectorized_elementwise_kernel<(int)4, at::native::CUDAFunctorOnSelf_add<f",
      "ms": 0.37449699999999997
     }
    ]
   },
   {
    "unit": "vae/up_blocks.1.resnets.1.norm1/WanRMS_norm",
    "ms": 2.689859,
    "share": 0.001179231935625467,
    "count": 408,
    "families": {
     "elementwise_reduce": 2.689859
    },
    "top_kernels": [
     {
      "name": "void at::native::elementwise_kernel<(int)128, (int)2, void at::native::gpu_kernel_impl_noc",
      "ms": 1.3335050000000002
     },
     {
      "name": "void at::native::reduce_kernel<(int)128, (int)4, at::native::ReduceOp<float, at::native::N",
      "ms": 0.54877
     },
     {
      "name": "void at::native::vectorized_elementwise_kernel<(int)4, at::native::CUDAFunctorOnSelf_add<f",
      "ms": 0.375136
     },
     {
      "name": "void at::native::vectorized_elementwise_kernel<(int)4, at::native::AUnaryFunctor<float, fl",
      "ms": 0.374336
     }
    ]
   },
   {
    "unit": "vae/up_blocks.1.resnets.0.norm2/WanRMS_norm",
    "ms": 2.665991,
    "share": 0.001168768224390228,
    "count": 407,
    "families": {
     "elementwise_reduce": 2.665991
    },
    "top_kernels": [
     {
      "name": "void at::native::elementwise_kernel<(int)128, (int)2, void at::native::gpu_kernel_impl_noc",
      "ms": 1.338532
     },
     {
      "name": "void at::native::reduce_kernel<(int)128, (int)4, at::native::ReduceOp<float, at::native::N",
      "ms": 0.519009
     },
     {
      "name": "void at::native::vectorized_elementwise_kernel<(int)4, at::native::AUnaryFunctor<float, fl",
      "ms": 0.376929
     },
     {
      "name": "void at::native::vectorized_elementwise_kernel<(int)4, at::native::CUDAFunctorOnSelf_add<f",
      "ms": 0.374176
     }
    ]
   },
   {
    "unit": "vae/up_blocks.1.resnets.1.norm2/WanRMS_norm",
    "ms": 2.6632369999999996,
    "share": 0.001167560873093854,
    "count": 408,
    "families": {
     "elementwise_reduce": 2.6632369999999996
    },
    "top_kernels": [
     {
      "name": "void at::native::elementwise_kernel<(int)128, (int)2, void at::native::gpu_kernel_impl_noc",
      "ms": 1.332225
     },
     {
      "name": "void at::native::reduce_kernel<(int)128, (int)4, at::native::ReduceOp<float, at::native::N",
      "ms": 0.522211
     },
     {
      "name": "void at::native::vectorized_elementwise_kernel<(int)4, at::native::AUnaryFunctor<float, fl",
      "ms": 0.37600100000000003
     },
     {
      "name": "void at::native::vectorized_elementwise_kernel<(int)4, at::native::CUDAFunctorOnSelf_add<f",
      "ms": 0.374112
     }
    ]
   },
   {
    "unit": "vae/up_blocks.1.resnets.2.norm2/WanRMS_norm",
    "ms": 2.6610579999999997,
    "share": 0.0011666056013165127,
    "count": 408,
    "families": {
     "elementwise_reduce": 2.6610579999999997
    },
    "top_kernels": [
     {
      "name": "void at::native::elementwise_kernel<(int)128, (int)2, void at::native::gpu_kernel_impl_noc",
      "ms": 1.3313920000000004
     },
     {
      "name": "void at::native::reduce_kernel<(int)128, (int)4, at::native::ReduceOp<float, at::native::N",
      "ms": 0.519841
     },
     {
      "name": "void at::native::vectorized_elementwise_kernel<(int)4, at::native::AUnaryFunctor<float, fl",
      "ms": 0.376768
     },
     {
      "name": "void at::native::vectorized_elementwise_kernel<(int)4, at::native::CUDAFunctorOnSelf_add<f",
      "ms": 0.37568
     }
    ]
   },
   {
    "unit": "vae/up_blocks.0.upsamplers.0.time_conv/WanDistCausalConv3d",
    "ms": 2.577031,
    "share": 0.001129768234802208,
    "count": 468,
    "families": {
     "conv": 1.9838440000000002,
     "elementwise_reduce": 0.593187
    },
    "top_kernels": [
     {
      "name": "void cudnn::engines_precompiled::nchwToNhwcKernel<float, float, float, (bool)0, (bool)1, (",
      "ms": 0.97197
     },
     {
      "name": "sm90_xmma_fprop_implicit_gemm_f32f32_tf32f32_f32_nhwckrsc_nhwc_tilesize128x128x32_warpgrou",
      "ms": 0.859202
     },
     {
      "name": "void at::native::elementwise_kernel<(int)128, (int)2, void at::native::gpu_kernel_impl_noc",
      "ms": 0.593187
     },
     {
      "name": "void cudnn::engines_precompiled::nhwcToNchwKernel<float, float, float, (bool)1, (bool)0, (",
      "ms": 0.152672
     }
    ]
   },
   {
    "unit": "vae/mid_block.attentions.0.proj/Conv2d",
    "ms": 2.111746,
    "share": 0.0009257876799971067,
    "count": 216,
    "families": {
     "elementwise_reduce": 1.22797,
     "gemm": 0.8837760000000001
    },
    "top_kernels": [
     {
      "name": "void at::native::elementwise_kernel<(int)128, (int)2, void at::native::gpu_kernel_impl_noc",
      "ms": 1.22797
     },
     {
      "name": "void cutlass::Kernel2<cutlass_80_tensorop_s1688gemm_64x128_32x3_nn_align4>(T1::Params)",
      "ms": 0.8837760000000001
     }
    ]
   },
   {
    "unit": "vae/up_blocks.1.resnets.0.conv_shortcut/WanDistCausalConv3d",
    "ms": 1.9152980000000002,
    "share": 0.0008396650411191017,
    "count": 204,
    "families": {
     "elementwise_reduce": 0.9855050000000001,
     "gemm": 0.929793
    },
    "top_kernels": [
     {
      "name": "void at::native::elementwise_kernel<(int)128, (int)2, void at::native::gpu_kernel_impl_noc",
      "ms": 0.9855050000000001
     },
     {
      "name": "void cutlass::Kernel2<cutlass_80_tensorop_s1688gemm_128x64_16x6_nn_align4>(T1::Params)",
      "ms": 0.929793
     }
    ]
   },
   {
    "unit": "vae/up_blocks.1.resnets.0.norm1/WanRMS_norm",
    "ms": 1.8379559999999997,
    "share": 0.0008057583730130242,
    "count": 408,
    "families": {
     "elementwise_reduce": 1.8379559999999997
    },
    "top_kernels": [
     {
      "name": "void at::native::elementwise_kernel<(int)128, (int)2, void at::native::gpu_kernel_impl_noc",
      "ms": 0.8131849999999999
     },
     {
      "name": "void at::native::reduce_kernel<(int)128, (int)4, at::native::ReduceOp<float, at::native::N",
      "ms": 0.532835
     },
     {
      "name": "void at::native::vectorized_elementwise_kernel<(int)4, at::native::AUnaryFunctor<float, fl",
      "ms": 0.21728
     },
     {
      "name": "void at::native::vectorized_elementwise_kernel<(int)4, at::native::CUDAFunctorOnSelf_add<f",
      "ms": 0.216
     }
    ]
   },
   {
    "unit": "vae/mid_block.attentions.0.norm/WanRMS_norm",
    "ms": 1.757348,
    "share": 0.0007704198932388437,
    "count": 432,
    "families": {
     "elementwise_reduce": 1.757348
    },
    "top_kernels": [
     {
      "name": "void at::native::elementwise_kernel<(int)128, (int)2, void at::native::gpu_kernel_impl_noc",
      "ms": 0.742401
     },
     {
      "name": "void at::native::reduce_kernel<(int)128, (int)4, at::native::ReduceOp<float, at::native::N",
      "ms": 0.525697
     },
     {
      "name": "void at::native::vectorized_elementwise_kernel<(int)4, at::native::CUDAFunctorOnSelf_add<f",
      "ms": 0.21769599999999997
     },
     {
      "name": "void at::native::vectorized_elementwise_kernel<(int)4, at::native::AUnaryFunctor<float, fl",
      "ms": 0.216386
     }
    ]
   },
   {
    "unit": "vae/mid_block.resnets.1.norm1/WanRMS_norm",
    "ms": 1.056834,
    "share": 0.00046331514159470987,
    "count": 432,
    "families": {
     "elementwise_reduce": 1.056834
    },
    "top_kernels": [
     {
      "name": "void at::native::reduce_kernel<(int)128, (int)4, at::native::ReduceOp<float, at::native::N",
      "ms": 0.497281
     },
     {
      "name": "void at::native::elementwise_kernel<(int)128, (int)2, void at::native::gpu_kernel_impl_noc",
      "ms": 0.316288
     },
     {
      "name": "void at::native::vectorized_elementwise_kernel<(int)4, at::native::AUnaryFunctor<float, fl",
      "ms": 0.08944
     },
     {
      "name": "void at::native::vectorized_elementwise_kernel<(int)4, at::native::CUDAFunctorOnSelf_add<f",
      "ms": 0.088289
     }
    ]
   },
   {
    "unit": "vae/mid_block.resnets.0.norm2/WanRMS_norm",
    "ms": 1.0556830000000001,
    "share": 0.00046281054415748185,
    "count": 430,
    "families": {
     "elementwise_reduce": 1.0556830000000001
    },
    "top_kernels": [
     {
      "name": "void at::native::reduce_kernel<(int)128, (int)4, at::native::ReduceOp<float, at::native::N",
      "ms": 0.496259
     },
     {
      "name": "void at::native::elementwise_kernel<(int)128, (int)2, void at::native::gpu_kernel_impl_noc",
      "ms": 0.317536
     },
     {
      "name": "void at::native::vectorized_elementwise_kernel<(int)4, at::native::CUDAFunctorOnSelf_add<f",
      "ms": 0.08976
     },
     {
      "name": "void at::native::vectorized_elementwise_kernel<(int)4, at::native::AUnaryFunctor<float, fl",
      "ms": 0.08915200000000001
     }
    ]
   },
   {
    "unit": "vae/mid_block.resnets.0.norm1/WanRMS_norm",
    "ms": 1.053665,
    "share": 0.0004619258546454694,
    "count": 432,
    "families": {
     "elementwise_reduce": 1.053665
    },
    "top_kernels": [
     {
      "name": "void at::native::reduce_kernel<(int)128, (int)4, at::native::ReduceOp<float, at::native::N",
      "ms": 0.496064
     },
     {
      "name": "void at::native::elementwise_kernel<(int)128, (int)2, void at::native::gpu_kernel_impl_noc",
      "ms": 0.317121
     },
     {
      "name": "void at::native::vectorized_elementwise_kernel<(int)4, at::native::AUnaryFunctor<float, fl",
      "ms": 0.08992
     },
     {
      "name": "void at::native::vectorized_elementwise_kernel<(int)4, at::native::CUDAFunctorOnSelf_add<f",
      "ms": 0.088128
     }
    ]
   },
   {
    "unit": "vae/up_blocks.0.resnets.2.norm2/WanRMS_norm",
    "ms": 1.051331,
    "share": 0.0004609026309977801,
    "count": 428,
    "families": {
     "elementwise_reduce": 1.051331
    },
    "top_kernels": [
     {
      "name": "void at::native::reduce_kernel<(int)128, (int)4, at::native::ReduceOp<float, at::native::N",
      "ms": 0.49580900000000006
     },
     {
      "name": "void at::native::elementwise_kernel<(int)128, (int)2, void at::native::gpu_kernel_impl_noc",
      "ms": 0.315904
     },
     {
      "name": "void at::native::vectorized_elementwise_kernel<(int)4, at::native::CUDAFunctorOnSelf_add<f",
      "ms": 0.08851300000000001
     },
     {
      "name": "void at::native::vectorized_elementwise_kernel<(int)4, at::native::AUnaryFunctor<float, fl",
      "ms": 0.08847999999999999
     }
    ]
   },
   {
    "unit": "vae/mid_block.resnets.1.norm2/WanRMS_norm",
    "ms": 1.0510419999999998,
    "share": 0.0004607759336395186,
    "count": 432,
    "families": {
     "elementwise_reduce": 1.0510419999999998
    },
    "top_kernels": [
     {
      "name": "void at::native::reduce_kernel<(int)128, (int)4, at::native::ReduceOp<float, at::native::N",
      "ms": 0.49462500000000004
     },
     {
      "name": "void at::native::elementwise_kernel<(int)128, (int)2, void at::native::gpu_kernel_impl_noc",
      "ms": 0.317025
     },
     {
      "name": "void at::native::vectorized_elementwise_kernel<(int)4, at::native::CUDAFunctorOnSelf_add<f",
      "ms": 0.08854399999999998
     },
     {
      "name": "void at::native::vectorized_elementwise_kernel<(int)4, at::native::AUnaryFunctor<float, fl",
      "ms": 0.088096
     }
    ]
   },
   {
    "unit": "vae/up_blocks.0.resnets.0.norm2/WanRMS_norm",
    "ms": 1.0506589999999998,
    "share": 0.0004606080267598849,
    "count": 432,
    "families": {
     "elementwise_reduce": 1.0506589999999998
    },
    "top_kernels": [
     {
      "name": "void at::native::reduce_kernel<(int)128, (int)4, at::native::ReduceOp<float, at::native::N",
      "ms": 0.49328
     },
     {
      "name": "void at::native::elementwise_kernel<(int)128, (int)2, void at::native::gpu_kernel_impl_noc",
      "ms": 0.315842
     },
     {
      "name": "void at::native::vectorized_elementwise_kernel<(int)4, at::native::AUnaryFunctor<float, fl",
      "ms": 0.089152
     },
     {
      "name": "void at::native::vectorized_elementwise_kernel<(int)4, at::native::CUDAFunctorOnSelf_add<f",
      "ms": 0.08870399999999999
     }
    ]
   },
   {
    "unit": "vae/up_blocks.0.resnets.1.norm2/WanRMS_norm",
    "ms": 1.0493140000000003,
    "share": 0.0004600183798849313,
    "count": 432,
    "families": {
     "elementwise_reduce": 1.0493140000000003
    },
    "top_kernels": [
     {
      "name": "void at::native::reduce_kernel<(int)128, (int)4, at::native::ReduceOp<float, at::native::N",
      "ms": 0.49353800000000003
     },
     {
      "name": "void at::native::elementwise_kernel<(int)128, (int)2, void at::native::gpu_kernel_impl_noc",
      "ms": 0.315264
     },
     {
      "name": "void at::native::vectorized_elementwise_kernel<(int)4, at::native::CUDAFunctorOnSelf_add<f",
      "ms": 0.08873600000000001
     },
     {
      "name": "void at::native::vectorized_elementwise_kernel<(int)4, at::native::AUnaryFunctor<float, fl",
      "ms": 0.08851200000000001
     }
    ]
   },
   {
    "unit": "vae/up_blocks.0.resnets.1.norm1/WanRMS_norm",
    "ms": 1.0391700000000001,
    "share": 0.0004555712587700384,
    "count": 432,
    "families": {
     "elementwise_reduce": 1.0391700000000001
    },
    "top_kernels": [
     {
      "name": "void at::native::reduce_kernel<(int)128, (int)4, at::native::ReduceOp<float, at::native::N",
      "ms": 0.482849
     },
     {
      "name": "void at::native::elementwise_kernel<(int)128, (int)2, void at::native::gpu_kernel_impl_noc",
      "ms": 0.315872
     },
     {
      "name": "void at::native::vectorized_elementwise_kernel<(int)4, at::native::AUnaryFunctor<float, fl",
      "ms": 0.089152
     },
     {
      "name": "void at::native::vectorized_elementwise_kernel<(int)4, at::native::CUDAFunctorOnSelf_add<f",
      "ms": 0.087681
     }
    ]
   },
   {
    "unit": "vae/up_blocks.0.resnets.0.norm1/WanRMS_norm",
    "ms": 1.037984,
    "share": 0.0004550513173620865,
    "count": 432,
    "families": {
     "elementwise_reduce": 1.037984
    },
    "top_kernels": [
     {
      "name": "void at::native::reduce_kernel<(int)128, (int)4, at::native::ReduceOp<float, at::native::N",
      "ms": 0.481952
     },
     {
      "name": "void at::native::elementwise_kernel<(int)128, (int)2, void at::native::gpu_kernel_impl_noc",
      "ms": 0.315744
     },
     {
      "name": "void at::native::vectorized_elementwise_kernel<(int)4, at::native::AUnaryFunctor<float, fl",
      "ms": 0.08902399999999999
     },
     {
      "name": "void at::native::vectorized_elementwise_kernel<(int)4, at::native::CUDAFunctorOnSelf_add<f",
      "ms": 0.08736000000000001
     }
    ]
   },
   {
    "unit": "vae/up_blocks.0.resnets.2.norm1/WanRMS_norm",
    "ms": 1.0352659999999998,
    "share": 0.00045385974843559993,
    "count": 432,
    "families": {
     "elementwise_reduce": 1.0352659999999998
    },
    "top_kernels": [
     {
      "name": "void at::native::reduce_kernel<(int)128, (int)4, at::native::ReduceOp<float, at::native::N",
      "ms": 0.481409
     },
     {
      "name": "void at::native::elementwise_kernel<(int)128, (int)2, void at::native::gpu_kernel_impl_noc",
      "ms": 0.316001
     },
     {
      "name": "void at::native::vectorized_elementwise_kernel<(int)4, at::native::CUDAFunctorOnSelf_add<f",
      "ms": 0.088352
     },
     {
      "name": "void at::native::vectorized_elementwise_kernel<(int)4, at::native::AUnaryFunctor<float, fl",
      "ms": 0.08678399999999999
     }
    ]
   },
   {
    "unit": "vae/up_blocks.0.upsamplers.0.resample.0/WanUpsample",
    "ms": 0.830498,
    "share": 0.0003640896285169888,
    "count": 136,
    "families": {
     "gather_index_cat": 0.618497,
     "elementwise_reduce": 0.212001
    },
    "top_kernels": [
     {
      "name": "void at::native::<unnamed>::upsample_nearest2d_out_frame<float, &at::native::nearest_neigh",
      "ms": 0.618497
     },
     {
      "name": "void at::native::elementwise_kernel<(int)128, (int)2, void at::native::gpu_kernel_impl_noc",
      "ms": 0.212001
     }
    ]
   },
   {
    "unit": "vae/up_blocks.1.resnets.1.nonlinearity/SiLU",
    "ms": 0.771906,
    "share": 0.0003384029447271815,
    "count": 136,
    "families": {
     "elementwise_reduce": 0.771906
    },
    "top_kernels": [
     {
      "name": "void at::native::vectorized_elementwise_kernel<(int)4, at::native::<unnamed>::silu_kernel(",
      "ms": 0.771906
     }
    ]
   },
   {
    "unit": "vae/up_blocks.1.resnets.2.nonlinearity/SiLU",
    "ms": 0.7701439999999999,
    "share": 0.0003376304854010338,
    "count": 136,
    "families": {
     "elementwise_reduce": 0.7701439999999999
    },
    "top_kernels": [
     {
      "name": "void at::native::vectorized_elementwise_kernel<(int)4, at::native::<unnamed>::silu_kernel(",
      "ms": 0.7701439999999999
     }
    ]
   },
   {
    "unit": "vae_convert/ep*_chunk*",
    "ms": 0.628737,
    "share": 0.0002756377748831255,
    "count": 25,
    "families": {
     "elementwise_reduce": 0.628737
    },
    "top_kernels": [
     {
      "name": "void at::native::unrolled_elementwise_kernel<at::native::direct_copy_kernel_cuda(at::Tenso",
      "ms": 0.16768
     },
     {
      "name": "void at::native::vectorized_elementwise_kernel<(int)4, at::native::<unnamed>::launch_clamp",
      "ms": 0.118848
     },
     {
      "name": "void at::native::vectorized_elementwise_kernel<(int)4, at::native::round_kernel_cuda(at::T",
      "ms": 0.115072
     },
     {
      "name": "void at::native::vectorized_elementwise_kernel<(int)4, at::native::CUDAFunctorOnSelf_add<f",
      "ms": 0.113825
     }
    ]
   },
   {
    "unit": "vae/up_blocks.1.resnets.0.nonlinearity/SiLU",
    "ms": 0.6046090000000001,
    "share": 0.00026506007986536764,
    "count": 140,
    "families": {
     "elementwise_reduce": 0.6046090000000001
    },
    "top_kernels": [
     {
      "name": "void at::native::vectorized_elementwise_kernel<(int)4, at::native::<unnamed>::silu_kernel(",
      "ms": 0.6046090000000001
     }
    ]
   },
   {
    "unit": "no context",
    "ms": 0.357504,
    "share": 0.00015672945456020072,
    "count": 92,
    "families": {
     "elementwise_reduce": 0.357504
    },
    "top_kernels": [
     {
      "name": "void at::native::vectorized_elementwise_kernel<(int)4, at::native::AbsFunctor<float>, std:",
      "ms": 0.106912
     },
     {
      "name": "void at::native::vectorized_elementwise_kernel<(int)8, at::native::AUnaryFunctor<float, fl",
      "ms": 0.076928
     },
     {
      "name": "void at::native::vectorized_elementwise_kernel<(int)8, at::native::BinaryFunctor<float, fl",
      "ms": 0.073472
     },
     {
      "name": "void at::native::vectorized_elementwise_kernel<(int)8, at::native::BinaryFunctor<bool, boo",
      "ms": 0.050848
     }
    ]
   },
   {
    "unit": "vae_decode/ep*_chunk*",
    "ms": 0.300032,
    "share": 0.0001315337778335519,
    "count": 22,
    "families": {
     "elementwise_reduce": 0.300032
    },
    "top_kernels": [
     {
      "name": "void at::native::elementwise_kernel<(int)128, (int)2, void at::native::gpu_kernel_impl_noc",
      "ms": 0.19328
     },
     {
      "name": "void at::native::vectorized_elementwise_kernel<(int)4, at::native::<unnamed>::launch_clamp",
      "ms": 0.10675199999999999
     }
    ]
   },
   {
    "unit": "vae/mid_block.resnets.0.nonlinearity/SiLU",
    "ms": 0.209985,
    "share": 9.205724835477014e-05,
    "count": 144,
    "families": {
     "elementwise_reduce": 0.209985
    },
    "top_kernels": [
     {
      "name": "void at::native::vectorized_elementwise_kernel<(int)4, at::native::<unnamed>::silu_kernel(",
      "ms": 0.209985
     }
    ]
   },
   {
    "unit": "vae/up_blocks.0.resnets.2.nonlinearity/SiLU",
    "ms": 0.2096,
    "share": 9.18884646768094e-05,
    "count": 144,
    "families": {
     "elementwise_reduce": 0.2096
    },
    "top_kernels": [
     {
      "name": "void at::native::vectorized_elementwise_kernel<(int)4, at::native::<unnamed>::silu_kernel(",
      "ms": 0.2096
     }
    ]
   },
   {
    "unit": "vae/up_blocks.0.resnets.0.nonlinearity/SiLU",
    "ms": 0.209218,
    "share": 9.172099619633926e-05,
    "count": 144,
    "families": {
     "elementwise_reduce": 0.209218
    },
    "top_kernels": [
     {
      "name": "void at::native::vectorized_elementwise_kernel<(int)4, at::native::<unnamed>::silu_kernel(",
      "ms": 0.209218
     }
    ]
   },
   {
    "unit": "vae/mid_block.resnets.1.nonlinearity/SiLU",
    "ms": 0.20912000000000003,
    "share": 9.16780330783129e-05,
    "count": 144,
    "families": {
     "elementwise_reduce": 0.20912000000000003
    },
    "top_kernels": [
     {
      "name": "void at::native::vectorized_elementwise_kernel<(int)4, at::native::<unnamed>::silu_kernel(",
      "ms": 0.20912000000000003
     }
    ]
   },
   {
    "unit": "vae/up_blocks.0.resnets.1.nonlinearity/SiLU",
    "ms": 0.20860800000000002,
    "share": 9.145357270658328e-05,
    "count": 144,
    "families": {
     "elementwise_reduce": 0.20860800000000002
    },
    "top_kernels": [
     {
      "name": "void at::native::vectorized_elementwise_kernel<(int)4, at::native::<unnamed>::silu_kernel(",
      "ms": 0.20860800000000002
     }
    ]
   },
   {
    "unit": "vae_pqc/root/WanCausalConv3d",
    "ms": 0.10464,
    "share": 4.5874088472239193e-05,
    "count": 48,
    "families": {
     "gemm": 0.06704,
     "elementwise_reduce": 0.0376
    },
    "top_kernels": [
     {
      "name": "void cutlass::Kernel2<cutlass_80_tensorop_s1688gemm_128x64_16x6_nn_align4>(T1::Params)",
      "ms": 0.06704
     },
     {
      "name": "void at::native::elementwise_kernel<(int)128, (int)2, void at::native::gpu_kernel_impl_noc",
      "ms": 0.0376
     }
    ]
   },
   {
    "unit": "vae_h2d/ep*_chunk*",
    "ms": 0.08748800000000001,
    "share": 3.835466601929724e-05,
    "count": 160,
    "families": {
     "elementwise_reduce": 0.08748800000000001
    },
    "top_kernels": [
     {
      "name": "void at::native::elementwise_kernel<(int)128, (int)2, void at::native::gpu_kernel_impl_noc",
      "ms": 0.08748800000000001
     }
    ]
   },
   {
    "unit": "vae_barrier/ep*_chunk*",
    "ms": 0.012896,
    "share": 5.653595612939571e-06,
    "count": 80,
    "families": {
     "elementwise_reduce": 0.012896
    },
    "top_kernels": [
     {
      "name": "void at::native::vectorized_elementwise_kernel<(int)4, at::native::FillFunctor<float>, std",
      "ms": 0.012896
     }
    ]
   },
   {
    "unit": "vae_barrier2/ep*_chunk*",
    "ms": 0.012064,
    "share": 5.2888475088789534e-06,
    "count": 20,
    "families": {
     "elementwise_reduce": 0.012064
    },
    "top_kernels": [
     {
      "name": "void at::native::vectorized_elementwise_kernel<(int)4, at::native::FillFunctor<float>, std",
      "ms": 0.012064
     }
    ]
   }
  ],
  "by_family": {
   "conv": 1005.1786940000004,
   "elementwise_reduce": 789.6852959999994,
   "nccl": 350.5019339999999,
   "attention_core": 92.169048,
   "gather_index_cat": 39.67242900000001,
   "gemm": 3.818849
  }
 },
 "dit_measured": {
  "gpu_ms": 29800.088928,
  "groups": 2019,
  "by_unit": [
   {
    "unit": "non-graph: no context",
    "ms": 7035.666609999995,
    "share": 0.23609549041947056,
    "count": 124888,
    "families": {
     "elementwise_reduce": 3831.153821999999,
     "conv": 3063.4081859999997,
     "attention_core": 65.379479,
     "gemm": 51.587457,
     "gather_index_cat": 19.535740999999998,
     "other": 4.601925
    },
    "top_kernels": [
     {
      "name": "sm80_xmma_fprop_implicit_gemm_bf16bf16_bf16f32_f32_nhwckrsc_nhwc_tilesize256x32x32_stage4_",
      "ms": 1267.49115
     },
     {
      "name": "void at::native::elementwise_kernel<(int)128, (int)2, void at::native::gpu_kernel_impl_noc",
      "ms": 1228.813669
     },
     {
      "name": "void at::native::elementwise_kernel<(int)128, (int)4, void at::native::gpu_kernel_impl<at:",
      "ms": 971.3213580000001
     },
     {
      "name": "sm90_xmma_fprop_implicit_gemm_bf16bf16_bf16f32_f32_nhwckrsc_nhwc_tilesize256x128x64_warpgr",
      "ms": 775.493553
     }
    ]
   },
   {
    "unit": "vllm_omni.ar_diffusion_paged_write_attn: ar_diffusion_paged_write_attn",
    "ms": 7016.660940999996,
    "share": 0.23545771819516884,
    "count": 142560,
    "families": {
     "attention_core": 5925.231643999999,
     "gather_index_cat": 1052.9008859999992,
     "elementwise_reduce": 20.68831999999999,
     "attention_prep": 17.840091
    },
    "top_kernels": [
     {
      "name": "void cutlass::device_kernel<flash::enable_sm90_or_later<flash::FlashAttnFwdSm90<flash::Col",
      "ms": 5925.231643999999
     },
     {
      "name": "void at::native::vectorized_gather_kernel<(int)16, long>(char *, char *, T2 *, int, long, ",
      "ms": 566.7460310000001
     },
     {
      "name": "void at::native::index_elementwise_kernel<(int)128, (int)4, void at::native::gpu_index_ker",
      "ms": 474.67772
     },
     {
      "name": "void flash::prepare_varlen_num_blocks_kernel<(int)1, (bool)0>(int, int, int, const int *, ",
      "ms": 17.840091
     }
    ]
   },
   {
    "unit": "addmm: float_1, modulation, chunk, float_2, layer_norm, scale_msa, add_1, mul",
    "ms": 2120.3969540000003,
    "share": 0.0711540478662024,
    "count": 15840,
    "families": {
     "gemm": 2120.3969540000003
    },
    "top_kernels": [
     {
      "name": "nvjet_sm90_tst_256x152_64x4_1x2_h_bz_coopA_bias_TNT",
      "ms": 2120.3969540000003
     }
    ]
   },
   {
    "unit": "addmm: output_parallel_6, input_1, output_parallel_7",
    "ms": 1793.1134260000001,
    "share": 0.06017141191532487,
    "count": 15840,
    "families": {
     "gemm": 1793.1134260000001
    },
    "top_kernels": [
     {
      "name": "nvjet_sm90_tst_128x200_64x5_2x1_v_bz_coopA_bias_TNT",
      "ms": 1793.1134260000001
     }
    ]
   },
   {
    "unit": "addmm: output_parallel_6",
    "ms": 1739.3397689999997,
    "share": 0.0583669321659549,
    "count": 15840,
    "families": {
     "gemm": 1739.3397689999997
    },
    "top_kernels": [
     {
      "name": "nvjet_sm90_tst_128x200_64x5_2x1_v_bz_coopA_bias_TNT",
      "ms": 1739.3397689999997
     }
    ]
   },
   {
    "unit": "_c10d_functional.all_to_all_single: output_parallel, split, value_7, stack, reshape, transpose, input_t, t",
    "ms": 1076.143541,
    "share": 0.03611209159811806,
    "count": 15840,
    "families": {
     "nccl": 1076.143541
    },
    "top_kernels": [
     {
      "name": "ncclDevKernel_SendRecv(ncclDevKernelArgsStorage<(unsigned long)4096>)",
      "ms": 1076.143541
     }
    ]
   },
   {
    "unit": "addmm: output_parallel_2, input_parallel_1, output_parallel_3",
    "ms": 771.477926,
    "share": 0.025888443751425304,
    "count": 15840,
    "families": {
     "gemm": 771.477926
    },
    "top_kernels": [
     {
      "name": "nvjet_sm90_tst_128x200_64x5_2x1_v_bz_coopA_bias_TNT",
      "ms": 771.477926
     }
    ]
   },
   {
    "unit": "addmm: contiguous_12, input_parallel_2, output_parallel_5",
    "ms": 730.049711,
    "share": 0.024498239342972204,
    "count": 15840,
    "families": {
     "gemm": 730.049711
    },
    "top_kernels": [
     {
      "name": "nvjet_sm90_tst_128x200_64x5_2x1_v_bz_coopA_bias_TNT",
      "ms": 730.049711
     }
    ]
   },
   {
    "unit": "_c10d_functional.all_to_all_single: output_5, reshape_3, transpose_3, transpose_4, contiguous_7, tensor_1",
    "ms": 710.5049840000001,
    "share": 0.023842377978020513,
    "count": 15840,
    "families": {
     "nccl": 710.5049840000001
    },
    "top_kernels": [
     {
      "name": "ncclDevKernel_SendRecv(ncclDevKernelArgsStorage<(unsigned long)4096>)",
      "ms": 710.5049840000001
     }
    ]
   },
   {
    "unit": "addmm: float_8, layer_norm_1, float_9, float_10, to_6, output_parallel_4",
    "ms": 695.2959430000001,
    "share": 0.02333200899768805,
    "count": 15840,
    "families": {
     "gemm": 695.2959430000001
    },
    "top_kernels": [
     {
      "name": "nvjet_sm90_tst_128x200_64x5_2x1_v_bz_coopA_bias_TNT",
      "ms": 695.2959430000001
     }
    ]
   },
   {
    "unit": "addmm: contiguous_8, input_parallel, output_parallel_1",
    "ms": 692.878441,
    "share": 0.023250885011587168,
    "count": 15840,
    "families": {
     "gemm": 692.878441
    },
    "top_kernels": [
     {
      "name": "nvjet_sm90_tst_128x200_64x5_2x1_v_bz_coopA_bias_TNT",
      "ms": 692.878441
     }
    ]
   },
   {
    "unit": "addmm: output_parallel_2",
    "ms": 690.2759149999999,
    "share": 0.023163552184954066,
    "count": 15840,
    "families": {
     "gemm": 690.2759149999999
    },
    "top_kernels": [
     {
      "name": "nvjet_sm90_tst_128x200_64x5_2x1_v_bz_coopA_bias_TNT",
      "ms": 690.2759149999999
     }
    ]
   },
   {
    "unit": "addmm: output_parallel_3, camera_features, camera_scale",
    "ms": 689.1163670000001,
    "share": 0.023124641294359028,
    "count": 15840,
    "families": {
     "gemm": 689.1163670000001
    },
    "top_kernels": [
     {
      "name": "nvjet_sm90_tst_128x200_64x5_2x1_v_bz_coopA_bias_TNT",
      "ms": 689.1163670000001
     }
    ]
   },
   {
    "unit": "addmm: camera_shift",
    "ms": 682.6895989999998,
    "share": 0.022908978582226593,
    "count": 15840,
    "families": {
     "gemm": 682.6895989999998
    },
    "top_kernels": [
     {
      "name": "nvjet_sm90_tst_128x200_64x5_2x1_v_bz_coopA_bias_TNT",
      "ms": 682.6895989999998
     }
    ]
   },
   {
    "unit": "_c10d_functional.all_to_all_single: output_parallel_4, value_float_2, truediv_4, rms_2, truediv_5, float_1",
    "ms": 486.16485600000004,
    "share": 0.01631420822852653,
    "count": 15840,
    "families": {
     "nccl": 486.16485600000004
    },
    "top_kernels": [
     {
      "name": "ncclDevKernel_SendRecv(ncclDevKernelArgsStorage<(unsigned long)4096>)",
      "ms": 486.16485600000004
     }
    ]
   },
   {
    "unit": "_c10d_functional.all_to_all_single: reshape_10, transpose_8, transpose_9, contiguous_11, tensor_3",
    "ms": 411.7792360000001,
    "share": 0.013818053932486576,
    "count": 15840,
    "families": {
     "nccl": 411.7792360000001
    },
    "top_kernels": [
     {
      "name": "ncclDevKernel_SendRecv(ncclDevKernelArgsStorage<(unsigned long)4096>)",
      "ms": 411.7792360000001
     }
    ]
   },
   {
    "unit": "fa3_fwd._flash_attn_forward: _flash_attn_forward_default, key_4, value_9",
    "ms": 230.251862,
    "share": 0.007726549493067338,
    "count": 15840,
    "families": {
     "attention_core": 230.251862
    },
    "top_kernels": [
     {
      "name": "void cutlass::device_kernel<flash::enable_sm90_or_later<flash::FlashAttnFwdSm90<flash::Col",
      "ms": 230.251862
     }
    ]
   },
   {
    "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": 194.04683600000004,
    "share": 0.006511619360225288,
    "count": 15840,
    "families": {
     "triton_fused": 194.04683600000004
    },
    "top_kernels": [
     {
      "name": "triton_red_fused__to_copy_add_mul_native_layer_norm_split_squeeze_view_9",
      "ms": 194.04683600000004
     }
    ]
   },
   {
    "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": 183.15405800000002,
    "share": 0.006146090987933579,
    "count": 15840,
    "families": {
     "triton_fused": 183.15405800000002
    },
    "top_kernels": [
     {
      "name": "triton_red_fused__to_copy_add_mul_native_layer_norm_split_squeeze_view_12",
      "ms": 183.15405800000002
     }
    ]
   },
   {
    "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": 177.21933899999996,
    "share": 0.005946939937937086,
    "count": 15840,
    "families": {
     "triton_fused": 177.21933899999996
    },
    "top_kernels": [
     {
      "name": "triton_red_fused__to_copy_add_mul_native_layer_norm_split_squeeze_1",
      "ms": 177.21933899999996
     }
    ]
   },
   {
    "unit": "triton_poi_fused_gelu_view_13: output_parallel_6, input_1",
    "ms": 176.454267,
    "share": 0.005921266457503908,
    "count": 15840,
    "families": {
     "triton_fused": 176.454267
    },
    "top_kernels": [
     {
      "name": "triton_poi_fused_gelu_view_13",
      "ms": 176.454267
     }
    ]
   },
   {
    "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": 146.35722800000002,
    "share": 0.004911301719723513,
    "count": 15840,
    "families": {
     "triton_fused": 146.35722800000002
    },
    "top_kernels": [
     {
      "name": "triton_poi_fused_all_to_all_single_clone_split_with_sizes_stack_transpose_view_4",
      "ms": 146.35722800000002
     }
    ]
   },
   {
    "unit": "triton_poi_fused__to_copy_add_mul_split_squeeze_view_14: float_1, modulation, chunk, output_parallel_5, hidden_states_2, output",
    "ms": 122.97947499999997,
    "share": 0.004126815705051441,
    "count": 15840,
    "families": {
     "triton_fused": 122.97947499999997
    },
    "top_kernels": [
     {
      "name": "triton_poi_fused__to_copy_add_mul_split_squeeze_view_14",
      "ms": 122.97947499999997
     }
    ]
   },
   {
    "unit": "triton_poi_fused_clone_transpose_view_8: contiguous_8",
    "ms": 86.921572,
    "share": 0.0029168225708994097,
    "count": 15840,
    "families": {
     "triton_fused": 86.921572
    },
    "top_kernels": [
     {
      "name": "triton_poi_fused_clone_transpose_view_8",
      "ms": 86.921572
     }
    ]
   },
   {
    "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": 86.73247199999999,
    "share": 0.002910476952251865,
    "count": 15840,
    "families": {
     "triton_fused": 86.73247199999999
    },
    "top_kernels": [
     {
      "name": "triton_poi_fused__to_copy_all_to_all_single_clone_div_mul_sqrt_transpose_view_11",
      "ms": 86.73247199999999
     }
    ]
   },
   {
    "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": 85.09991100000005,
    "share": 0.0028556931895609423,
    "count": 15840,
    "families": {
     "triton_fused": 85.09991100000005
    },
    "top_kernels": [
     {
      "name": "triton_red_fused__to_copy_div_mul_pow_split_with_sizes_sqrt_sum_view_3",
      "ms": 85.09991100000005
     }
    ]
   },
   {
    "unit": "triton_poi_fused_silu_view_6: output_parallel_2, input_parallel_1",
    "ms": 81.621832,
    "share": 0.002738979477450773,
    "count": 15840,
    "families": {
     "triton_fused": 81.621832
    },
    "top_kernels": [
     {
      "name": "triton_poi_fused_silu_view_6",
      "ms": 81.621832
     }
    ]
   },
   {
    "unit": "rotary_kernel_0: output_parallel, split, value_float, truediv, rms, truediv_1, float_5,",
    "ms": 77.21610600000001,
    "share": 0.0025911367642748265,
    "count": 15840,
    "families": {
     "triton_fused": 77.21610600000001
    },
    "top_kernels": [
     {
      "name": "rotary_kernel",
      "ms": 77.21610600000001
     }
    ]
   },
   {
    "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": 77.16982200000001,
    "share": 0.0025895836145472593,
    "count": 15840,
    "families": {
     "triton_fused": 77.16982200000001
    },
    "top_kernels": [
     {
      "name": "triton_red_fused__to_copy_div_mul_pow_split_with_sizes_sqrt_sum_view_2",
      "ms": 77.16982200000001
     }
    ]
   },
   {
    "unit": "rotary_kernel_0: output_parallel, split, value_float_1, truediv_2, rms_1, truediv_3, fl",
    "ms": 76.932741,
    "share": 0.002581627900033359,
    "count": 15840,
    "families": {
     "triton_fused": 76.932741
    },
    "top_kernels": [
     {
      "name": "rotary_kernel",
      "ms": 76.932741
     }
    ]
   },
   {
    "unit": "non-graph: tfm_sub/head/_LingBotHead",
    "ms": 74.547019,
    "share": 0.0025015703537030736,
    "count": 3600,
    "families": {
     "elementwise_reduce": 71.22551099999998,
     "gemm": 3.3215079999999997
    },
    "top_kernels": [
     {
      "name": "void at::native::elementwise_kernel<(int)128, (int)2, void at::native::gpu_kernel_impl_noc",
      "ms": 32.908922000000004
     },
     {
      "name": "void at::native::unrolled_elementwise_kernel<at::native::direct_copy_kernel_cuda(at::Tenso",
      "ms": 15.465117000000001
     },
     {
      "name": "void at::native::<unnamed>::vectorized_layer_norm_kernel<float, float, (bool)0>(int, T2, c",
      "ms": 14.985373000000003
     },
     {
      "name": "void at::native::vectorized_elementwise_kernel<(int)8, at::native::bfloat16_copy_kernel_cu",
      "ms": 6.830163000000001
     }
    ]
   },
   {
    "unit": "triton_poi_fused_add_view_7: output_parallel_3, camera_features",
    "ms": 73.63035999999998,
    "share": 0.0024708100763691783,
    "count": 15840,
    "families": {
     "triton_fused": 73.63035999999998
    },
    "top_kernels": [
     {
      "name": "triton_poi_fused_add_view_7",
      "ms": 73.63035999999998
     }
    ]
   },
   {
    "unit": "triton_poi_fused_clone_transpose_view_8: contiguous_12",
    "ms": 72.840447,
    "share": 0.002444303007819534,
    "count": 15840,
    "families": {
     "triton_fused": 72.840447
    },
    "top_kernels": [
     {
      "name": "triton_poi_fused_clone_transpose_view_8",
      "ms": 72.840447
     }
    ]
   },
   {
    "unit": "non-graph: tfm_sub/sp_prepare/_LingBotSPPrepare",
    "ms": 63.065367,
    "share": 0.002116281167897594,
    "count": 400,
    "families": {
     "elementwise_reduce": 63.065367
    },
    "top_kernels": [
     {
      "name": "void at::native::elementwise_kernel<(int)128, (int)4, void at::native::gpu_kernel_impl_noc",
      "ms": 63.065367
     }
    ]
   },
   {
    "unit": "non-graph: tfm_sub/c2ws_hidden_states_layer1/ColumnParallelLinear",
    "ms": 60.276903000000004,
    "share": 0.002022708829682859,
    "count": 400,
    "families": {
     "gemm": 60.276903000000004
    },
    "top_kernels": [
     {
      "name": "nvjet_sm90_tst_256x144_64x4_2x1_v_bz_coopA_bias_TNT",
      "ms": 60.276903000000004
     }
    ]
   },
   {
    "unit": "non-graph: tfm_sub/c2ws_hidden_states_layer2/RowParallelLinear",
    "ms": 60.212307,
    "share": 0.002020541185144748,
    "count": 400,
    "families": {
     "gemm": 60.212307
    },
    "top_kernels": [
     {
      "name": "nvjet_sm90_tst_256x144_64x4_2x1_v_bz_coopA_bias_TNT",
      "ms": 60.212307
     }
    ]
   },
   {
    "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": 55.18050699999999,
    "share": 0.0018516893400325623,
    "count": 15840,
    "families": {
     "triton_fused": 55.18050699999999
    },
    "top_kernels": [
     {
      "name": "triton_poi_fused_all_to_all_single_clone_transpose_unsqueeze_view_5",
      "ms": 55.18050699999999
     }
    ]
   },
   {
    "unit": "triton_poi_fused_all_to_all_single_clone_transpose_unsqueeze_view_5: reshape_10, transpose_8, transpose_9, contiguous_11, tensor_3",
    "ms": 50.839527,
    "share": 0.0017060193049367532,
    "count": 15840,
    "families": {
     "triton_fused": 50.839527
    },
    "top_kernels": [
     {
      "name": "triton_poi_fused_all_to_all_single_clone_transpose_unsqueeze_view_5",
      "ms": 50.839527
     }
    ]
   },
   {
    "unit": "triton_red_fused__to_copy_pow_sum_view_10: output_parallel_4, value_float_2, pow_3, sum_of_squares_2",
    "ms": 40.821319,
    "share": 0.0013698388316433685,
    "count": 15840,
    "families": {
     "triton_fused": 40.821319
    },
    "top_kernels": [
     {
      "name": "triton_red_fused__to_copy_pow_sum_view_10",
      "ms": 40.821319
     }
    ]
   },
   {
    "unit": "non-graph: ncclAllGather",
    "ms": 29.233268,
    "share": 0.0009809792202510033,
    "count": 400,
    "families": {
     "nccl": 29.233268
    },
    "top_kernels": [
     {
      "name": "ncclDevKernel_AllGather_RING_LL(ncclDevKernelArgsStorage<(unsigned long)4096>)",
      "ms": 29.233268
     }
    ]
   },
   {
    "unit": "non-graph: tfm_sub/patch_embedding_wancamctrl/_LingBotCameraPatchEmbedding",
    "ms": 26.315661000000002,
    "share": 0.0008830732372504416,
    "count": 800,
    "families": {
     "gemm": 19.990693,
     "elementwise_reduce": 6.324967999999999
    },
    "top_kernels": [
     {
      "name": "nvjet_sm90_tst_256x144_64x4_2x1_v_bz_coopA_bias_TNT",
      "ms": 19.990693
     },
     {
      "name": "void at::native::elementwise_kernel<(int)128, (int)4, void at::native::gpu_kernel_impl_noc",
      "ms": 6.324967999999999
     }
    ]
   },
   {
    "unit": "non-graph: tfm_sub/_timestep_embeddings",
    "ms": 24.737824,
    "share": 0.0008301258449184182,
    "count": 6400,
    "families": {
     "gemm": 20.284921999999995,
     "elementwise_reduce": 4.070372999999997,
     "gather_index_cat": 0.38252900000000006
    },
    "top_kernels": [
     {
      "name": "nvjet_sm90_tst_64x8_64x16_2x1_v_bz_bias_TNT",
      "ms": 15.466581
     },
     {
      "name": "nvjet_sm90_tst_64x8_64x16_4x1_v_bz_bias_TNT",
      "ms": 4.818341000000001
     },
     {
      "name": "void at::native::elementwise_kernel<(int)128, (int)4, void at::native::gpu_kernel_impl<at:",
      "ms": 1.020834
     },
     {
      "name": "void at::native::vectorized_elementwise_kernel<(int)8, at::native::<unnamed>::silu_kernel(",
      "ms": 0.6073919999999999
     }
    ]
   },
   {
    "unit": "non-graph: tfm_sub/patch_embedding/Conv3dLayer",
    "ms": 5.998921999999999,
    "share": 0.0002013055066544934,
    "count": 800,
    "families": {
     "gemm": 5.215209000000001,
     "elementwise_reduce": 0.7837129999999999
    },
    "top_kernels": [
     {
      "name": "nvjet_sm90_tst_256x144_64x4_2x1_v_bz_coopA_bias_TNT",
      "ms": 5.215209000000001
     },
     {
      "name": "void at::native::elementwise_kernel<(int)128, (int)4, void at::native::gpu_kernel_impl_noc",
      "ms": 0.7837129999999999
     }
    ]
   },
   {
    "unit": "non-graph: gen/f=*",
    "ms": 3.615429,
    "share": 0.00012132275875871505,
    "count": 3200,
    "families": {
     "elementwise_reduce": 3.615429
    },
    "top_kernels": [
     {
      "name": "void at::native::<unnamed>::distribution_elementwise_grid_stride_kernel<float, (int)4, voi",
      "ms": 0.854016
     },
     {
      "name": "void at::native::elementwise_kernel<(int)128, (int)4, void at::native::gpu_kernel_impl_noc",
      "ms": 0.721921
     },
     {
      "name": "void at::native::vectorized_elementwise_kernel<(int)4, at::native::AUnaryFunctor<float, fl",
      "ms": 0.569184
     },
     {
      "name": "void at::native::vectorized_elementwise_kernel<(int)4, at::native::CUDAFunctor_add<float>,",
      "ms": 0.529379
     }
    ]
   },
   {
    "unit": "non-graph: tfm_sub/_rotary_embedding",
    "ms": 2.2199699999999996,
    "share": 7.449541527757415e-05,
    "count": 1600,
    "families": {
     "gather_index_cat": 1.6786580000000002,
     "elementwise_reduce": 0.541312
    },
    "top_kernels": [
     {
      "name": "void at::native::<unnamed>::CatArrayBatchedCopy<at::native::<unnamed>::OpaqueType<(unsigne",
      "ms": 1.6786580000000002
     },
     {
      "name": "void at::native::vectorized_elementwise_kernel<(int)8, at::native::bfloat16_copy_kernel_cu",
      "ms": 0.541312
     }
    ]
   },
   {
    "unit": "non-graph: tfm/f=15/c=0/s=*",
    "ms": 0.951778,
    "share": 3.1938763749987155e-05,
    "count": 64,
    "families": {
     "elementwise_reduce": 0.951778
    },
    "top_kernels": [
     {
      "name": "void at::native::vectorized_elementwise_kernel<(int)8, at::native::CUDAFunctor_add<c10::BF",
      "ms": 0.523201
     },
     {
      "name": "void at::native::vectorized_elementwise_kernel<(int)8, at::native::<unnamed>::silu_kernel(",
      "ms": 0.428577
     }
    ]
   },
   {
    "unit": "non-graph: tfm/f=12/c=0/s=*",
    "ms": 0.95021,
    "share": 3.18861464573412e-05,
    "count": 64,
    "families": {
     "elementwise_reduce": 0.95021
    },
    "top_kernels": [
     {
      "name": "void at::native::vectorized_elementwise_kernel<(int)8, at::native::CUDAFunctor_add<c10::BF",
      "ms": 0.528
     },
     {
      "name": "void at::native::vectorized_elementwise_kernel<(int)8, at::native::<unnamed>::silu_kernel(",
      "ms": 0.42221
     }
    ]
   },
   {
    "unit": "non-graph: tfm/f=9/c=0/s=*",
    "ms": 0.949478,
    "share": 3.1861582772253935e-05,
    "count": 64,
    "families": {
     "elementwise_reduce": 0.949478
    },
    "top_kernels": [
     {
      "name": "void at::native::vectorized_elementwise_kernel<(int)8, at::native::CUDAFunctor_add<c10::BF",
      "ms": 0.523267
     },
     {
      "name": "void at::native::vectorized_elementwise_kernel<(int)8, at::native::<unnamed>::silu_kernel(",
      "ms": 0.426211
     }
    ]
   },
   {
    "unit": "non-graph: tfm/f=24/c=0/s=*",
    "ms": 0.949441,
    "share": 3.186034116522083e-05,
    "count": 64,
    "families": {
     "elementwise_reduce": 0.949441
    },
    "top_kernels": [
     {
      "name": "void at::native::vectorized_elementwise_kernel<(int)8, at::native::CUDAFunctor_add<c10::BF",
      "ms": 0.523329
     },
     {
      "name": "void at::native::vectorized_elementwise_kernel<(int)8, at::native::<unnamed>::silu_kernel(",
      "ms": 0.426112
     }
    ]
   },
   {
    "unit": "non-graph: tfm/f=21/c=0/s=*",
    "ms": 0.94909,
    "share": 3.1848562676879805e-05,
    "count": 64,
    "families": {
     "elementwise_reduce": 0.94909
    },
    "top_kernels": [
     {
      "name": "void at::native::vectorized_elementwise_kernel<(int)8, at::native::CUDAFunctor_add<c10::BF",
      "ms": 0.523873
     },
     {
      "name": "void at::native::vectorized_elementwise_kernel<(int)8, at::native::<unnamed>::silu_kernel(",
      "ms": 0.425217
     }
    ]
   },
   {
    "unit": "non-graph: tfm/f=27/c=0/s=*",
    "ms": 0.948418,
    "share": 3.182601240860297e-05,
    "count": 64,
    "families": {
     "elementwise_reduce": 0.948418
    },
    "top_kernels": [
     {
      "name": "void at::native::vectorized_elementwise_kernel<(int)8, at::native::CUDAFunctor_add<c10::BF",
      "ms": 0.521889
     },
     {
      "name": "void at::native::vectorized_elementwise_kernel<(int)8, at::native::<unnamed>::silu_kernel(",
      "ms": 0.426529
     }
    ]
   },
   {
    "unit": "non-graph: tfm/f=18/c=0/s=*",
    "ms": 0.945697,
    "share": 3.1734703956249885e-05,
    "count": 64,
    "families": {
     "elementwise_reduce": 0.945697
    },
    "top_kernels": [
     {
      "name": "void at::native::vectorized_elementwise_kernel<(int)8, at::native::CUDAFunctor_add<c10::BF",
      "ms": 0.520385
     },
     {
      "name": "void at::native::vectorized_elementwise_kernel<(int)8, at::native::<unnamed>::silu_kernel(",
      "ms": 0.425312
     }
    ]
   },
   {
    "unit": "non-graph: tfm/f=6/c=0/s=*",
    "ms": 0.945441,
    "share": 3.1726113377858704e-05,
    "count": 64,
    "families": {
     "elementwise_reduce": 0.945441
    },
    "top_kernels": [
     {
      "name": "void at::native::vectorized_elementwise_kernel<(int)8, at::native::CUDAFunctor_add<c10::BF",
      "ms": 0.524129
     },
     {
      "name": "void at::native::vectorized_elementwise_kernel<(int)8, at::native::<unnamed>::silu_kernel(",
      "ms": 0.421312
     }
    ]
   },
   {
    "unit": "non-graph: tfm/f=3/c=0/s=*",
    "ms": 0.944608,
    "share": 3.1698160441140544e-05,
    "count": 64,
    "families": {
     "elementwise_reduce": 0.944608
    },
    "top_kernels": [
     {
      "name": "void at::native::vectorized_elementwise_kernel<(int)8, at::native::CUDAFunctor_add<c10::BF",
      "ms": 0.523104
     },
     {
      "name": "void at::native::vectorized_elementwise_kernel<(int)8, at::native::<unnamed>::silu_kernel(",
      "ms": 0.421504
     }
    ]
   },
   {
    "unit": "non-graph: tfm/f=0/c=0/s=*",
    "ms": 0.9358090000000001,
    "share": 3.1402892865890715e-05,
    "count": 64,
    "families": {
     "elementwise_reduce": 0.9358090000000001
    },
    "top_kernels": [
     {
      "name": "void at::native::vectorized_elementwise_kernel<(int)8, at::native::CUDAFunctor_add<c10::BF",
      "ms": 0.519777
     },
     {
      "name": "void at::native::vectorized_elementwise_kernel<(int)8, at::native::<unnamed>::silu_kernel(",
      "ms": 0.416032
     }
    ]
   },
   {
    "unit": "non-graph: tfm_sub/_unpatchify",
    "ms": 0.6999070000000001,
    "share": 2.348674199231571e-05,
    "count": 400,
    "families": {
     "elementwise_reduce": 0.6999070000000001
    },
    "top_kernels": [
     {
      "name": "void at::native::elementwise_kernel<(int)128, (int)4, void at::native::gpu_kernel_impl_noc",
      "ms": 0.6999070000000001
     }
    ]
   },
   {
    "unit": "non-graph: tfm_sub/text_embedding/Sequential",
    "ms": 0.296641,
    "share": 9.954366267722032e-06,
    "count": 12,
    "families": {
     "gemm": 0.281153,
     "elementwise_reduce": 0.015488
    },
    "top_kernels": [
     {
      "name": "nvjet_sm90_tst_168x128_64x5_1x2_h_bz_coopA_bias_TNN",
      "ms": 0.281153
     },
     {
      "name": "void at::native::vectorized_elementwise_kernel<(int)8, at::native::GeluCUDAKernelImpl(at::",
      "ms": 0.015488
     }
    ]
   },
   {
    "unit": "non-graph: tfm/f=24/c=1/s=None",
    "ms": 0.23852800000000002,
    "share": 8.004271415978239e-06,
    "count": 16,
    "families": {
     "elementwise_reduce": 0.23852800000000002
    },
    "top_kernels": [
     {
      "name": "void at::native::vectorized_elementwise_kernel<(int)8, at::native::CUDAFunctor_add<c10::BF",
      "ms": 0.131136
     },
     {
      "name": "void at::native::vectorized_elementwise_kernel<(int)8, at::native::<unnamed>::silu_kernel(",
      "ms": 0.107392
     }
    ]
   },
   {
    "unit": "non-graph: tfm/f=15/c=1/s=None",
    "ms": 0.238528,
    "share": 8.004271415978239e-06,
    "count": 16,
    "families": {
     "elementwise_reduce": 0.238528
    },
    "top_kernels": [
     {
      "name": "void at::native::vectorized_elementwise_kernel<(int)8, at::native::CUDAFunctor_add<c10::BF",
      "ms": 0.131872
     },
     {
      "name": "void at::native::vectorized_elementwise_kernel<(int)8, at::native::<unnamed>::silu_kernel(",
      "ms": 0.106656
     }
    ]
   },
   {
    "unit": "non-graph: tfm/f=18/c=1/s=None",
    "ms": 0.238368,
    "share": 7.998902304483753e-06,
    "count": 16,
    "families": {
     "elementwise_reduce": 0.238368
    },
    "top_kernels": [
     {
      "name": "void at::native::vectorized_elementwise_kernel<(int)8, at::native::CUDAFunctor_add<c10::BF",
      "ms": 0.130688
     },
     {
      "name": "void at::native::vectorized_elementwise_kernel<(int)8, at::native::<unnamed>::silu_kernel(",
      "ms": 0.10768
     }
    ]
   },
   {
    "unit": "non-graph: tfm/f=6/c=1/s=None",
    "ms": 0.238176,
    "share": 7.992459370690372e-06,
    "count": 16,
    "families": {
     "elementwise_reduce": 0.238176
    },
    "top_kernels": [
     {
      "name": "void at::native::vectorized_elementwise_kernel<(int)8, at::native::CUDAFunctor_add<c10::BF",
      "ms": 0.13248
     },
     {
      "name": "void at::native::vectorized_elementwise_kernel<(int)8, at::native::<unnamed>::silu_kernel(",
      "ms": 0.105696
     }
    ]
   },
   {
    "unit": "non-graph: tfm/f=12/c=1/s=None",
    "ms": 0.23792000000000002,
    "share": 7.983868792299197e-06,
    "count": 16,
    "families": {
     "elementwise_reduce": 0.23792000000000002
    },
    "top_kernels": [
     {
      "name": "void at::native::vectorized_elementwise_kernel<(int)8, at::native::CUDAFunctor_add<c10::BF",
      "ms": 0.129504
     },
     {
      "name": "void at::native::vectorized_elementwise_kernel<(int)8, at::native::<unnamed>::silu_kernel(",
      "ms": 0.108416
     }
    ]
   },
   {
    "unit": "non-graph: tfm/f=21/c=1/s=None",
    "ms": 0.237664,
    "share": 7.97527821390802e-06,
    "count": 16,
    "families": {
     "elementwise_reduce": 0.237664
    },
    "top_kernels": [
     {
      "name": "void at::native::vectorized_elementwise_kernel<(int)8, at::native::CUDAFunctor_add<c10::BF",
      "ms": 0.131136
     },
     {
      "name": "void at::native::vectorized_elementwise_kernel<(int)8, at::native::<unnamed>::silu_kernel(",
      "ms": 0.106528
     }
    ]
   },
   {
    "unit": "non-graph: tfm/f=27/c=1/s=None",
    "ms": 0.237152,
    "share": 7.958097057125668e-06,
    "count": 16,
    "families": {
     "elementwise_reduce": 0.237152
    },
    "top_kernels": [
     {
      "name": "void at::native::vectorized_elementwise_kernel<(int)8, at::native::CUDAFunctor_add<c10::BF",
      "ms": 0.13008
     },
     {
      "name": "void at::native::vectorized_elementwise_kernel<(int)8, at::native::<unnamed>::silu_kernel(",
      "ms": 0.107072
     }
    ]
   },
   {
    "unit": "non-graph: tfm/f=3/c=1/s=None",
    "ms": 0.23683200000000001,
    "share": 7.947358834136697e-06,
    "count": 16,
    "families": {
     "elementwise_reduce": 0.23683200000000001
    },
    "top_kernels": [
     {
      "name": "void at::native::vectorized_elementwise_kernel<(int)8, at::native::CUDAFunctor_add<c10::BF",
      "ms": 0.130592
     },
     {
      "name": "void at::native::vectorized_elementwise_kernel<(int)8, at::native::<unnamed>::silu_kernel(",
      "ms": 0.10624
     }
    ]
   },
   {
    "unit": "non-graph: tfm/f=0/c=1/s=None",
    "ms": 0.23641600000000002,
    "share": 7.933399144251037e-06,
    "count": 16,
    "families": {
     "elementwise_reduce": 0.23641600000000002
    },
    "top_kernels": [
     {
      "name": "void at::native::vectorized_elementwise_kernel<(int)8, at::native::CUDAFunctor_add<c10::BF",
      "ms": 0.129664
     },
     {
      "name": "void at::native::vectorized_elementwise_kernel<(int)8, at::native::<unnamed>::silu_kernel(",
      "ms": 0.106752
     }
    ]
   },
   {
    "unit": "non-graph: tfm/f=9/c=1/s=None",
    "ms": 0.23625599999999997,
    "share": 7.928030032756551e-06,
    "count": 16,
    "families": {
     "elementwise_reduce": 0.23625599999999997
    },
    "top_kernels": [
     {
      "name": "void at::native::vectorized_elementwise_kernel<(int)8, at::native::CUDAFunctor_add<c10::BF",
      "ms": 0.12976
     },
     {
      "name": "void at::native::vectorized_elementwise_kernel<(int)8, at::native::<unnamed>::silu_kernel(",
      "ms": 0.106496
     }
    ]
   }
  ],
  "by_family": {
   "gemm": 10825.804202999996,
   "attention_core": 6220.862985,
   "elementwise_reduce": 4014.0300199999947,
   "conv": 3063.4081859999997,
   "nccl": 2713.8258849999984,
   "triton_fused": 1865.2178189999984,
   "gather_index_cat": 1074.4978139999998,
   "attention_prep": 17.840091,
   "other": 4.601925
  }
 },
 "vae_measured": {
  "gpu_ms": 5520.008762,
  "groups": 2644,
  "by_unit": [
   {
    "unit": "vae_frame/*",
    "ms": 380.883649,
    "share": 0.06900055152484919,
    "count": 14796,
    "families": {
     "attention_core": 230.010368,
     "elementwise_reduce": 150.87328099999993
    },
    "top_kernels": [
     {
      "name": "fmha_cutlassF_f32_aligned_32x128_gmem_sm80(PyTorchMemEffAttention::AttentionKernel<float, ",
      "ms": 230.010368
     },
     {
      "name": "void at::native::elementwise_kernel<(int)128, (int)2, void at::native::gpu_kernel_impl_noc",
      "ms": 105.53319500000003
     },
     {
      "name": "void at::native::vectorized_elementwise_kernel<(int)4, at::native::CUDAFunctor_add<float>,",
      "ms": 45.334582000000005
     },
     {
      "name": "void at::native::vectorized_elementwise_kernel<(int)4, at::native::FillFunctor<float>, std",
      "ms": 0.005504
     }
    ]
   },
   {
    "unit": "vae/up_blocks.3.resnets.0.conv1/WanDistCausalConv3d",
    "ms": 278.1185479999999,
    "share": 0.05038371495251621,
    "count": 3552,
    "families": {
     "conv": 183.13709399999996,
     "elementwise_reduce": 85.71977199999998,
     "nccl": 9.261682
    },
    "top_kernels": [
     {
      "name": "sm90_xmma_fprop_implicit_gemm_f32f32_tf32f32_f32_nhwckrsc_nhwc_tilesize256x32x32_warpgroup",
      "ms": 155.919773
     },
     {
      "name": "void at::native::elementwise_kernel<(int)128, (int)2, void at::native::gpu_kernel_impl_noc",
      "ms": 79.73415500000002
     },
     {
      "name": "void cudnn::engines_precompiled::nchwToNhwcKernel<float, float, float, (bool)0, (bool)1, (",
      "ms": 17.657370000000004
     },
     {
      "name": "void cudnn::engines_precompiled::nhwcToNchwKernel<float, float, float, (bool)1, (bool)0, (",
      "ms": 9.559951
     }
    ]
   },
   {
    "unit": "vae/up_blocks.3.resnets.0.conv2/WanDistCausalConv3d",
    "ms": 275.33221100000014,
    "share": 0.049878944558095634,
    "count": 3552,
    "families": {
     "conv": 183.11924899999994,
     "elementwise_reduce": 84.820051,
     "nccl": 7.392911
    },
    "top_kernels": [
     {
      "name": "sm90_xmma_fprop_implicit_gemm_f32f32_tf32f32_f32_nhwckrsc_nhwc_tilesize256x32x32_warpgroup",
      "ms": 155.96770099999998
     },
     {
      "name": "void at::native::elementwise_kernel<(int)128, (int)2, void at::native::gpu_kernel_impl_noc",
      "ms": 78.84010300000003
     },
     {
      "name": "void cudnn::engines_precompiled::nchwToNhwcKernel<float, float, float, (bool)0, (bool)1, (",
      "ms": 17.614282
     },
     {
      "name": "void cudnn::engines_precompiled::nhwcToNchwKernel<float, float, float, (bool)1, (bool)0, (",
      "ms": 9.537265999999999
     }
    ]
   },
   {
    "unit": "vae/up_blocks.3.resnets.1.conv1/WanDistCausalConv3d",
    "ms": 275.0318940000002,
    "share": 0.0498245393908308,
    "count": 3552,
    "families": {
     "conv": 183.01705999999993,
     "elementwise_reduce": 84.76902899999999,
     "nccl": 7.245805
    },
    "top_kernels": [
     {
      "name": "sm90_xmma_fprop_implicit_gemm_f32f32_tf32f32_f32_nhwckrsc_nhwc_tilesize256x32x32_warpgroup",
      "ms": 155.794423
     },
     {
      "name": "void at::native::elementwise_kernel<(int)128, (int)2, void at::native::gpu_kernel_impl_noc",
      "ms": 78.78008899999999
     },
     {
      "name": "void cudnn::engines_precompiled::nchwToNhwcKernel<float, float, float, (bool)0, (bool)1, (",
      "ms": 17.665248999999996
     },
     {
      "name": "void cudnn::engines_precompiled::nhwcToNchwKernel<float, float, float, (bool)1, (bool)0, (",
      "ms": 9.557388
     }
    ]
   },
   {
    "unit": "vae/up_blocks.3.resnets.1.conv2/WanDistCausalConv3d",
    "ms": 274.54782699999987,
    "share": 0.0497368462329263,
    "count": 3552,
    "families": {
     "conv": 183.09568900000005,
     "elementwise_reduce": 84.75503800000001,
     "nccl": 6.6971
    },
    "top_kernels": [
     {
      "name": "sm90_xmma_fprop_implicit_gemm_f32f32_tf32f32_f32_nhwckrsc_nhwc_tilesize256x32x32_warpgroup",
      "ms": 155.954483
     },
     {
      "name": "void at::native::elementwise_kernel<(int)128, (int)2, void at::native::gpu_kernel_impl_noc",
      "ms": 78.776781
     },
     {
      "name": "void cudnn::engines_precompiled::nchwToNhwcKernel<float, float, float, (bool)0, (bool)1, (",
      "ms": 17.594312
     },
     {
      "name": "void cudnn::engines_precompiled::nhwcToNchwKernel<float, float, float, (bool)1, (bool)0, (",
      "ms": 9.546894
     }
    ]
   },
   {
    "unit": "vae/up_blocks.3.resnets.2.conv2/WanDistCausalConv3d",
    "ms": 274.3119270000001,
    "share": 0.049694110793514726,
    "count": 3552,
    "families": {
     "conv": 182.994067,
     "elementwise_reduce": 84.68651700000001,
     "nccl": 6.631343
    },
    "top_kernels": [
     {
      "name": "sm90_xmma_fprop_implicit_gemm_f32f32_tf32f32_f32_nhwckrsc_nhwc_tilesize256x32x32_warpgroup",
      "ms": 155.83628199999998
     },
     {
      "name": "void at::native::elementwise_kernel<(int)128, (int)2, void at::native::gpu_kernel_impl_noc",
      "ms": 78.71463400000002
     },
     {
      "name": "void cudnn::engines_precompiled::nchwToNhwcKernel<float, float, float, (bool)0, (bool)1, (",
      "ms": 17.612996
     },
     {
      "name": "void cudnn::engines_precompiled::nhwcToNchwKernel<float, float, float, (bool)1, (bool)0, (",
      "ms": 9.544789
     }
    ]
   },
   {
    "unit": "vae/up_blocks.3.resnets.2.conv1/WanDistCausalConv3d",
    "ms": 274.2058099999999,
    "share": 0.049674886729826506,
    "count": 3552,
    "families": {
     "conv": 182.823625,
     "elementwise_reduce": 84.74038200000004,
     "nccl": 6.641803
    },
    "top_kernels": [
     {
      "name": "sm90_xmma_fprop_implicit_gemm_f32f32_tf32f32_f32_nhwckrsc_nhwc_tilesize256x32x32_warpgroup",
      "ms": 155.645496
     },
     {
      "name": "void at::native::elementwise_kernel<(int)128, (int)2, void at::native::gpu_kernel_impl_noc",
      "ms": 78.76718900000002
     },
     {
      "name": "void cudnn::engines_precompiled::nchwToNhwcKernel<float, float, float, (bool)0, (bool)1, (",
      "ms": 17.624033999999998
     },
     {
      "name": "void cudnn::engines_precompiled::nhwcToNchwKernel<float, float, float, (bool)1, (bool)0, (",
      "ms": 9.554095
     }
    ]
   },
   {
    "unit": "vae/up_blocks.2.upsamplers.0.resample.1/WanDistConv2d",
    "ms": 208.217369,
    "share": 0.03772047798789345,
    "count": 2612,
    "families": {
     "conv": 109.76874599999998,
     "elementwise_reduce": 51.04821700000002,
     "gather_index_cat": 31.820246,
     "nccl": 15.58016
    },
    "top_kernels": [
     {
      "name": "sm90_xmma_fprop_implicit_gemm_f32f32_tf32f32_f32_nhwckrsc_nhwc_tilesize64x192x32_warpgroup",
      "ms": 77.44813199999999
     },
     {
      "name": "void at::native::elementwise_kernel<(int)128, (int)2, void at::native::gpu_kernel_impl_noc",
      "ms": 43.201257000000005
     },
     {
      "name": "void at::native::<unnamed>::CatArrayBatchedCopy_alignedK_contig<at::native::<unnamed>::Opa",
      "ms": 31.820246
     },
     {
      "name": "void cudnn::engines_precompiled::nchwToNhwcKernel<float, float, float, (bool)0, (bool)1, (",
      "ms": 22.478417
     }
    ]
   },
   {
    "unit": "vae/up_blocks.2.resnets.2.conv2/WanDistCausalConv3d",
    "ms": 166.4903139999999,
    "share": 0.03016124089261,
    "count": 3552,
    "families": {
     "conv": 103.35132400000002,
     "elementwise_reduce": 45.956232,
     "nccl": 17.182758
    },
    "top_kernels": [
     {
      "name": "sm90_xmma_fprop_implicit_gemm_f32f32_tf32f32_f32_nhwckrsc_nhwc_tilesize128x128x32_warpgrou",
      "ms": 89.45622699999998
     },
     {
      "name": "void at::native::elementwise_kernel<(int)128, (int)2, void at::native::gpu_kernel_impl_noc",
      "ms": 42.843652000000006
     },
     {
      "name": "ncclDevKernel_SendRecv(ncclDevKernelArgsStorage<(unsigned long)4096>)",
      "ms": 17.182758
     },
     {
      "name": "void cudnn::engines_precompiled::nchwToNhwcKernel<float, float, float, (bool)0, (bool)1, (",
      "ms": 9.286511
     }
    ]
   },
   {
    "unit": "vae/up_blocks.2.resnets.1.conv2/WanDistCausalConv3d",
    "ms": 165.73648599999993,
    "share": 0.03002467806590049,
    "count": 3552,
    "families": {
     "conv": 103.293467,
     "elementwise_reduce": 45.956584,
     "nccl": 16.486435
    },
    "top_kernels": [
     {
      "name": "sm90_xmma_fprop_implicit_gemm_f32f32_tf32f32_f32_nhwckrsc_nhwc_tilesize128x128x32_warpgrou",
      "ms": 89.39139599999999
     },
     {
      "name": "void at::native::elementwise_kernel<(int)128, (int)2, void at::native::gpu_kernel_impl_noc",
      "ms": 42.846879
     },
     {
      "name": "ncclDevKernel_SendRecv(ncclDevKernelArgsStorage<(unsigned long)4096>)",
      "ms": 16.486435
     },
     {
      "name": "void cudnn::engines_precompiled::nchwToNhwcKernel<float, float, float, (bool)0, (bool)1, (",
      "ms": 9.29576
     }
    ]
   },
   {
    "unit": "vae/up_blocks.2.resnets.2.conv1/WanDistCausalConv3d",
    "ms": 164.99242299999995,
    "share": 0.029889884258122112,
    "count": 3552,
    "families": {
     "conv": 103.108446,
     "elementwise_reduce": 45.948076,
     "nccl": 15.935901000000001
    },
    "top_kernels": [
     {
      "name": "sm90_xmma_fprop_implicit_gemm_f32f32_tf32f32_f32_nhwckrsc_nhwc_tilesize128x128x32_warpgrou",
      "ms": 89.19420499999998
     },
     {
      "name": "void at::native::elementwise_kernel<(int)128, (int)2, void at::native::gpu_kernel_impl_noc",
      "ms": 42.842853999999996
     },
     {
      "name": "ncclDevKernel_SendRecv(ncclDevKernelArgsStorage<(unsigned long)4096>)",
      "ms": 15.935901000000001
     },
     {
      "name": "void cudnn::engines_precompiled::nchwToNhwcKernel<float, float, float, (bool)0, (bool)1, (",
      "ms": 9.315131999999998
     }
    ]
   },
   {
    "unit": "vae/up_blocks.2.resnets.1.conv1/WanDistCausalConv3d",
    "ms": 164.90634500000002,
    "share": 0.029874290442294775,
    "count": 3552,
    "families": {
     "conv": 103.63531599999999,
     "elementwise_reduce": 45.991643999999994,
     "nccl": 15.279385
    },
    "top_kernels": [
     {
      "name": "sm90_xmma_fprop_implicit_gemm_f32f32_tf32f32_f32_nhwckrsc_nhwc_tilesize128x128x32_warpgrou",
      "ms": 89.713471
     },
     {
      "name": "void at::native::elementwise_kernel<(int)128, (int)2, void at::native::gpu_kernel_impl_noc",
      "ms": 42.88776599999999
     },
     {
      "name": "ncclDevKernel_SendRecv(ncclDevKernelArgsStorage<(unsigned long)4096>)",
      "ms": 15.279385
     },
     {
      "name": "void cudnn::engines_precompiled::nchwToNhwcKernel<float, float, float, (bool)0, (bool)1, (",
      "ms": 9.315853
     }
    ]
   },
   {
    "unit": "vae/up_blocks.2.resnets.0.conv1/WanDistCausalConv3d",
    "ms": 164.49076200000005,
    "share": 0.029799003786436386,
    "count": 3552,
    "families": {
     "conv": 101.588344,
     "elementwise_reduce": 46.43057700000001,
     "nccl": 16.471841
    },
    "top_kernels": [
     {
      "name": "sm90_xmma_fprop_implicit_gemm_f32f32_tf32f32_f32_nhwckrsc_nhwc_tilesize128x128x32_warpgrou",
      "ms": 87.65010699999999
     },
     {
      "name": "void at::native::elementwise_kernel<(int)128, (int)2, void at::native::gpu_kernel_impl_noc",
      "ms": 43.32119599999999
     },
     {
      "name": "ncclDevKernel_SendRecv(ncclDevKernelArgsStorage<(unsigned long)4096>)",
      "ms": 16.471841
     },
     {
      "name": "void cudnn::engines_precompiled::nchwToNhwcKernel<float, float, float, (bool)0, (bool)1, (",
      "ms": 9.328114
     }
    ]
   },
   {
    "unit": "vae/conv_out/WanDistCausalConv3d",
    "ms": 162.087913,
    "share": 0.0293637057455091,
    "count": 3548,
    "families": {
     "conv": 83.57807399999999,
     "elementwise_reduce": 71.853417,
     "nccl": 6.656422000000001
    },
    "top_kernels": [
     {
      "name": "void at::native::elementwise_kernel<(int)128, (int)2, void at::native::gpu_kernel_impl_noc",
      "ms": 65.874266
     },
     {
      "name": "sm80_xmma_fprop_implicit_gemm_indexed_wo_smem_tf32f32_tf32f32_f32_nhwckrsc_nhwc_tilesize12",
      "ms": 63.685654
     },
     {
      "name": "void cudnn::engines_precompiled::nchwToNhwcKernel<float, float, float, (bool)0, (bool)1, (",
      "ms": 17.682816000000003
     },
     {
      "name": "ncclDevKernel_SendRecv(ncclDevKernelArgsStorage<(unsigned long)4096>)",
      "ms": 6.656422000000001
     }
    ]
   },
   {
    "unit": "vae/up_blocks.2.resnets.0.conv2/WanDistCausalConv3d",
    "ms": 160.883554,
    "share": 0.02914552511357046,
    "count": 3552,
    "families": {
     "conv": 100.67387899999997,
     "elementwise_reduce": 45.98452799999999,
     "nccl": 14.225147000000002
    },
    "top_kernels": [
     {
      "name": "sm90_xmma_fprop_implicit_gemm_f32f32_tf32f32_f32_nhwckrsc_nhwc_tilesize128x128x32_warpgrou",
      "ms": 86.75628499999999
     },
     {
      "name": "void at::native::elementwise_kernel<(int)128, (int)2, void at::native::gpu_kernel_impl_noc",
      "ms": 42.870377
     },
     {
      "name": "ncclDevKernel_SendRecv(ncclDevKernelArgsStorage<(unsigned long)4096>)",
      "ms": 14.225147000000002
     },
     {
      "name": "void cudnn::engines_precompiled::nchwToNhwcKernel<float, float, float, (bool)0, (bool)1, (",
      "ms": 9.303439
     }
    ]
   },
   {
    "unit": "vae/up_blocks.1.upsamplers.0.resample.1/WanDistConv2d",
    "ms": 135.64877700000002,
    "share": 0.024574014797551152,
    "count": 2856,
    "families": {
     "conv": 67.190222,
     "elementwise_reduce": 26.994279999999996,
     "nccl": 22.169038,
     "gather_index_cat": 19.295237
    },
    "top_kernels": [
     {
      "name": "sm90_xmma_fprop_implicit_gemm_f32f32_tf32f32_f32_nhwckrsc_nhwc_tilesize128x128x32_warpgrou",
      "ms": 50.904859
     },
     {
      "name": "void at::native::elementwise_kernel<(int)128, (int)2, void at::native::gpu_kernel_impl_noc",
      "ms": 22.938405
     },
     {
      "name": "ncclDevKernel_SendRecv(ncclDevKernelArgsStorage<(unsigned long)4096>)",
      "ms": 22.169038
     },
     {
      "name": "void at::native::<unnamed>::CatArrayBatchedCopy_alignedK_contig<at::native::<unnamed>::Opa",
      "ms": 19.295237
     }
    ]
   },
   {
    "unit": "vae/up_blocks.1.resnets.0.conv2/WanDistCausalConv3d",
    "ms": 109.782148,
    "share": 0.01988803872119651,
    "count": 3552,
    "families": {
     "conv": 56.90928699999999,
     "nccl": 36.024612,
     "elementwise_reduce": 16.848248999999985
    },
    "top_kernels": [
     {
      "name": "sm90_xmma_fprop_implicit_gemm_f32f32_tf32f32_f32_nhwckrsc_nhwc_tilesize128x128x32_warpgrou",
      "ms": 51.17679
     },
     {
      "name": "ncclDevKernel_SendRecv(ncclDevKernelArgsStorage<(unsigned long)4096>)",
      "ms": 36.024612
     },
     {
      "name": "void at::native::elementwise_kernel<(int)128, (int)2, void at::native::gpu_kernel_impl_noc",
      "ms": 15.685656
     },
     {
      "name": "void cudnn::engines_precompiled::nchwToNhwcKernel<float, float, float, (bool)0, (bool)1, (",
      "ms": 4.539022000000001
     }
    ]
   },
   {
    "unit": "vae/up_blocks.1.resnets.1.conv1/WanDistCausalConv3d",
    "ms": 102.01730900000004,
    "share": 0.018481367222148628,
    "count": 3552,
    "families": {
     "conv": 56.80816599999999,
     "nccl": 28.332981999999998,
     "elementwise_reduce": 16.876160999999993
    },
    "top_kernels": [
     {
      "name": "sm90_xmma_fprop_implicit_gemm_f32f32_tf32f32_f32_nhwckrsc_nhwc_tilesize128x128x32_warpgrou",
      "ms": 51.091514
     },
     {
      "name": "ncclDevKernel_SendRecv(ncclDevKernelArgsStorage<(unsigned long)4096>)",
      "ms": 28.332981999999998
     },
     {
      "name": "void at::native::elementwise_kernel<(int)128, (int)2, void at::native::gpu_kernel_impl_noc",
      "ms": 15.704320000000006
     },
     {
      "name": "void cudnn::engines_precompiled::nchwToNhwcKernel<float, float, float, (bool)0, (bool)1, (",
      "ms": 4.519785000000001
     }
    ]
   },
   {
    "unit": "vae/up_blocks.1.resnets.1.conv2/WanDistCausalConv3d",
    "ms": 97.48417500000001,
    "share": 0.017660148598148184,
    "count": 3552,
    "families": {
     "conv": 56.807656,
     "nccl": 23.824873999999998,
     "elementwise_reduce": 16.851644999999998
    },
    "top_kernels": [
     {
      "name": "sm90_xmma_fprop_implicit_gemm_f32f32_tf32f32_f32_nhwckrsc_nhwc_tilesize128x128x32_warpgrou",
      "ms": 51.073699
     },
     {
      "name": "ncclDevKernel_SendRecv(ncclDevKernelArgsStorage<(unsigned long)4096>)",
      "ms": 23.824873999999998
     },
     {
      "name": "void at::native::elementwise_kernel<(int)128, (int)2, void at::native::gpu_kernel_impl_noc",
      "ms": 15.681178999999997
     },
     {
      "name": "void cudnn::engines_precompiled::nchwToNhwcKernel<float, float, float, (bool)0, (bool)1, (",
      "ms": 4.535108999999999
     }
    ]
   },
   {
    "unit": "vae/up_blocks.1.resnets.2.conv1/WanDistCausalConv3d",
    "ms": 95.91130600000001,
    "share": 0.01737520901420627,
    "count": 3552,
    "families": {
     "conv": 56.579114000000004,
     "nccl": 22.470948,
     "elementwise_reduce": 16.861243999999992
    },
    "top_kernels": [
     {
      "name": "sm90_xmma_fprop_implicit_gemm_f32f32_tf32f32_f32_nhwckrsc_nhwc_tilesize128x128x32_warpgrou",
      "ms": 50.860448
     },
     {
      "name": "ncclDevKernel_SendRecv(ncclDevKernelArgsStorage<(unsigned long)4096>)",
      "ms": 22.470948
     },
     {
      "name": "void at::native::elementwise_kernel<(int)128, (int)2, void at::native::gpu_kernel_impl_noc",
      "ms": 15.692442999999997
     },
     {
      "name": "void cudnn::engines_precompiled::nchwToNhwcKernel<float, float, float, (bool)0, (bool)1, (",
      "ms": 4.524357
     }
    ]
   },
   {
    "unit": "vae/up_blocks.1.resnets.2.conv2/WanDistCausalConv3d",
    "ms": 95.796163,
    "share": 0.017354349808186047,
    "count": 3552,
    "families": {
     "conv": 56.867425000000004,
     "nccl": 22.055809,
     "elementwise_reduce": 16.872928999999996
    },
    "top_kernels": [
     {
      "name": "sm90_xmma_fprop_implicit_gemm_f32f32_tf32f32_f32_nhwckrsc_nhwc_tilesize128x128x32_warpgrou",
      "ms": 51.167640000000006
     },
     {
      "name": "ncclDevKernel_SendRecv(ncclDevKernelArgsStorage<(unsigned long)4096>)",
      "ms": 22.055809
     },
     {
      "name": "void at::native::elementwise_kernel<(int)128, (int)2, void at::native::gpu_kernel_impl_noc",
      "ms": 15.700671999999999
     },
     {
      "name": "void cudnn::engines_precompiled::nchwToNhwcKernel<float, float, float, (bool)0, (bool)1, (",
      "ms": 4.503846
     }
    ]
   },
   {
    "unit": "vae/up_blocks.1.resnets.0.conv1/WanDistCausalConv3d",
    "ms": 81.93986200000003,
    "share": 0.01484415433614488,
    "count": 3548,
    "families": {
     "nccl": 44.270227,
     "conv": 28.852657,
     "elementwise_reduce": 8.816978000000004
    },
    "top_kernels": [
     {
      "name": "ncclDevKernel_SendRecv(ncclDevKernelArgsStorage<(unsigned long)4096>)",
      "ms": 44.270227
     },
     {
      "name": "sm90_xmma_fprop_implicit_gemm_f32f32_tf32f32_f32_nhwckrsc_nhwc_tilesize128x128x32_warpgrou",
      "ms": 25.549959
     },
     {
      "name": "void at::native::elementwise_kernel<(int)128, (int)2, void at::native::gpu_kernel_impl_noc",
      "ms": 8.136816999999999
     },
     {
      "name": "void cudnn::engines_precompiled::nchwToNhwcKernel<float, float, float, (bool)0, (bool)1, (",
      "ms": 2.117253
     }
    ]
   },
   {
    "unit": "vae/up_blocks.0.upsamplers.0.resample.1/WanDistConv2d",
    "ms": 64.66126900000002,
    "share": 0.011713979413426159,
    "count": 2612,
    "families": {
     "nccl": 47.617115999999996,
     "conv": 11.236205,
     "elementwise_reduce": 3.6958470000000014,
     "gather_index_cat": 2.112101
    },
    "top_kernels": [
     {
      "name": "ncclDevKernel_SendRecv(ncclDevKernelArgsStorage<(unsigned long)4096>)",
      "ms": 47.617115999999996
     },
     {
      "name": "sm90_xmma_fprop_implicit_gemm_f32f32_tf32f32_f32_nhwckrsc_nhwc_tilesize128x128x32_warpgrou",
      "ms": 8.655724
     },
     {
      "name": "void at::native::elementwise_kernel<(int)128, (int)2, void at::native::gpu_kernel_impl_noc",
      "ms": 3.024230000000001
     },
     {
      "name": "void at::native::<unnamed>::CatArrayBatchedCopy_alignedK_contig<at::native::<unnamed>::Opa",
      "ms": 2.112101
     }
    ]
   },
   {
    "unit": "ncclAllReduce",
    "ms": 61.302459,
    "share": 0.011105500306812737,
    "count": 160,
    "families": {
     "nccl": 61.302459
    },
    "top_kernels": [
     {
      "name": "ncclDevKernel_AllReduce_Sum_f32_RING_LL(ncclDevKernelArgsStorage<(unsigned long)4096>)",
      "ms": 61.302459
     }
    ]
   },
   {
    "unit": "vae/up_blocks.0.resnets.2.conv2/WanDistCausalConv3d",
    "ms": 58.201639,
    "share": 0.01054375844485299,
    "count": 3584,
    "families": {
     "nccl": 40.542537,
     "conv": 13.712887,
     "elementwise_reduce": 3.946215000000001
    },
    "top_kernels": [
     {
      "name": "ncclDevKernel_SendRecv(ncclDevKernelArgsStorage<(unsigned long)4096>)",
      "ms": 40.542537
     },
     {
      "name": "sm90_xmma_fprop_implicit_gemm_f32f32_tf32f32_f32_nhwckrsc_nhwc_tilesize64x64x32_warpgroups",
      "ms": 11.417141
     },
     {
      "name": "void at::native::elementwise_kernel<(int)128, (int)2, void at::native::gpu_kernel_impl_noc",
      "ms": 3.580807
     },
     {
      "name": "void cudnn::engines_precompiled::nchwToNhwcKernel<float, float, float, (bool)0, (bool)1, (",
      "ms": 1.9849929999999998
     }
    ]
   },
   {
    "unit": "vae/up_blocks.0.resnets.2.conv1/WanDistCausalConv3d",
    "ms": 56.995236000000006,
    "share": 0.010325207523647044,
    "count": 3584,
    "families": {
     "nccl": 39.315177,
     "conv": 13.736213000000001,
     "elementwise_reduce": 3.943846
    },
    "top_kernels": [
     {
      "name": "ncclDevKernel_SendRecv(ncclDevKernelArgsStorage<(unsigned long)4096>)",
      "ms": 39.315177
     },
     {
      "name": "sm90_xmma_fprop_implicit_gemm_f32f32_tf32f32_f32_nhwckrsc_nhwc_tilesize64x64x32_warpgroups",
      "ms": 11.441327000000001
     },
     {
      "name": "void at::native::elementwise_kernel<(int)128, (int)2, void at::native::gpu_kernel_impl_noc",
      "ms": 3.577733
     },
     {
      "name": "void cudnn::engines_precompiled::nchwToNhwcKernel<float, float, float, (bool)0, (bool)1, (",
      "ms": 1.984294
     }
    ]
   },
   {
    "unit": "vae/up_blocks.0.resnets.1.conv2/WanDistCausalConv3d",
    "ms": 56.54240799999999,
    "share": 0.010243173595890024,
    "count": 3584,
    "families": {
     "nccl": 38.810117999999996,
     "conv": 13.780761,
     "elementwise_reduce": 3.9515290000000003
    },
    "top_kernels": [
     {
      "name": "ncclDevKernel_SendRecv(ncclDevKernelArgsStorage<(unsigned long)4096>)",
      "ms": 38.810117999999996
     },
     {
      "name": "sm90_xmma_fprop_implicit_gemm_f32f32_tf32f32_f32_nhwckrsc_nhwc_tilesize64x64x32_warpgroups",
      "ms": 11.471253
     },
     {
      "name": "void at::native::elementwise_kernel<(int)128, (int)2, void at::native::gpu_kernel_impl_noc",
      "ms": 3.5858320000000004
     },
     {
      "name": "void cudnn::engines_precompiled::nchwToNhwcKernel<float, float, float, (bool)0, (bool)1, (",
      "ms": 1.998915
     }
    ]
   },
   {
    "unit": "vae/up_blocks.0.resnets.1.conv1/WanDistCausalConv3d",
    "ms": 56.034414,
    "share": 0.010151145843416687,
    "count": 3584,
    "families": {
     "nccl": 38.321672,
     "conv": 13.770623,
     "elementwise_reduce": 3.942119
    },
    "top_kernels": [
     {
      "name": "ncclDevKernel_SendRecv(ncclDevKernelArgsStorage<(unsigned long)4096>)",
      "ms": 38.321672
     },
     {
      "name": "sm90_xmma_fprop_implicit_gemm_f32f32_tf32f32_f32_nhwckrsc_nhwc_tilesize64x64x32_warpgroups",
      "ms": 11.474842
     },
     {
      "name": "void at::native::elementwise_kernel<(int)128, (int)2, void at::native::gpu_kernel_impl_noc",
      "ms": 3.5756539999999997
     },
     {
      "name": "void cudnn::engines_precompiled::nchwToNhwcKernel<float, float, float, (bool)0, (bool)1, (",
      "ms": 1.9825330000000005
     }
    ]
   },
   {
    "unit": "vae/up_blocks.0.resnets.0.conv2/WanDistCausalConv3d",
    "ms": 54.33062900000001,
    "share": 0.009842489630453961,
    "count": 3584,
    "families": {
     "nccl": 36.650052,
     "conv": 13.727996999999998,
     "elementwise_reduce": 3.95258
    },
    "top_kernels": [
     {
      "name": "ncclDevKernel_SendRecv(ncclDevKernelArgsStorage<(unsigned long)4096>)",
      "ms": 36.650052
     },
     {
      "name": "sm90_xmma_fprop_implicit_gemm_f32f32_tf32f32_f32_nhwckrsc_nhwc_tilesize64x64x32_warpgroups",
      "ms": 11.415638
     },
     {
      "name": "void at::native::elementwise_kernel<(int)128, (int)2, void at::native::gpu_kernel_impl_noc",
      "ms": 3.586148
     },
     {
      "name": "void cudnn::engines_precompiled::nchwToNhwcKernel<float, float, float, (bool)0, (bool)1, (",
      "ms": 2.002407
     }
    ]
   },
   {
    "unit": "vae/up_blocks.3.resnets.0.norm1/WanRMS_norm",
    "ms": 52.00108199999999,
    "share": 0.009420470916274242,
    "count": 1416,
    "families": {
     "elementwise_reduce": 52.00108199999999
    },
    "top_kernels": [
     {
      "name": "void at::native::elementwise_kernel<(int)128, (int)2, void at::native::gpu_kernel_impl_noc",
      "ms": 29.328347000000004
     },
     {
      "name": "void at::native::vectorized_elementwise_kernel<(int)4, at::native::CUDAFunctorOnSelf_add<f",
      "ms": 8.372526
     },
     {
      "name": "void at::native::vectorized_elementwise_kernel<(int)4, at::native::AUnaryFunctor<float, fl",
      "ms": 8.363751
     },
     {
      "name": "void at::native::reduce_kernel<(int)128, (int)4, at::native::ReduceOp<float, at::native::N",
      "ms": 5.692009
     }
    ]
   },
   {
    "unit": "vae/up_blocks.3.resnets.0.norm2/WanRMS_norm",
    "ms": 50.10191900000001,
    "share": 0.009076420194276498,
    "count": 1416,
    "families": {
     "elementwise_reduce": 50.10191900000001
    },
    "top_kernels": [
     {
      "name": "void at::native::elementwise_kernel<(int)128, (int)2, void at::native::gpu_kernel_impl_noc",
      "ms": 27.515384
     },
     {
      "name": "void at::native::vectorized_elementwise_kernel<(int)4, at::native::AUnaryFunctor<float, fl",
      "ms": 8.391437
     },
     {
      "name": "void at::native::vectorized_elementwise_kernel<(int)4, at::native::CUDAFunctorOnSelf_add<f",
      "ms": 8.379246
     },
     {
      "name": "void at::native::reduce_kernel<(int)128, (int)4, at::native::ReduceOp<float, at::native::N",
      "ms": 5.574316
     }
    ]
   },
   {
    "unit": "vae/up_blocks.3.resnets.1.norm2/WanRMS_norm",
    "ms": 50.080473999999995,
    "share": 0.009072535236675046,
    "count": 1416,
    "families": {
     "elementwise_reduce": 50.080473999999995
    },
    "top_kernels": [
     {
      "name": "void at::native::elementwise_kernel<(int)128, (int)2, void at::native::gpu_kernel_impl_noc",
      "ms": 27.486711999999994
     },
     {
      "name": "void at::native::vectorized_elementwise_kernel<(int)4, at::native::AUnaryFunctor<float, fl",
      "ms": 8.396071000000001
     },
     {
      "name": "void at::native::vectorized_elementwise_kernel<(int)4, at::native::CUDAFunctorOnSelf_add<f",
      "ms": 8.378991
     },
     {
      "name": "void at::native::reduce_kernel<(int)128, (int)4, at::native::ReduceOp<float, at::native::N",
      "ms": 5.577195
     }
    ]
   },
   {
    "unit": "vae/up_blocks.3.resnets.2.norm2/WanRMS_norm",
    "ms": 50.06856999999998,
    "share": 0.009070378718358994,
    "count": 1416,
    "families": {
     "elementwise_reduce": 50.06856999999998
    },
    "top_kernels": [
     {
      "name": "void at::native::elementwise_kernel<(int)128, (int)2, void at::native::gpu_kernel_impl_noc",
      "ms": 27.471826
     },
     {
      "name": "void at::native::vectorized_elementwise_kernel<(int)4, at::native::AUnaryFunctor<float, fl",
      "ms": 8.395436
     },
     {
      "name": "void at::native::vectorized_elementwise_kernel<(int)4, at::native::CUDAFunctorOnSelf_add<f",
      "ms": 8.384493
     },
     {
      "name": "void at::native::reduce_kernel<(int)128, (int)4, at::native::ReduceOp<float, at::native::N",
      "ms": 5.574479
     }
    ]
   },
   {
    "unit": "vae/up_blocks.3.resnets.2.norm1/WanRMS_norm",
    "ms": 50.013113999999995,
    "share": 0.009060332357494181,
    "count": 1416,
    "families": {
     "elementwise_reduce": 50.013113999999995
    },
    "top_kernels": [
     {
      "name": "void at::native::elementwise_kernel<(int)128, (int)2, void at::native::gpu_kernel_impl_noc",
      "ms": 27.469714000000007
     },
     {
      "name": "void at::native::vectorized_elementwise_kernel<(int)4, at::native::AUnaryFunctor<float, fl",
      "ms": 8.407726
     },
     {
      "name": "void at::native::vectorized_elementwise_kernel<(int)4, at::native::CUDAFunctorOnSelf_add<f",
      "ms": 8.391792
     },
     {
      "name": "void at::native::reduce_kernel<(int)128, (int)4, at::native::ReduceOp<float, at::native::N",
      "ms": 5.5017700000000005
     }
    ]
   },
   {
    "unit": "vae/norm_out/WanRMS_norm",
    "ms": 50.00370000000001,
    "share": 0.00905862692541864,
    "count": 1416,
    "families": {
     "elementwise_reduce": 50.00370000000001
    },
    "top_kernels": [
     {
      "name": "void at::native::elementwise_kernel<(int)128, (int)2, void at::native::gpu_kernel_impl_noc",
      "ms": 27.458037
     },
     {
      "name": "void at::native::vectorized_elementwise_kernel<(int)4, at::native::AUnaryFunctor<float, fl",
      "ms": 8.411437
     },
     {
      "name": "void at::native::vectorized_elementwise_kernel<(int)4, at::native::CUDAFunctorOnSelf_add<f",
      "ms": 8.39729
     },
     {
      "name": "void at::native::reduce_kernel<(int)128, (int)4, at::native::ReduceOp<float, at::native::N",
      "ms": 5.496262999999999
     }
    ]
   },
   {
    "unit": "vae/up_blocks.0.resnets.0.conv1/WanDistCausalConv3d",
    "ms": 49.81516400000001,
    "share": 0.009024471907169774,
    "count": 3584,
    "families": {
     "nccl": 32.112443000000006,
     "conv": 13.748759999999999,
     "elementwise_reduce": 3.9539609999999996
    },
    "top_kernels": [
     {
      "name": "ncclDevKernel_SendRecv(ncclDevKernelArgsStorage<(unsigned long)4096>)",
      "ms": 32.112443000000006
     },
     {
      "name": "sm90_xmma_fprop_implicit_gemm_f32f32_tf32f32_f32_nhwckrsc_nhwc_tilesize64x64x32_warpgroups",
      "ms": 11.44143
     },
     {
      "name": "void at::native::elementwise_kernel<(int)128, (int)2, void at::native::gpu_kernel_impl_noc",
      "ms": 3.586377
     },
     {
      "name": "void cudnn::engines_precompiled::nchwToNhwcKernel<float, float, float, (bool)0, (bool)1, (",
      "ms": 1.9965460000000002
     }
    ]
   },
   {
    "unit": "vae/up_blocks.3.resnets.1.norm1/WanRMS_norm",
    "ms": 49.78880100000001,
    "share": 0.009019696008953545,
    "count": 1416,
    "families": {
     "elementwise_reduce": 49.78880100000001
    },
    "top_kernels": [
     {
      "name": "void at::native::elementwise_kernel<(int)128, (int)2, void at::native::gpu_kernel_impl_noc",
      "ms": 27.506066
     },
     {
      "name": "void at::native::vectorized_elementwise_kernel<(int)4, at::native::AUnaryFunctor<float, fl",
      "ms": 8.37064
     },
     {
      "name": "void at::native::vectorized_elementwise_kernel<(int)4, at::native::CUDAFunctorOnSelf_add<f",
      "ms": 8.364563
     },
     {
      "name": "void at::native::reduce_kernel<(int)128, (int)4, at::native::ReduceOp<float, at::native::N",
      "ms": 5.303209
     }
    ]
   },
   {
    "unit": "vae/mid_block.resnets.1.conv2/WanDistCausalConv3d",
    "ms": 43.154345000000006,
    "share": 0.00781780371384128,
    "count": 3584,
    "families": {
     "nccl": 25.457102,
     "conv": 13.749304,
     "elementwise_reduce": 3.9479390000000008
    },
    "top_kernels": [
     {
      "name": "ncclDevKernel_SendRecv(ncclDevKernelArgsStorage<(unsigned long)4096>)",
      "ms": 25.457102
     },
     {
      "name": "sm90_xmma_fprop_implicit_gemm_f32f32_tf32f32_f32_nhwckrsc_nhwc_tilesize64x64x32_warpgroups",
      "ms": 11.441077
     },
     {
      "name": "void at::native::elementwise_kernel<(int)128, (int)2, void at::native::gpu_kernel_impl_noc",
      "ms": 3.582211000000001
     },
     {
      "name": "void cudnn::engines_precompiled::nchwToNhwcKernel<float, float, float, (bool)0, (bool)1, (",
      "ms": 1.99885
     }
    ]
   },
   {
    "unit": "vae/up_blocks.2.upsamplers.0.resample.0/WanUpsample",
    "ms": 34.912124,
    "share": 0.006324650105691262,
    "count": 468,
    "families": {
     "gather_index_cat": 27.926994999999998,
     "elementwise_reduce": 6.985129
    },
    "top_kernels": [
     {
      "name": "void at::native::<unnamed>::upsample_nearest2d_out_frame<float, &at::native::nearest_neigh",
      "ms": 27.926994999999998
     },
     {
      "name": "void at::native::elementwise_kernel<(int)128, (int)2, void at::native::gpu_kernel_impl_noc",
      "ms": 6.985129
     }
    ]
   },
   {
    "unit": "vae/mid_block.resnets.1.conv1/WanDistCausalConv3d",
    "ms": 34.18986200000001,
    "share": 0.006193805748165588,
    "count": 3584,
    "families": {
     "nccl": 16.51302,
     "conv": 13.733730000000001,
     "elementwise_reduce": 3.943112
    },
    "top_kernels": [
     {
      "name": "ncclDevKernel_SendRecv(ncclDevKernelArgsStorage<(unsigned long)4096>)",
      "ms": 16.51302
     },
     {
      "name": "sm90_xmma_fprop_implicit_gemm_f32f32_tf32f32_f32_nhwckrsc_nhwc_tilesize64x64x32_warpgroups",
      "ms": 11.422877
     },
     {
      "name": "void at::native::elementwise_kernel<(int)128, (int)2, void at::native::gpu_kernel_impl_noc",
      "ms": 3.577798
     },
     {
      "name": "void cudnn::engines_precompiled::nchwToNhwcKernel<float, float, float, (bool)0, (bool)1, (",
      "ms": 2.000549
     }
    ]
   },
   {
    "unit": "vae/mid_block.resnets.0.conv2/WanDistCausalConv3d",
    "ms": 30.116887999999992,
    "share": 0.0054559493106833565,
    "count": 3584,
    "families": {
     "conv": 13.734428999999999,
     "nccl": 12.421846,
     "elementwise_reduce": 3.960613000000001
    },
    "top_kernels": [
     {
      "name": "ncclDevKernel_SendRecv(ncclDevKernelArgsStorage<(unsigned long)4096>)",
      "ms": 12.421846
     },
     {
      "name": "sm90_xmma_fprop_implicit_gemm_f32f32_tf32f32_f32_nhwckrsc_nhwc_tilesize64x64x32_warpgroups",
      "ms": 11.431835000000001
     },
     {
      "name": "void at::native::elementwise_kernel<(int)128, (int)2, void at::native::gpu_kernel_impl_noc",
      "ms": 3.593892
     },
     {
      "name": "void cudnn::engines_precompiled::nchwToNhwcKernel<float, float, float, (bool)0, (bool)1, (",
      "ms": 1.991426
     }
    ]
   },
   {
    "unit": "vae/mid_block.resnets.0.conv1/WanDistCausalConv3d",
    "ms": 29.254321,
    "share": 0.005299687421039641,
    "count": 3584,
    "families": {
     "conv": 13.825076999999999,
     "nccl": 11.488468000000001,
     "elementwise_reduce": 3.940776
    },
    "top_kernels": [
     {
      "name": "sm90_xmma_fprop_implicit_gemm_f32f32_tf32f32_f32_nhwckrsc_nhwc_tilesize64x64x32_warpgroups",
      "ms": 11.510129
     },
     {
      "name": "ncclDevKernel_SendRecv(ncclDevKernelArgsStorage<(unsigned long)4096>)",
      "ms": 11.488468000000001
     },
     {
      "name": "void at::native::elementwise_kernel<(int)128, (int)2, void at::native::gpu_kernel_impl_noc",
      "ms": 3.578919
     },
     {
      "name": "void cudnn::engines_precompiled::nchwToNhwcKernel<float, float, float, (bool)0, (bool)1, (",
      "ms": 2.002179
     }
    ]
   },
   {
    "unit": "ncclAllGather",
    "ms": 28.532878,
    "share": 0.005168991432843671,
    "count": 476,
    "families": {
     "nccl": 28.532878
    },
    "top_kernels": [
     {
      "name": "ncclDevKernel_AllGather_RING_LL(ncclDevKernelArgsStorage<(unsigned long)4096>)",
      "ms": 28.532878
     }
    ]
   },
   {
    "unit": "vae/up_blocks.2.resnets.0.norm1/WanRMS_norm",
    "ms": 26.204659999999993,
    "share": 0.004747213479151357,
    "count": 1416,
    "families": {
     "elementwise_reduce": 26.204659999999993
    },
    "top_kernels": [
     {
      "name": "void at::native::elementwise_kernel<(int)128, (int)2, void at::native::gpu_kernel_impl_noc",
      "ms": 14.923520000000002
     },
     {
      "name": "void at::native::vectorized_elementwise_kernel<(int)4, at::native::AUnaryFunctor<float, fl",
      "ms": 4.224360999999999
     },
     {
      "name": "void at::native::vectorized_elementwise_kernel<(int)4, at::native::CUDAFunctorOnSelf_add<f",
      "ms": 4.2233979999999995
     },
     {
      "name": "void at::native::reduce_kernel<(int)128, (int)4, at::native::ReduceOp<float, at::native::N",
      "ms": 2.6662129999999995
     }
    ]
   },
   {
    "unit": "vae/up_blocks.2.resnets.0.norm2/WanRMS_norm",
    "ms": 25.682729000000005,
    "share": 0.004652660911845126,
    "count": 1416,
    "families": {
     "elementwise_reduce": 25.682729000000005
    },
    "top_kernels": [
     {
      "name": "void at::native::elementwise_kernel<(int)128, (int)2, void at::native::gpu_kernel_impl_noc",
      "ms": 13.994423
     },
     {
      "name": "void at::native::vectorized_elementwise_kernel<(int)4, at::native::CUDAFunctorOnSelf_add<f",
      "ms": 4.234632
     },
     {
      "name": "void at::native::vectorized_elementwise_kernel<(int)4, at::native::AUnaryFunctor<float, fl",
      "ms": 4.232832999999999
     },
     {
      "name": "void at::native::reduce_kernel<(int)128, (int)4, at::native::ReduceOp<float, at::native::N",
      "ms": 3.0590479999999998
     }
    ]
   },
   {
    "unit": "vae/up_blocks.2.resnets.2.norm1/WanRMS_norm",
    "ms": 25.681846999999998,
    "share": 0.004652501129490054,
    "count": 1416,
    "families": {
     "elementwise_reduce": 25.681846999999998
    },
    "top_kernels": [
     {
      "name": "void at::native::elementwise_kernel<(int)128, (int)2, void at::native::gpu_kernel_impl_noc",
      "ms": 13.974198999999999
     },
     {
      "name": "void at::native::vectorized_elementwise_kernel<(int)4, at::native::CUDAFunctorOnSelf_add<f",
      "ms": 4.2235
     },
     {
      "name": "void at::native::vectorized_elementwise_kernel<(int)4, at::native::AUnaryFunctor<float, fl",
      "ms": 4.221551000000001
     },
     {
      "name": "void at::native::reduce_kernel<(int)128, (int)4, at::native::ReduceOp<float, at::native::N",
      "ms": 3.097861
     }
    ]
   },
   {
    "unit": "vae/up_blocks.2.resnets.2.norm2/WanRMS_norm",
    "ms": 25.671888,
    "share": 0.004650696965687171,
    "count": 1416,
    "families": {
     "elementwise_reduce": 25.671888
    },
    "top_kernels": [
     {
      "name": "void at::native::elementwise_kernel<(int)128, (int)2, void at::native::gpu_kernel_impl_noc",
      "ms": 14.015357000000002
     },
     {
      "name": "void at::native::vectorized_elementwise_kernel<(int)4, at::native::AUnaryFunctor<float, fl",
      "ms": 4.220422999999999
     },
     {
      "name": "void at::native::vectorized_elementwise_kernel<(int)4, at::native::CUDAFunctorOnSelf_add<f",
      "ms": 4.2150810000000005
     },
     {
      "name": "void at::native::reduce_kernel<(int)128, (int)4, at::native::ReduceOp<float, at::native::N",
      "ms": 3.057923
     }
    ]
   },
   {
    "unit": "vae/up_blocks.2.resnets.1.norm2/WanRMS_norm",
    "ms": 25.656399,
    "share": 0.004647890991880277,
    "count": 1416,
    "families": {
     "elementwise_reduce": 25.656399
    },
    "top_kernels": [
     {
      "name": "void at::native::elementwise_kernel<(int)128, (int)2, void at::native::gpu_kernel_impl_noc",
      "ms": 13.99372
     },
     {
      "name": "void at::native::vectorized_elementwise_kernel<(int)4, at::native::AUnaryFunctor<float, fl",
      "ms": 4.224357
     },
     {
      "name": "void at::native::vectorized_elementwise_kernel<(int)4, at::native::CUDAFunctorOnSelf_add<f",
      "ms": 4.223177
     },
     {
      "name": "void at::native::reduce_kernel<(int)128, (int)4, at::native::ReduceOp<float, at::native::N",
      "ms": 3.050856
     }
    ]
   },
   {
    "unit": "vae/up_blocks.2.resnets.1.norm1/WanRMS_norm",
    "ms": 25.515151999999997,
    "share": 0.004622302807859201,
    "count": 1416,
    "families": {
     "elementwise_reduce": 25.515151999999997
    },
    "top_kernels": [
     {
      "name": "void at::native::elementwise_kernel<(int)128, (int)2, void at::native::gpu_kernel_impl_noc",
      "ms": 13.981561999999998
     },
     {
      "name": "void at::native::vectorized_elementwise_kernel<(int)4, at::native::CUDAFunctorOnSelf_add<f",
      "ms": 4.210565
     },
     {
      "name": "void at::native::vectorized_elementwise_kernel<(int)4, at::native::AUnaryFunctor<float, fl",
      "ms": 4.205611
     },
     {
      "name": "void at::native::reduce_kernel<(int)128, (int)4, at::native::ReduceOp<float, at::native::N",
      "ms": 2.95063
     }
    ]
   },
   {
    "unit": "vae/up_blocks.1.upsamplers.0.time_conv/WanDistCausalConv3d",
    "ms": 23.679242000000002,
    "share": 0.004289710944484186,
    "count": 1856,
    "families": {
     "conv": 16.541213,
     "elementwise_reduce": 7.138029
    },
    "top_kernels": [
     {
      "name": "sm90_xmma_fprop_implicit_gemm_f32f32_tf32f32_f32_nhwckrsc_nhwc_tilesize128x128x32_warpgrou",
      "ms": 9.306419
     },
     {
      "name": "void at::native::elementwise_kernel<(int)128, (int)2, void at::native::gpu_kernel_impl_noc",
      "ms": 7.100973000000001
     },
     {
      "name": "void cudnn::engines_precompiled::nchwToNhwcKernel<float, float, float, (bool)0, (bool)1, (",
      "ms": 4.522373
     },
     {
      "name": "void cudnn::engines_precompiled::nhwcToNchwKernel<float, float, float, (bool)1, (bool)0, (",
      "ms": 2.45722
     }
    ]
   },
   {
    "unit": "vae/up_blocks.1.upsamplers.0.resample.0/WanUpsample",
    "ms": 17.534402,
    "share": 0.0031765170593038997,
    "count": 468,
    "families": {
     "gather_index_cat": 14.000471,
     "elementwise_reduce": 3.533931
    },
    "top_kernels": [
     {
      "name": "void at::native::<unnamed>::upsample_nearest2d_out_frame<float, &at::native::nearest_neigh",
      "ms": 14.000471
     },
     {
      "name": "void at::native::elementwise_kernel<(int)128, (int)2, void at::native::gpu_kernel_impl_noc",
      "ms": 3.533931
     }
    ]
   },
   {
    "unit": "vae/up_blocks.3.resnets.0.nonlinearity/SiLU",
    "ms": 17.239046000000002,
    "share": 0.0031230106224965446,
    "count": 472,
    "families": {
     "elementwise_reduce": 17.239046000000002
    },
    "top_kernels": [
     {
      "name": "void at::native::vectorized_elementwise_kernel<(int)4, at::native::<unnamed>::silu_kernel(",
      "ms": 17.239046000000002
     }
    ]
   },
   {
    "unit": "vae/up_blocks.3.resnets.1.nonlinearity/SiLU",
    "ms": 17.228633,
    "share": 0.0031211242124474,
    "count": 472,
    "families": {
     "elementwise_reduce": 17.228633
    },
    "top_kernels": [
     {
      "name": "void at::native::vectorized_elementwise_kernel<(int)4, at::native::<unnamed>::silu_kernel(",
      "ms": 17.228633
     }
    ]
   },
   {
    "unit": "vae/up_blocks.3.resnets.2.nonlinearity/SiLU",
    "ms": 17.223222,
    "share": 0.0031201439603801844,
    "count": 472,
    "families": {
     "elementwise_reduce": 17.223222
    },
    "top_kernels": [
     {
      "name": "void at::native::vectorized_elementwise_kernel<(int)4, at::native::<unnamed>::silu_kernel(",
      "ms": 17.223222
     }
    ]
   },
   {
    "unit": "vae/conv_in/WanDistCausalConv3d",
    "ms": 13.410519,
    "share": 0.002429437991533391,
    "count": 3344,
    "families": {
     "nccl": 9.115252,
     "conv": 2.3960670000000004,
     "elementwise_reduce": 1.8991999999999993
    },
    "top_kernels": [
     {
      "name": "ncclDevKernel_SendRecv(ncclDevKernelArgsStorage<(unsigned long)4096>)",
      "ms": 9.115252
     },
     {
      "name": "sm80_xmma_fprop_implicit_gemm_tf32f32_tf32f32_f32_nhwckrsc_nchw_tilesize128x64x32_stage5_w",
      "ms": 1.901987
     },
     {
      "name": "void at::native::elementwise_kernel<(int)128, (int)2, void at::native::gpu_kernel_impl_noc",
      "ms": 1.7359999999999998
     },
     {
      "name": "void cudnn::engines_precompiled::nchwToNhwcKernel<float, float, float, (bool)0, (bool)1, (",
      "ms": 0.49408
     }
    ]
   },
   {
    "unit": "vae/up_blocks.2.resnets.1.nonlinearity/SiLU",
    "ms": 8.878157999999999,
    "share": 0.0016083594035425552,
    "count": 472,
    "families": {
     "elementwise_reduce": 8.878157999999999
    },
    "top_kernels": [
     {
      "name": "void at::native::vectorized_elementwise_kernel<(int)4, at::native::<unnamed>::silu_kernel(",
      "ms": 8.878157999999999
     }
    ]
   },
   {
    "unit": "vae/up_blocks.2.resnets.2.nonlinearity/SiLU",
    "ms": 8.873456,
    "share": 0.0016075075933004467,
    "count": 472,
    "families": {
     "elementwise_reduce": 8.873456
    },
    "top_kernels": [
     {
      "name": "void at::native::vectorized_elementwise_kernel<(int)4, at::native::<unnamed>::silu_kernel(",
      "ms": 8.873456
     }
    ]
   },
   {
    "unit": "vae/up_blocks.2.resnets.0.nonlinearity/SiLU",
    "ms": 8.863985999999999,
    "share": 0.0016057920163134694,
    "count": 472,
    "families": {
     "elementwise_reduce": 8.863985999999999
    },
    "top_kernels": [
     {
      "name": "void at::native::vectorized_elementwise_kernel<(int)4, at::native::<unnamed>::silu_kernel(",
      "ms": 8.863985999999999
     }
    ]
   },
   {
    "unit": "vae/nonlinearity/SiLU",
    "ms": 8.629166,
    "share": 0.001563252228765212,
    "count": 236,
    "families": {
     "elementwise_reduce": 8.629166
    },
    "top_kernels": [
     {
      "name": "void at::native::vectorized_elementwise_kernel<(int)4, at::native::<unnamed>::silu_kernel(",
      "ms": 8.629166
     }
    ]
   },
   {
    "unit": "vae/mid_block.attentions.0.to_qkv/Conv2d",
    "ms": 7.164199,
    "share": 0.0012978600775634059,
    "count": 480,
    "families": {
     "gemm": 4.840673,
     "elementwise_reduce": 2.323526
    },
    "top_kernels": [
     {
      "name": "void cutlass::Kernel2<cutlass_80_tensorop_s1688gemm_128x128_32x3_nn_align4>(T1::Params)",
      "ms": 4.840673
     },
     {
      "name": "void at::native::elementwise_kernel<(int)128, (int)2, void at::native::gpu_kernel_impl_noc",
      "ms": 2.323526
     }
    ]
   },
   {
    "unit": "vae/up_blocks.1.resnets.2.norm1/WanRMS_norm",
    "ms": 6.650377999999999,
    "share": 0.0012047767108236337,
    "count": 1416,
    "families": {
     "elementwise_reduce": 6.650377999999999
    },
    "top_kernels": [
     {
      "name": "void at::native::elementwise_kernel<(int)128, (int)2, void at::native::gpu_kernel_impl_noc",
      "ms": 3.2929950000000003
     },
     {
      "name": "void at::native::reduce_kernel<(int)128, (int)4, at::native::ReduceOp<float, at::native::N",
      "ms": 1.364003
     },
     {
      "name": "void at::native::vectorized_elementwise_kernel<(int)4, at::native::AUnaryFunctor<float, fl",
      "ms": 0.9280659999999998
     },
     {
      "name": "void at::native::vectorized_elementwise_kernel<(int)4, at::native::CUDAFunctorOnSelf_add<f",
      "ms": 0.9204180000000001
     }
    ]
   },
   {
    "unit": "vae/up_blocks.1.resnets.1.norm1/WanRMS_norm",
    "ms": 6.644136000000001,
    "share": 0.0012036459155171176,
    "count": 1416,
    "families": {
     "elementwise_reduce": 6.644136000000001
    },
    "top_kernels": [
     {
      "name": "void at::native::elementwise_kernel<(int)128, (int)2, void at::native::gpu_kernel_impl_noc",
      "ms": 3.280645
     },
     {
      "name": "void at::native::reduce_kernel<(int)128, (int)4, at::native::ReduceOp<float, at::native::N",
      "ms": 1.369378
     },
     {
      "name": "void at::native::vectorized_elementwise_kernel<(int)4, at::native::AUnaryFunctor<float, fl",
      "ms": 0.924929
     },
     {
      "name": "void at::native::vectorized_elementwise_kernel<(int)4, at::native::CUDAFunctorOnSelf_add<f",
      "ms": 0.924544
     }
    ]
   },
   {
    "unit": "vae/up_blocks.1.resnets.0.norm2/WanRMS_norm",
    "ms": 6.583337,
    "share": 0.001192631621406111,
    "count": 1416,
    "families": {
     "elementwise_reduce": 6.583337
    },
    "top_kernels": [
     {
      "name": "void at::native::elementwise_kernel<(int)128, (int)2, void at::native::gpu_kernel_impl_noc",
      "ms": 3.291524
     },
     {
      "name": "void at::native::reduce_kernel<(int)128, (int)4, at::native::ReduceOp<float, at::native::N",
      "ms": 1.29613
     },
     {
      "name": "void at::native::vectorized_elementwise_kernel<(int)4, at::native::CUDAFunctorOnSelf_add<f",
      "ms": 0.925505
     },
     {
      "name": "void at::native::vectorized_elementwise_kernel<(int)4, at::native::AUnaryFunctor<float, fl",
      "ms": 0.924289
     }
    ]
   },
   {
    "unit": "vae/up_blocks.1.resnets.1.norm2/WanRMS_norm",
    "ms": 6.578251000000001,
    "share": 0.0011917102460570334,
    "count": 1416,
    "families": {
     "elementwise_reduce": 6.578251000000001
    },
    "top_kernels": [
     {
      "name": "void at::native::elementwise_kernel<(int)128, (int)2, void at::native::gpu_kernel_impl_noc",
      "ms": 3.281252
     },
     {
      "name": "void at::native::reduce_kernel<(int)128, (int)4, at::native::ReduceOp<float, at::native::N",
      "ms": 1.3018599999999998
     },
     {
      "name": "void at::native::vectorized_elementwise_kernel<(int)4, at::native::AUnaryFunctor<float, fl",
      "ms": 0.927266
     },
     {
      "name": "void at::native::vectorized_elementwise_kernel<(int)4, at::native::CUDAFunctorOnSelf_add<f",
      "ms": 0.922816
     }
    ]
   },
   {
    "unit": "vae/up_blocks.1.resnets.2.norm2/WanRMS_norm",
    "ms": 6.572971000000003,
    "share": 0.001190753725836206,
    "count": 1416,
    "families": {
     "elementwise_reduce": 6.572971000000003
    },
    "top_kernels": [
     {
      "name": "void at::native::elementwise_kernel<(int)128, (int)2, void at::native::gpu_kernel_impl_noc",
      "ms": 3.2760039999999995
     },
     {
      "name": "void at::native::reduce_kernel<(int)128, (int)4, at::native::ReduceOp<float, at::native::N",
      "ms": 1.301989
     },
     {
      "name": "void at::native::vectorized_elementwise_kernel<(int)4, at::native::AUnaryFunctor<float, fl",
      "ms": 0.9270720000000001
     },
     {
      "name": "void at::native::vectorized_elementwise_kernel<(int)4, at::native::CUDAFunctorOnSelf_add<f",
      "ms": 0.9232649999999999
     }
    ]
   },
   {
    "unit": "vae/up_blocks.0.upsamplers.0.time_conv/WanDistCausalConv3d",
    "ms": 6.216590999999999,
    "share": 0.0011261922341129788,
    "count": 1624,
    "families": {
     "conv": 4.788488999999999,
     "elementwise_reduce": 1.4281019999999998
    },
    "top_kernels": [
     {
      "name": "void cudnn::engines_precompiled::nchwToNhwcKernel<float, float, float, (bool)0, (bool)1, (",
      "ms": 2.343526
     },
     {
      "name": "sm90_xmma_fprop_implicit_gemm_f32f32_tf32f32_f32_nhwckrsc_nhwc_tilesize128x128x32_warpgrou",
      "ms": 2.07341
     },
     {
      "name": "void at::native::elementwise_kernel<(int)128, (int)2, void at::native::gpu_kernel_impl_noc",
      "ms": 1.4185339999999997
     },
     {
      "name": "void cudnn::engines_precompiled::nhwcToNchwKernel<float, float, float, (bool)1, (bool)0, (",
      "ms": 0.371553
     }
    ]
   },
   {
    "unit": "vae/mid_block.attentions.0.proj/Conv2d",
    "ms": 5.273060999999999,
    "share": 0.0009552631576058355,
    "count": 720,
    "families": {
     "elementwise_reduce": 3.067748,
     "gemm": 2.205313
    },
    "top_kernels": [
     {
      "name": "void at::native::elementwise_kernel<(int)128, (int)2, void at::native::gpu_kernel_impl_noc",
      "ms": 3.067748
     },
     {
      "name": "void cutlass::Kernel2<cutlass_80_tensorop_s1688gemm_64x128_32x3_nn_align4>(T1::Params)",
      "ms": 2.205313
     }
    ]
   },
   {
    "unit": "vae/up_blocks.1.resnets.0.conv_shortcut/WanDistCausalConv3d",
    "ms": 4.696998999999999,
    "share": 0.0008509042652856569,
    "count": 704,
    "families": {
     "elementwise_reduce": 2.4091579999999997,
     "gemm": 2.2878410000000002
    },
    "top_kernels": [
     {
      "name": "void at::native::elementwise_kernel<(int)128, (int)2, void at::native::gpu_kernel_impl_noc",
      "ms": 2.4091579999999997
     },
     {
      "name": "void cutlass::Kernel2<cutlass_80_tensorop_s1688gemm_128x64_16x6_nn_align4>(T1::Params)",
      "ms": 2.241601
     },
     {
      "name": "void cutlass::Kernel2<cutlass_80_tensorop_s1688gemm_64x128_32x3_nn_align4>(T1::Params)",
      "ms": 0.04624
     }
    ]
   },
   {
    "unit": "vae/up_blocks.1.resnets.0.norm1/WanRMS_norm",
    "ms": 4.526349000000001,
    "share": 0.0008199894592848475,
    "count": 1416,
    "families": {
     "elementwise_reduce": 4.526349000000001
    },
    "top_kernels": [
     {
      "name": "void at::native::elementwise_kernel<(int)128, (int)2, void at::native::gpu_kernel_impl_noc",
      "ms": 1.9959410000000002
     },
     {
      "name": "void at::native::reduce_kernel<(int)128, (int)4, at::native::ReduceOp<float, at::native::N",
      "ms": 1.317253
     },
     {
      "name": "void at::native::vectorized_elementwise_kernel<(int)4, at::native::AUnaryFunctor<float, fl",
      "ms": 0.5351049999999999
     },
     {
      "name": "void at::native::vectorized_elementwise_kernel<(int)4, at::native::CUDAFunctorOnSelf_add<f",
      "ms": 0.531554
     }
    ]
   },
   {
    "unit": "vae/mid_block.attentions.0.norm/WanRMS_norm",
    "ms": 4.389802000000001,
    "share": 0.000795252723187616,
    "count": 1440,
    "families": {
     "elementwise_reduce": 4.389802000000001
    },
    "top_kernels": [
     {
      "name": "void at::native::elementwise_kernel<(int)128, (int)2, void at::native::gpu_kernel_impl_noc",
      "ms": 1.859139
     },
     {
      "name": "void at::native::reduce_kernel<(int)128, (int)4, at::native::ReduceOp<float, at::native::N",
      "ms": 1.307203
     },
     {
      "name": "void at::native::vectorized_elementwise_kernel<(int)4, at::native::CUDAFunctorOnSelf_add<f",
      "ms": 0.543233
     },
     {
      "name": "void at::native::vectorized_elementwise_kernel<(int)4, at::native::AUnaryFunctor<float, fl",
      "ms": 0.541731
     }
    ]
   },
   {
    "unit": "vae/mid_block.resnets.0.norm2/WanRMS_norm",
    "ms": 2.6398780000000004,
    "share": 0.0004782380089997401,
    "count": 1440,
    "families": {
     "elementwise_reduce": 2.6398780000000004
    },
    "top_kernels": [
     {
      "name": "void at::native::reduce_kernel<(int)128, (int)4, at::native::ReduceOp<float, at::native::N",
      "ms": 1.2402289999999998
     },
     {
      "name": "void at::native::elementwise_kernel<(int)128, (int)2, void at::native::gpu_kernel_impl_noc",
      "ms": 0.7941450000000001
     },
     {
      "name": "void at::native::vectorized_elementwise_kernel<(int)4, at::native::CUDAFunctorOnSelf_add<f",
      "ms": 0.224192
     },
     {
      "name": "void at::native::vectorized_elementwise_kernel<(int)4, at::native::AUnaryFunctor<float, fl",
      "ms": 0.222464
     }
    ]
   },
   {
    "unit": "vae/mid_block.resnets.1.norm1/WanRMS_norm",
    "ms": 2.639651,
    "share": 0.00047819688587660977,
    "count": 1440,
    "families": {
     "elementwise_reduce": 2.639651
    },
    "top_kernels": [
     {
      "name": "void at::native::reduce_kernel<(int)128, (int)4, at::native::ReduceOp<float, at::native::N",
      "ms": 1.241026
     },
     {
      "name": "void at::native::elementwise_kernel<(int)128, (int)2, void at::native::gpu_kernel_impl_noc",
      "ms": 0.7924479999999999
     },
     {
      "name": "void at::native::vectorized_elementwise_kernel<(int)4, at::native::AUnaryFunctor<float, fl",
      "ms": 0.223968
     },
     {
      "name": "void at::native::vectorized_elementwise_kernel<(int)4, at::native::CUDAFunctorOnSelf_add<f",
      "ms": 0.22192099999999998
     }
    ]
   },
   {
    "unit": "vae/up_blocks.0.resnets.0.norm2/WanRMS_norm",
    "ms": 2.6309499999999995,
    "share": 0.0004766206202627037,
    "count": 1440,
    "families": {
     "elementwise_reduce": 2.6309499999999995
    },
    "top_kernels": [
     {
      "name": "void at::native::reduce_kernel<(int)128, (int)4, at::native::ReduceOp<float, at::native::N",
      "ms": 1.236289
     },
     {
      "name": "void at::native::elementwise_kernel<(int)128, (int)2, void at::native::gpu_kernel_impl_noc",
      "ms": 0.789508
     },
     {
      "name": "void at::native::vectorized_elementwise_kernel<(int)4, at::native::AUnaryFunctor<float, fl",
      "ms": 0.224064
     },
     {
      "name": "void at::native::vectorized_elementwise_kernel<(int)4, at::native::CUDAFunctorOnSelf_add<f",
      "ms": 0.22246399999999997
     }
    ]
   },
   {
    "unit": "vae/mid_block.resnets.1.norm2/WanRMS_norm",
    "ms": 2.6304050000000005,
    "share": 0.0004765218885353647,
    "count": 1440,
    "families": {
     "elementwise_reduce": 2.6304050000000005
    },
    "top_kernels": [
     {
      "name": "void at::native::reduce_kernel<(int)128, (int)4, at::native::ReduceOp<float, at::native::N",
      "ms": 1.236515
     },
     {
      "name": "void at::native::elementwise_kernel<(int)128, (int)2, void at::native::gpu_kernel_impl_noc",
      "ms": 0.792834
     },
     {
      "name": "void at::native::vectorized_elementwise_kernel<(int)4, at::native::AUnaryFunctor<float, fl",
      "ms": 0.221824
     },
     {
      "name": "void at::native::vectorized_elementwise_kernel<(int)4, at::native::CUDAFunctorOnSelf_add<f",
      "ms": 0.22128
     }
    ]
   },
   {
    "unit": "vae/up_blocks.0.resnets.2.norm2/WanRMS_norm",
    "ms": 2.6289960000000008,
    "share": 0.00047626663531734456,
    "count": 1440,
    "families": {
     "elementwise_reduce": 2.6289960000000008
    },
    "top_kernels": [
     {
      "name": "void at::native::reduce_kernel<(int)128, (int)4, at::native::ReduceOp<float, at::native::N",
      "ms": 1.2368329999999998
     },
     {
      "name": "void at::native::elementwise_kernel<(int)128, (int)2, void at::native::gpu_kernel_impl_noc",
      "ms": 0.7893439999999999
     },
     {
      "name": "void at::native::vectorized_elementwise_kernel<(int)4, at::native::AUnaryFunctor<float, fl",
      "ms": 0.222689
     },
     {
      "name": "void at::native::vectorized_elementwise_kernel<(int)4, at::native::CUDAFunctorOnSelf_add<f",
      "ms": 0.22195299999999998
     }
    ]
   },
   {
    "unit": "vae/mid_block.resnets.0.norm1/WanRMS_norm",
    "ms": 2.6266909999999997,
    "share": 0.00047584906351639584,
    "count": 1440,
    "families": {
     "elementwise_reduce": 2.6266909999999997
    },
    "top_kernels": [
     {
      "name": "void at::native::reduce_kernel<(int)128, (int)4, at::native::ReduceOp<float, at::native::N",
      "ms": 1.2364819999999999
     },
     {
      "name": "void at::native::elementwise_kernel<(int)128, (int)2, void at::native::gpu_kernel_impl_noc",
      "ms": 0.7924810000000001
     },
     {
      "name": "void at::native::vectorized_elementwise_kernel<(int)4, at::native::AUnaryFunctor<float, fl",
      "ms": 0.22259199999999998
     },
     {
      "name": "void at::native::vectorized_elementwise_kernel<(int)4, at::native::CUDAFunctorOnSelf_add<f",
      "ms": 0.21929600000000002
     }
    ]
   },
   {
    "unit": "vae/up_blocks.0.resnets.1.norm2/WanRMS_norm",
    "ms": 2.6215740000000003,
    "share": 0.0004749220722342035,
    "count": 1440,
    "families": {
     "elementwise_reduce": 2.6215740000000003
    },
    "top_kernels": [
     {
      "name": "void at::native::reduce_kernel<(int)128, (int)4, at::native::ReduceOp<float, at::native::N",
      "ms": 1.233444
     },
     {
      "name": "void at::native::elementwise_kernel<(int)128, (int)2, void at::native::gpu_kernel_impl_noc",
      "ms": 0.788578
     },
     {
      "name": "void at::native::vectorized_elementwise_kernel<(int)4, at::native::CUDAFunctorOnSelf_add<f",
      "ms": 0.22160000000000002
     },
     {
      "name": "void at::native::vectorized_elementwise_kernel<(int)4, at::native::AUnaryFunctor<float, fl",
      "ms": 0.220544
     }
    ]
   },
   {
    "unit": "vae/up_blocks.0.resnets.0.norm1/WanRMS_norm",
    "ms": 2.599043,
    "share": 0.0004708403758146063,
    "count": 1440,
    "families": {
     "elementwise_reduce": 2.599043
    },
    "top_kernels": [
     {
      "name": "void at::native::reduce_kernel<(int)128, (int)4, at::native::ReduceOp<float, at::native::N",
      "ms": 1.207841
     },
     {
      "name": "void at::native::elementwise_kernel<(int)128, (int)2, void at::native::gpu_kernel_impl_noc",
      "ms": 0.7899530000000001
     },
     {
      "name": "void at::native::vectorized_elementwise_kernel<(int)4, at::native::AUnaryFunctor<float, fl",
      "ms": 0.22195299999999998
     },
     {
      "name": "void at::native::vectorized_elementwise_kernel<(int)4, at::native::CUDAFunctorOnSelf_add<f",
      "ms": 0.21923199999999998
     }
    ]
   },
   {
    "unit": "vae/up_blocks.0.resnets.1.norm1/WanRMS_norm",
    "ms": 2.5975379999999992,
    "share": 0.0004705677313198437,
    "count": 1440,
    "families": {
     "elementwise_reduce": 2.5975379999999992
    },
    "top_kernels": [
     {
      "name": "void at::native::reduce_kernel<(int)128, (int)4, at::native::ReduceOp<float, at::native::N",
      "ms": 1.208897
     },
     {
      "name": "void at::native::elementwise_kernel<(int)128, (int)2, void at::native::gpu_kernel_impl_noc",
      "ms": 0.7901119999999999
     },
     {
      "name": "void at::native::vectorized_elementwise_kernel<(int)4, at::native::AUnaryFunctor<float, fl",
      "ms": 0.220736
     },
     {
      "name": "void at::native::vectorized_elementwise_kernel<(int)4, at::native::CUDAFunctorOnSelf_add<f",
      "ms": 0.219041
     }
    ]
   },
   {
    "unit": "vae/up_blocks.0.resnets.2.norm1/WanRMS_norm",
    "ms": 2.592198,
    "share": 0.00046960034155105196,
    "count": 1440,
    "families": {
     "elementwise_reduce": 2.592198
    },
    "top_kernels": [
     {
      "name": "void at::native::reduce_kernel<(int)128, (int)4, at::native::ReduceOp<float, at::native::N",
      "ms": 1.20493
     },
     {
      "name": "void at::native::elementwise_kernel<(int)128, (int)2, void at::native::gpu_kernel_impl_noc",
      "ms": 0.789795
     },
     {
      "name": "void at::native::vectorized_elementwise_kernel<(int)4, at::native::CUDAFunctorOnSelf_add<f",
      "ms": 0.221152
     },
     {
      "name": "void at::native::vectorized_elementwise_kernel<(int)4, at::native::AUnaryFunctor<float, fl",
      "ms": 0.219296
     }
    ]
   },
   {
    "unit": "vae/up_blocks.0.upsamplers.0.resample.0/WanUpsample",
    "ms": 2.0353,
    "share": 0.00036871318285056006,
    "count": 468,
    "families": {
     "gather_index_cat": 1.5237769999999997,
     "elementwise_reduce": 0.5115230000000001
    },
    "top_kernels": [
     {
      "name": "void at::native::<unnamed>::upsample_nearest2d_out_frame<float, &at::native::nearest_neigh",
      "ms": 1.5237769999999997
     },
     {
      "name": "void at::native::elementwise_kernel<(int)128, (int)2, void at::native::gpu_kernel_impl_noc",
      "ms": 0.5115230000000001
     }
    ]
   },
   {
    "unit": "vae/up_blocks.1.resnets.2.nonlinearity/SiLU",
    "ms": 1.906338,
    "share": 0.00034535053877510495,
    "count": 472,
    "families": {
     "elementwise_reduce": 1.906338
    },
    "top_kernels": [
     {
      "name": "void at::native::vectorized_elementwise_kernel<(int)4, at::native::<unnamed>::silu_kernel(",
      "ms": 1.906338
     }
    ]
   },
   {
    "unit": "vae/up_blocks.1.resnets.1.nonlinearity/SiLU",
    "ms": 1.905443,
    "share": 0.0003451884013513093,
    "count": 472,
    "families": {
     "elementwise_reduce": 1.905443
    },
    "top_kernels": [
     {
      "name": "void at::native::vectorized_elementwise_kernel<(int)4, at::native::<unnamed>::silu_kernel(",
      "ms": 1.905443
     }
    ]
   },
   {
    "unit": "vae_convert/ep*_chunk*",
    "ms": 1.5325169999999997,
    "share": 0.0002776294506178901,
    "count": 100,
    "families": {
     "elementwise_reduce": 1.5325169999999997
    },
    "top_kernels": [
     {
      "name": "void at::native::unrolled_elementwise_kernel<at::native::direct_copy_kernel_cuda(at::Tenso",
      "ms": 0.412258
     },
     {
      "name": "void at::native::vectorized_elementwise_kernel<(int)4, at::native::<unnamed>::launch_clamp",
      "ms": 0.284704
     },
     {
      "name": "void at::native::vectorized_elementwise_kernel<(int)4, at::native::round_kernel_cuda(at::T",
      "ms": 0.279168
     },
     {
      "name": "void at::native::vectorized_elementwise_kernel<(int)4, at::native::CUDAFunctorOnSelf_add<f",
      "ms": 0.278402
     }
    ]
   },
   {
    "unit": "vae/up_blocks.1.resnets.0.nonlinearity/SiLU",
    "ms": 1.474275,
    "share": 0.0002670783804092809,
    "count": 476,
    "families": {
     "elementwise_reduce": 1.474275
    },
    "top_kernels": [
     {
      "name": "void at::native::vectorized_elementwise_kernel<(int)4, at::native::<unnamed>::silu_kernel(",
      "ms": 1.474275
     }
    ]
   },
   {
    "unit": "no context",
    "ms": 0.8738570000000001,
    "share": 0.00015830717625226842,
    "count": 100,
    "families": {
     "elementwise_reduce": 0.8738570000000001
    },
    "top_kernels": [
     {
      "name": "void at::native::vectorized_elementwise_kernel<(int)4, at::native::AbsFunctor<float>, std:",
      "ms": 0.259744
     },
     {
      "name": "void at::native::vectorized_elementwise_kernel<(int)8, at::native::AUnaryFunctor<float, fl",
      "ms": 0.18544000000000002
     },
     {
      "name": "void at::native::vectorized_elementwise_kernel<(int)8, at::native::BinaryFunctor<float, fl",
      "ms": 0.180928
     },
     {
      "name": "void at::native::vectorized_elementwise_kernel<(int)8, at::native::BinaryFunctor<bool, boo",
      "ms": 0.12560100000000002
     }
    ]
   },
   {
    "unit": "vae_decode/ep*_chunk*",
    "ms": 0.7341759999999999,
    "share": 0.000133002687433053,
    "count": 78,
    "families": {
     "elementwise_reduce": 0.7341759999999999
    },
    "top_kernels": [
     {
      "name": "void at::native::elementwise_kernel<(int)128, (int)2, void at::native::gpu_kernel_impl_noc",
      "ms": 0.47344
     },
     {
      "name": "void at::native::vectorized_elementwise_kernel<(int)4, at::native::<unnamed>::launch_clamp",
      "ms": 0.260736
     }
    ]
   },
   {
    "unit": "vae/mid_block.resnets.0.nonlinearity/SiLU",
    "ms": 0.524129,
    "share": 9.495075507997897e-05,
    "count": 480,
    "families": {
     "elementwise_reduce": 0.524129
    },
    "top_kernels": [
     {
      "name": "void at::native::vectorized_elementwise_kernel<(int)4, at::native::<unnamed>::silu_kernel(",
      "ms": 0.524129
     }
    ]
   },
   {
    "unit": "vae/up_blocks.0.resnets.0.nonlinearity/SiLU",
    "ms": 0.522147,
    "share": 9.459169767890307e-05,
    "count": 480,
    "families": {
     "elementwise_reduce": 0.522147
    },
    "top_kernels": [
     {
      "name": "void at::native::vectorized_elementwise_kernel<(int)4, at::native::<unnamed>::silu_kernel(",
      "ms": 0.522147
     }
    ]
   },
   {
    "unit": "vae/up_blocks.0.resnets.1.nonlinearity/SiLU",
    "ms": 0.52208,
    "share": 9.457956001700998e-05,
    "count": 480,
    "families": {
     "elementwise_reduce": 0.52208
    },
    "top_kernels": [
     {
      "name": "void at::native::vectorized_elementwise_kernel<(int)4, at::native::<unnamed>::silu_kernel(",
      "ms": 0.52208
     }
    ]
   },
   {
    "unit": "vae/up_blocks.0.resnets.2.nonlinearity/SiLU",
    "ms": 0.52192,
    "share": 9.455057455577279e-05,
    "count": 480,
    "families": {
     "elementwise_reduce": 0.52192
    },
    "top_kernels": [
     {
      "name": "void at::native::vectorized_elementwise_kernel<(int)4, at::native::<unnamed>::silu_kernel(",
      "ms": 0.52192
     }
    ]
   },
   {
    "unit": "vae/mid_block.resnets.1.nonlinearity/SiLU",
    "ms": 0.521024,
    "share": 9.438825597284442e-05,
    "count": 480,
    "families": {
     "elementwise_reduce": 0.521024
    },
    "top_kernels": [
     {
      "name": "void at::native::vectorized_elementwise_kernel<(int)4, at::native::<unnamed>::silu_kernel(",
      "ms": 0.521024
     }
    ]
   },
   {
    "unit": "vae_pqc/root/WanCausalConv3d",
    "ms": 0.26105599999999995,
    "share": 4.7292678554628705e-05,
    "count": 160,
    "families": {
     "gemm": 0.16691199999999998,
     "elementwise_reduce": 0.094144
    },
    "top_kernels": [
     {
      "name": "void cutlass::Kernel2<cutlass_80_tensorop_s1688gemm_128x64_16x6_nn_align4>(T1::Params)",
      "ms": 0.16691199999999998
     },
     {
      "name": "void at::native::elementwise_kernel<(int)128, (int)2, void at::native::gpu_kernel_impl_noc",
      "ms": 0.094144
     }
    ]
   },
   {
    "unit": "vae_h2d/ep*_chunk*",
    "ms": 0.218465,
    "share": 3.957692993241665e-05,
    "count": 160,
    "families": {
     "elementwise_reduce": 0.218465
    },
    "top_kernels": [
     {
      "name": "void at::native::elementwise_kernel<(int)128, (int)2, void at::native::gpu_kernel_impl_noc",
      "ms": 0.218465
     }
    ]
   },
   {
    "unit": "vae_barrier/ep*_chunk*",
    "ms": 0.031967999999999996,
    "share": 5.791295155194175e-06,
    "count": 80,
    "families": {
     "elementwise_reduce": 0.031967999999999996
    },
    "top_kernels": [
     {
      "name": "void at::native::vectorized_elementwise_kernel<(int)4, at::native::FillFunctor<float>, std",
      "ms": 0.031967999999999996
     }
    ]
   },
   {
    "unit": "vae_barrier2/ep*_chunk*",
    "ms": 0.030496000000000002,
    "share": 5.5246289118118616e-06,
    "count": 80,
    "families": {
     "elementwise_reduce": 0.030496000000000002
    },
    "top_kernels": [
     {
      "name": "void at::native::vectorized_elementwise_kernel<(int)4, at::native::FillFunctor<float>, std",
      "ms": 0.030496000000000002
     }
    ]
   }
  ],
  "by_family": {
   "conv": 2459.6806620000016,
   "elementwise_reduce": 1925.1008430000006,
   "nccl": 799.0373229999996,
   "attention_core": 230.010368,
   "gather_index_cat": 96.67882699999998,
   "gemm": 9.500739000000001
  }
 },
 "self_attention_by_frame": [
  {
   "frame": 15,
   "commit": false,
   "family": "attention_core",
   "ms": 628.414104,
   "count": 1280
  },
  {
   "frame": 15,
   "commit": false,
   "family": "attention_prep",
   "ms": 1.436354,
   "count": 1280
  },
  {
   "frame": 15,
   "commit": false,
   "family": "elementwise_reduce",
   "ms": 1.662366,
   "count": 2560
  },
  {
   "frame": 15,
   "commit": false,
   "family": "gather_index_cat",
   "ms": 88.774712,
   "count": 6400
  },
  {
   "frame": 15,
   "commit": true,
   "family": "attention_core",
   "ms": 158.100714,
   "count": 320
  },
  {
   "frame": 15,
   "commit": true,
   "family": "attention_prep",
   "ms": 0.360576,
   "count": 320
  },
  {
   "frame": 15,
   "commit": true,
   "family": "elementwise_reduce",
   "ms": 0.421321,
   "count": 640
  },
  {
   "frame": 15,
   "commit": true,
   "family": "gather_index_cat",
   "ms": 22.255119,
   "count": 1600
  },
  {
   "frame": 18,
   "commit": false,
   "family": "attention_core",
   "ms": 627.02782,
   "count": 1280
  },
  {
   "frame": 18,
   "commit": false,
   "family": "attention_prep",
   "ms": 1.428322,
   "count": 1280
  },
  {
   "frame": 18,
   "commit": false,
   "family": "elementwise_reduce",
   "ms": 1.656727,
   "count": 2560
  },
  {
   "frame": 18,
   "commit": false,
   "family": "gather_index_cat",
   "ms": 88.588039,
   "count": 6400
  },
  {
   "frame": 18,
   "commit": true,
   "family": "attention_core",
   "ms": 158.024579,
   "count": 320
  },
  {
   "frame": 18,
   "commit": true,
   "family": "attention_prep",
   "ms": 0.361856,
   "count": 320
  },
  {
   "frame": 18,
   "commit": true,
   "family": "elementwise_reduce",
   "ms": 0.417389,
   "count": 640
  },
  {
   "frame": 18,
   "commit": true,
   "family": "gather_index_cat",
   "ms": 22.238939,
   "count": 1600
  },
  {
   "frame": 21,
   "commit": false,
   "family": "attention_core",
   "ms": 627.493756,
   "count": 1280
  },
  {
   "frame": 21,
   "commit": false,
   "family": "attention_prep",
   "ms": 1.429504,
   "count": 1280
  },
  {
   "frame": 21,
   "commit": false,
   "family": "elementwise_reduce",
   "ms": 1.659735,
   "count": 2560
  },
  {
   "frame": 21,
   "commit": false,
   "family": "gather_index_cat",
   "ms": 88.554739,
   "count": 6400
  },
  {
   "frame": 21,
   "commit": true,
   "family": "attention_core",
   "ms": 158.273996,
   "count": 320
  },
  {
   "frame": 21,
   "commit": true,
   "family": "attention_prep",
   "ms": 0.362368,
   "count": 320
  },
  {
   "frame": 21,
   "commit": true,
   "family": "elementwise_reduce",
   "ms": 0.420653,
   "count": 640
  },
  {
   "frame": 21,
   "commit": true,
   "family": "gather_index_cat",
   "ms": 22.262064,
   "count": 1600
  },
  {
   "frame": 24,
   "commit": false,
   "family": "attention_core",
   "ms": 629.156141,
   "count": 1280
  },
  {
   "frame": 24,
   "commit": false,
   "family": "attention_prep",
   "ms": 1.436067,
   "count": 1280
  },
  {
   "frame": 24,
   "commit": false,
   "family": "elementwise_reduce",
   "ms": 1.659954,
   "count": 2560
  },
  {
   "frame": 24,
   "commit": false,
   "family": "gather_index_cat",
   "ms": 88.76074899999999,
   "count": 6400
  },
  {
   "frame": 24,
   "commit": true,
   "family": "attention_core",
   "ms": 158.179174,
   "count": 320
  },
  {
   "frame": 24,
   "commit": true,
   "family": "attention_prep",
   "ms": 0.360579,
   "count": 320
  },
  {
   "frame": 24,
   "commit": true,
   "family": "elementwise_reduce",
   "ms": 0.41629499999999997,
   "count": 640
  },
  {
   "frame": 24,
   "commit": true,
   "family": "gather_index_cat",
   "ms": 22.265556999999998,
   "count": 1600
  }
 ]
}