{
 "generated_at": "2026-09-10T16:16:56+0800",
 "tag": "480x832",
 "steady_gpu_ms": 12224.388316999997,
 "steady_share_with_shape": 0.947153237426072,
 "steady_share_by_ncu_status": {
  "exact_card": 0.6825559087808544,
  "no_card": 0.31744409121914574
 },
 "by_role": {
  "dit": {
   "exact_card": 0.7458144245407573,
   "no_card": 0.2541855754592427
  },
  "vae": {
   "no_card": 0.5931981848082637,
   "exact_card": 0.40680181519173625
  }
 },
 "uncovered_top": [
  {
   "role": "dit",
   "unit": "_c10d_functional.all_to_all_single",
   "kernel": "ncclDevKernel_SendRecv(ncclDevKernelArgsStorage<(unsigned lo",
   "steady_ms": 1073.5951300000002
  },
  {
   "role": "vae",
   "unit": "ncclGroupEnd",
   "kernel": "ncclDevKernel_SendRecv(ncclDevKernelArgsStorage<(unsigned lo",
   "steady_ms": 308.8656929999999
  },
  {
   "role": "dit",
   "unit": "vllm_omni.ar_diffusion_paged_write_attn",
   "kernel": "void at::native::vectorized_gather_kernel<(int)16, long>(cha",
   "steady_ms": 248.05808600000003
  },
  {
   "role": "dit",
   "unit": "vllm_omni.ar_diffusion_paged_write_attn",
   "kernel": "void at::native::index_elementwise_kernel<(int)128, (int)4, ",
   "steady_ms": 191.032256
  },
  {
   "role": "dit",
   "unit": "fa3_fwd._flash_attn_forward",
   "kernel": "void cutlass::device_kernel<flash::enable_sm90_or_later<flas",
   "steady_ms": 92.77299099999999
  },
  {
   "role": "vae",
   "unit": "vae_frame/*",
   "kernel": "fmha_cutlassF_f32_aligned_32x128_gmem_sm80(PyTorchMemEffAtte",
   "steady_ms": 92.169048
  },
  {
   "role": "dit",
   "unit": "triton_red_fused__to_copy_add_mul_native_layer_norm_split_squeeze_view_9",
   "kernel": "triton_red_fused__to_copy_add_mul_native_layer_norm_split_sq",
   "steady_ms": 77.775204
  },
  {
   "role": "dit",
   "unit": "triton_red_fused__to_copy_add_mul_native_layer_norm_split_squeeze_view_12",
   "kernel": "triton_red_fused__to_copy_add_mul_native_layer_norm_split_sq",
   "steady_ms": 73.54639600000002
  },
  {
   "role": "dit",
   "unit": "triton_red_fused__to_copy_add_mul_native_layer_norm_split_squeeze_1",
   "kernel": "triton_red_fused__to_copy_add_mul_native_layer_norm_split_sq",
   "steady_ms": 71.26247900000001
  },
  {
   "role": "dit",
   "unit": "triton_poi_fused_gelu_view_13",
   "kernel": "triton_poi_fused_gelu_view_13",
   "steady_ms": 71.048524
  },
  {
   "role": "dit",
   "unit": "triton_poi_fused_clone_transpose_view_8",
   "kernel": "triton_poi_fused_clone_transpose_view_8",
   "steady_ms": 64.162941
  },
  {
   "role": "dit",
   "unit": "rotary_kernel_0",
   "kernel": "rotary_kernel",
   "steady_ms": 62.080633
  },
  {
   "role": "dit",
   "unit": "triton_poi_fused_all_to_all_single_clone_split_with_sizes_stack_transpose_view_4",
   "kernel": "triton_poi_fused_all_to_all_single_clone_split_with_sizes_st",
   "steady_ms": 58.59643500000001
  },
  {
   "role": "dit",
   "unit": "triton_poi_fused__to_copy_add_mul_split_squeeze_view_14",
   "kernel": "triton_poi_fused__to_copy_add_mul_split_squeeze_view_14",
   "steady_ms": 49.27491000000001
  },
  {
   "role": "vae",
   "unit": "vae_frame/*",
   "kernel": "void at::native::elementwise_kernel<(int)128, (int)2, void a",
   "steady_ms": 43.53613899999998
  },
  {
   "role": "dit",
   "unit": "triton_poi_fused_all_to_all_single_clone_transpose_unsqueeze_view_5",
   "kernel": "triton_poi_fused_all_to_all_single_clone_transpose_unsqueeze",
   "steady_ms": 42.824946
  },
  {
   "role": "dit",
   "unit": "triton_poi_fused__to_copy_all_to_all_single_clone_div_mul_sqrt_transpose_view_11",
   "kernel": "triton_poi_fused__to_copy_all_to_all_single_clone_div_mul_sq",
   "steady_ms": 34.926145
  },
  {
   "role": "dit",
   "unit": "triton_red_fused__to_copy_div_mul_pow_split_with_sizes_sqrt_sum_view_3",
   "kernel": "triton_red_fused__to_copy_div_mul_pow_split_with_sizes_sqrt_",
   "steady_ms": 34.19756999999999
  },
  {
   "role": "dit",
   "unit": "triton_poi_fused_silu_view_6",
   "kernel": "triton_poi_fused_silu_view_6",
   "steady_ms": 33.03871
  },
  {
   "role": "vae",
   "unit": "vae/up_blocks.3.resnets.0.conv1/WanDistCausalConv3d",
   "kernel": "void at::native::elementwise_kernel<(int)128, (int)2, void a",
   "steady_ms": 32.809633
  },
  {
   "role": "vae",
   "unit": "vae/up_blocks.3.resnets.0.conv2/WanDistCausalConv3d",
   "kernel": "void at::native::elementwise_kernel<(int)128, (int)2, void a",
   "steady_ms": 32.430462999999996
  },
  {
   "role": "vae",
   "unit": "vae/up_blocks.3.resnets.1.conv1/WanDistCausalConv3d",
   "kernel": "void at::native::elementwise_kernel<(int)128, (int)2, void a",
   "steady_ms": 32.421701
  },
  {
   "role": "vae",
   "unit": "vae/up_blocks.3.resnets.2.conv1/WanDistCausalConv3d",
   "kernel": "void at::native::elementwise_kernel<(int)128, (int)2, void a",
   "steady_ms": 32.40233800000001
  },
  {
   "role": "vae",
   "unit": "vae/up_blocks.3.resnets.1.conv2/WanDistCausalConv3d",
   "kernel": "void at::native::elementwise_kernel<(int)128, (int)2, void a",
   "steady_ms": 32.395168
  },
  {
   "role": "vae",
   "unit": "vae/up_blocks.3.resnets.2.conv2/WanDistCausalConv3d",
   "kernel": "void at::native::elementwise_kernel<(int)128, (int)2, void a",
   "steady_ms": 32.372189
  },
  {
   "role": "vae",
   "unit": "vae/up_blocks.2.upsamplers.0.resample.1/WanDistConv2d",
   "kernel": "sm90_xmma_fprop_implicit_gemm_f32f32_tf32f32_f32_nhwckrsc_nh",
   "steady_ms": 32.085662
  },
  {
   "role": "dit",
   "unit": "triton_red_fused__to_copy_div_mul_pow_split_with_sizes_sqrt_sum_view_2",
   "kernel": "triton_red_fused__to_copy_div_mul_pow_split_with_sizes_sqrt_",
   "steady_ms": 31.027860999999994
  },
  {
   "role": "dit",
   "unit": "triton_poi_fused_add_view_7",
   "kernel": "triton_poi_fused_add_view_7",
   "steady_ms": 29.508433
  },
  {
   "role": "vae",
   "unit": "ncclAllReduce",
   "kernel": "ncclDevKernel_AllReduce_Sum_f32_RING_LL(ncclDevKernelArgsSto",
   "steady_ms": 29.287644
  },
  {
   "role": "vae",
   "unit": "vae/conv_out/WanDistCausalConv3d",
   "kernel": "void at::native::elementwise_kernel<(int)128, (int)2, void a",
   "steady_ms": 27.101589000000004
  }
 ],
 "positions_with_shape": 42,
 "vae_modules_with_shape": 39,
 "seconds": 0.28046250343322754,
 "scope": "Shapes come from the eager full-dispatch run of the same frozen source and configuration (not the traced run); compiled fused kernels get inferred activation shapes. NCU cards are isolated replays; exact_card means identical operator shape, representative_only means same operator family with a different KV length."
}