01_matmul_add/64_bf16_cold_eager.json

  {
    "schemaVersion": 1,
    "deviceProperties": [
    {
      "id": 0, "name": "NVIDIA GeForce RTX 5080", "totalGlobalMem": 17094475776,
      "computeMajor": 12, "computeMinor": 0,
      "maxThreadsPerBlock": 1024, "maxThreadsPerMultiprocessor": 1536,
      "regsPerBlock": 65536, "warpSize": 32,
      "sharedMemPerBlock": 49152, "numSms": 84
    , "regsPerMultiprocessor": 65536, "sharedMemPerBlockOptin": 101376, "sharedMemPerMultiprocessor": 102400
    }
    ],
      "trace_id": "210C9921042A41179797025EBB4175BF",
      "cuda_driver_version": 13000,
      "cuda_runtime_version": 13020,
      "cupti_version": 130201,
  "displayTimeUnit": "ms",
  "baseTimeNanoseconds": 1775078550000000000,
  "traceEvents": [
  {
    "ph": "X", "cat": "user_annotation", "name": "ProfilerStep#2", "pid": 92830, "tid": 92830,
    "ts": 6097771625072.097, "dur": 432.492,
    "args": {
      "External id": 1,"Record function id": 0, "Ev Idx": 0
    }
  },
  {
    "ph": "X", "cat": "user_annotation", "name": "matmul_add", "pid": 92830, "tid": 92830,
    "ts": 6097771625094.149, "dur": 402.525,
    "args": {
      "External id": 2,"Record function id": 0, "Ev Idx": 1
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::matmul", "pid": 92830, "tid": 92830,
    "ts": 6097771625406.601, "dur": 53.873,
    "args": {
      "External id": 3,"Record function id": 0, "Sequence number": 0, "Fwd thread id": 0, "Ev Idx": 2
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 92830, "tid": 92830,
    "ts": 6097771625408.886, "dur": 48.913,
    "args": {
      "External id": 4,"Record function id": 0, "Sequence number": 0, "Fwd thread id": 0, "Ev Idx": 3
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::add", "pid": 92830, "tid": 92830,
    "ts": 6097771625466.927, "dur": 12.874,
    "args": {
      "External id": 5,"Record function id": 0, "Sequence number": 0, "Fwd thread id": 0, "Ev Idx": 4
    }
  },
  {
    "ph": "X", "cat": "user_annotation", "name": "ProfilerStep#3", "pid": 92830, "tid": 92830,
    "ts": 6097771625519.868, "dur": 39.285,
    "args": {
      "External id": 6,"Record function id": 0, "Ev Idx": 5
    }
  },
  {
    "ph": "X", "cat": "user_annotation", "name": "matmul_add", "pid": 92830, "tid": 92830,
    "ts": 6097771625529.787, "dur": 26.040,
    "args": {
      "External id": 7,"Record function id": 0, "Ev Idx": 6
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::matmul", "pid": 92830, "tid": 92830,
    "ts": 6097771625533.093, "dur": 11.903,
    "args": {
      "External id": 8,"Record function id": 0, "Sequence number": 0, "Fwd thread id": 0, "Ev Idx": 7
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 92830, "tid": 92830,
    "ts": 6097771625533.474, "dur": 11.161,
    "args": {
      "External id": 9,"Record function id": 0, "Sequence number": 0, "Fwd thread id": 0, "Ev Idx": 8
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::add", "pid": 92830, "tid": 92830,
    "ts": 6097771625545.958, "dur": 5.631,
    "args": {
      "External id": 10,"Record function id": 0, "Sequence number": 0, "Fwd thread id": 0, "Ev Idx": 9
    }
  },
  {
    "ph": "X", "cat": "user_annotation", "name": "ProfilerStep#4", "pid": 92830, "tid": 92830,
    "ts": 6097771625566.016, "dur": 25.649,
    "args": {
      "External id": 11,"Record function id": 0, "Ev Idx": 10
    }
  },
  {
    "ph": "X", "cat": "user_annotation", "name": "matmul_add", "pid": 92830, "tid": 92830,
    "ts": 6097771625570.224, "dur": 18.596,
    "args": {
      "External id": 12,"Record function id": 0, "Ev Idx": 11
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::matmul", "pid": 92830, "tid": 92830,
    "ts": 6097771625572.839, "dur": 8.296,
    "args": {
      "External id": 13,"Record function id": 0, "Sequence number": 0, "Fwd thread id": 0, "Ev Idx": 12
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 92830, "tid": 92830,
    "ts": 6097771625573.100, "dur": 7.805,
    "args": {
      "External id": 14,"Record function id": 0, "Sequence number": 0, "Fwd thread id": 0, "Ev Idx": 13
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::add", "pid": 92830, "tid": 92830,
    "ts": 6097771625581.977, "dur": 4.218,
    "args": {
      "External id": 15,"Record function id": 0, "Sequence number": 0, "Fwd thread id": 0, "Ev Idx": 14
    }
  },
  {
    "ph": "X", "cat": "kernel", "name": "void cutlass::Kernel2<cutlass_80_wmma_tensorop_bf16_s161616gemm_bf16_32x32_64x2_nn_align8>(cutlass_80_wmma_tensorop_bf16_s161616gemm_bf16_32x32_64x2_nn_align8::Params)", "pid": 0, "tid": 7,
    "ts": 6097771625757.309, "dur": 2.018,
    "args": {
            "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 16,      "registers per thread": 68,      "shared memory": 18944,      "blocks per SM": 0.047619,      "warps per SM": 0.190476,      "grid": [4, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 0
    }
  },
  {
    "ph": "f", "id": 16, "pid": 0, "tid": 7, "ts": 6097771625757.309,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void at::native::vectorized_elementwise_kernel<4, at::native::CUDAFunctor_add<c10::BFloat16>, std::array<char*, 3ul> >(int, at::native::CUDAFunctor_add<c10::BFloat16>, std::array<char*, 3ul>)", "pid": 0, "tid": 7,
    "ts": 6097771625761.216, "dur": 0.897,
    "args": {
            "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 26,      "registers per thread": 40,      "shared memory": 0,      "blocks per SM": 0.047619,      "warps per SM": 0.190476,      "grid": [4, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 0
    }
  },
  {
    "ph": "f", "id": 26, "pid": 0, "tid": 7, "ts": 6097771625761.216,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "cudaDeviceGetAttribute", "pid": 92830, "tid": 92830,
    "ts": 6097771625442.129, "dur": 0.611,
    "args": {
      "External id": 4,      "cbid": 200, "correlation": 43
    }
  },
  {
    "ph": "f", "id": 43, "pid": 92830, "tid": 92830, "ts": 6097771625442.129,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void cutlass::Kernel2<cutlass_80_wmma_tensorop_bf16_s161616gemm_bf16_32x32_64x2_nn_align8>(cutlass_80_wmma_tensorop_bf16_s161616gemm_bf16_32x32_64x2_nn_align8::Params)", "pid": 0, "tid": 7,
    "ts": 6097771625763.265, "dur": 1.569,
    "args": {
      "External id": 4,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 45,      "registers per thread": 68,      "shared memory": 18944,      "blocks per SM": 0.047619,      "warps per SM": 0.190476,      "grid": [4, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 0
    }
  },
  {
    "ph": "f", "id": 45, "pid": 0, "tid": 7, "ts": 6097771625763.265,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "cuda_driver", "name": "cuLaunchKernel", "pid": 92830, "tid": 92830,
    "ts": 6097771625444.494, "dur": 11.672,
    "args": {
      "External id": 4,      "cbid": 307, "correlation": 45
    }
  },
  {
    "ph": "s", "id": 45, "pid": 92830, "tid": 92830, "ts": 6097771625444.494,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void at::native::vectorized_elementwise_kernel<4, at::native::CUDAFunctor_add<c10::BFloat16>, std::array<char*, 3ul> >(int, at::native::CUDAFunctor_add<c10::BFloat16>, std::array<char*, 3ul>)", "pid": 0, "tid": 7,
    "ts": 6097771625767.716, "dur": 0.897,
    "args": {
      "External id": 5,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 55,      "registers per thread": 40,      "shared memory": 0,      "blocks per SM": 0.047619,      "warps per SM": 0.190476,      "grid": [4, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 0
    }
  },
  {
    "ph": "f", "id": 55, "pid": 0, "tid": 7, "ts": 6097771625767.716,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 92830, "tid": 92830,
    "ts": 6097771625474.641, "dur": 3.988,
    "args": {
      "External id": 5,      "cbid": 211, "correlation": 55
    }
  },
  {
    "ph": "s", "id": 55, "pid": 92830, "tid": 92830, "ts": 6097771625474.641,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "cudaDeviceGetAttribute", "pid": 92830, "tid": 92830,
    "ts": 6097771625540.447, "dur": 0.201,
    "args": {
      "External id": 9,      "cbid": 200, "correlation": 72
    }
  },
  {
    "ph": "f", "id": 72, "pid": 92830, "tid": 92830, "ts": 6097771625540.447,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void cutlass::Kernel2<cutlass_80_wmma_tensorop_bf16_s161616gemm_bf16_32x32_64x2_nn_align8>(cutlass_80_wmma_tensorop_bf16_s161616gemm_bf16_32x32_64x2_nn_align8::Params)", "pid": 0, "tid": 7,
    "ts": 6097771625769.413, "dur": 1.570,
    "args": {
      "External id": 9,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 74,      "registers per thread": 68,      "shared memory": 18944,      "blocks per SM": 0.047619,      "warps per SM": 0.190476,      "grid": [4, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 0
    }
  },
  {
    "ph": "f", "id": 74, "pid": 0, "tid": 7, "ts": 6097771625769.413,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "cuda_driver", "name": "cuLaunchKernel", "pid": 92830, "tid": 92830,
    "ts": 6097771625541.339, "dur": 2.625,
    "args": {
      "External id": 9,      "cbid": 307, "correlation": 74
    }
  },
  {
    "ph": "s", "id": 74, "pid": 92830, "tid": 92830, "ts": 6097771625541.339,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void at::native::vectorized_elementwise_kernel<4, at::native::CUDAFunctor_add<c10::BFloat16>, std::array<char*, 3ul> >(int, at::native::CUDAFunctor_add<c10::BFloat16>, std::array<char*, 3ul>)", "pid": 0, "tid": 7,
    "ts": 6097771625773.480, "dur": 0.897,
    "args": {
      "External id": 10,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 84,      "registers per thread": 40,      "shared memory": 0,      "blocks per SM": 0.047619,      "warps per SM": 0.190476,      "grid": [4, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 0
    }
  },
  {
    "ph": "f", "id": 84, "pid": 0, "tid": 7, "ts": 6097771625773.480,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 92830, "tid": 92830,
    "ts": 6097771625548.793, "dur": 2.054,
    "args": {
      "External id": 10,      "cbid": 211, "correlation": 84
    }
  },
  {
    "ph": "s", "id": 84, "pid": 92830, "tid": 92830, "ts": 6097771625548.793,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "cudaDeviceGetAttribute", "pid": 92830, "tid": 92830,
    "ts": 6097771625577.909, "dur": 0.170,
    "args": {
      "External id": 14,      "cbid": 200, "correlation": 101
    }
  },
  {
    "ph": "f", "id": 101, "pid": 92830, "tid": 92830, "ts": 6097771625577.909,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void cutlass::Kernel2<cutlass_80_wmma_tensorop_bf16_s161616gemm_bf16_32x32_64x2_nn_align8>(cutlass_80_wmma_tensorop_bf16_s161616gemm_bf16_32x32_64x2_nn_align8::Params)", "pid": 0, "tid": 7,
    "ts": 6097771625775.530, "dur": 1.569,
    "args": {
      "External id": 14,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 103,      "registers per thread": 68,      "shared memory": 18944,      "blocks per SM": 0.047619,      "warps per SM": 0.190476,      "grid": [4, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 0
    }
  },
  {
    "ph": "f", "id": 103, "pid": 0, "tid": 7, "ts": 6097771625775.530,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "cuda_driver", "name": "cuLaunchKernel", "pid": 92830, "tid": 92830,
    "ts": 6097771625578.500, "dur": 1.884,
    "args": {
      "External id": 14,      "cbid": 307, "correlation": 103
    }
  },
  {
    "ph": "s", "id": 103, "pid": 92830, "tid": 92830, "ts": 6097771625578.500,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void at::native::vectorized_elementwise_kernel<4, at::native::CUDAFunctor_add<c10::BFloat16>, std::array<char*, 3ul> >(int, at::native::CUDAFunctor_add<c10::BFloat16>, std::array<char*, 3ul>)", "pid": 0, "tid": 7,
    "ts": 6097771625779.660, "dur": 0.865,
    "args": {
      "External id": 15,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 113,      "registers per thread": 40,      "shared memory": 0,      "blocks per SM": 0.047619,      "warps per SM": 0.190476,      "grid": [4, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 0
    }
  },
  {
    "ph": "f", "id": 113, "pid": 0, "tid": 7, "ts": 6097771625779.660,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 92830, "tid": 92830,
    "ts": 6097771625583.920, "dur": 1.754,
    "args": {
      "External id": 15,      "cbid": 211, "correlation": 113
    }
  },
  {
    "ph": "s", "id": 113, "pid": 92830, "tid": 92830, "ts": 6097771625583.920,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "cudaDeviceSynchronize", "pid": 92830, "tid": 92830,
    "ts": 6097771625630.439, "dur": 139.467,
    "args": {
            "cbid": 165, "correlation": 119
    }
  },
  {
    "ph": "s", "id": 119, "pid": 92830, "tid": 92830, "ts": 6097771625630.439,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "gpu_user_annotation", "name": "matmul_add", "pid": 0, "tid": 7,
    "ts": 6097771625775.529, "dur": 4.997,
    "args": {
      "External id": 12
    }
  },
  {
    "ph": "X", "cat": "gpu_user_annotation", "name": "matmul_add", "pid": 0, "tid": 7,
    "ts": 6097771625769.412, "dur": 4.966,
    "args": {
      "External id": 7
    }
  },
  {
    "ph": "X", "cat": "gpu_user_annotation", "name": "matmul_add", "pid": 0, "tid": 7,
    "ts": 6097771625763.264, "dur": 5.350,
    "args": {
      "External id": 2
    }
  },
  {
    "name": "process_name", "ph": "M", "ts": 6097771624940.547, "pid": 92830, "tid": 0,
    "args": {
      "name": "python"
    }
  },
  {
    "name": "process_labels", "ph": "M", "ts": 6097771624940.547, "pid": 92830, "tid": 0,
    "args": {
      "labels": "CPU"
    }
  },
  {
    "name": "process_sort_index", "ph": "M", "ts": 6097771624940.547, "pid": 92830, "tid": 0,
    "args": {
      "sort_index": 92830
    }
  },
  {
    "name": "process_name", "ph": "M", "ts": 6097771624940.547, "pid": 0, "tid": 0,
    "args": {
      "name": "python"
    }
  },
  {
    "name": "process_labels", "ph": "M", "ts": 6097771624940.547, "pid": 0, "tid": 0,
    "args": {
      "labels": "GPU 0"
    }
  },
  {
    "name": "process_sort_index", "ph": "M", "ts": 6097771624940.547, "pid": 0, "tid": 0,
    "args": {
      "sort_index": 5000000
    }
  },
  {
    "name": "process_name", "ph": "M", "ts": 6097771624940.547, "pid": 1, "tid": 0,
    "args": {
      "name": "python"
    }
  },
  {
    "name": "process_labels", "ph": "M", "ts": 6097771624940.547, "pid": 1, "tid": 0,
    "args": {
      "labels": "GPU 1"
    }
  },
  {
    "name": "process_sort_index", "ph": "M", "ts": 6097771624940.547, "pid": 1, "tid": 0,
    "args": {
      "sort_index": 5000001
    }
  },
  {
    "name": "process_name", "ph": "M", "ts": 6097771624940.547, "pid": 2, "tid": 0,
    "args": {
      "name": "python"
    }
  },
  {
    "name": "process_labels", "ph": "M", "ts": 6097771624940.547, "pid": 2, "tid": 0,
    "args": {
      "labels": "GPU 2"
    }
  },
  {
    "name": "process_sort_index", "ph": "M", "ts": 6097771624940.547, "pid": 2, "tid": 0,
    "args": {
      "sort_index": 5000002
    }
  },
  {
    "name": "process_name", "ph": "M", "ts": 6097771624940.547, "pid": 3, "tid": 0,
    "args": {
      "name": "python"
    }
  },
  {
    "name": "process_labels", "ph": "M", "ts": 6097771624940.547, "pid": 3, "tid": 0,
    "args": {
      "labels": "GPU 3"
    }
  },
  {
    "name": "process_sort_index", "ph": "M", "ts": 6097771624940.547, "pid": 3, "tid": 0,
    "args": {
      "sort_index": 5000003
    }
  },
  {
    "name": "process_name", "ph": "M", "ts": 6097771624940.547, "pid": 4, "tid": 0,
    "args": {
      "name": "python"
    }
  },
  {
    "name": "process_labels", "ph": "M", "ts": 6097771624940.547, "pid": 4, "tid": 0,
    "args": {
      "labels": "GPU 4"
    }
  },
  {
    "name": "process_sort_index", "ph": "M", "ts": 6097771624940.547, "pid": 4, "tid": 0,
    "args": {
      "sort_index": 5000004
    }
  },
  {
    "name": "process_name", "ph": "M", "ts": 6097771624940.547, "pid": 5, "tid": 0,
    "args": {
      "name": "python"
    }
  },
  {
    "name": "process_labels", "ph": "M", "ts": 6097771624940.547, "pid": 5, "tid": 0,
    "args": {
      "labels": "GPU 5"
    }
  },
  {
    "name": "process_sort_index", "ph": "M", "ts": 6097771624940.547, "pid": 5, "tid": 0,
    "args": {
      "sort_index": 5000005
    }
  },
  {
    "name": "process_name", "ph": "M", "ts": 6097771624940.547, "pid": 6, "tid": 0,
    "args": {
      "name": "python"
    }
  },
  {
    "name": "process_labels", "ph": "M", "ts": 6097771624940.547, "pid": 6, "tid": 0,
    "args": {
      "labels": "GPU 6"
    }
  },
  {
    "name": "process_sort_index", "ph": "M", "ts": 6097771624940.547, "pid": 6, "tid": 0,
    "args": {
      "sort_index": 5000006
    }
  },
  {
    "name": "process_name", "ph": "M", "ts": 6097771624940.547, "pid": 7, "tid": 0,
    "args": {
      "name": "python"
    }
  },
  {
    "name": "process_labels", "ph": "M", "ts": 6097771624940.547, "pid": 7, "tid": 0,
    "args": {
      "labels": "GPU 7"
    }
  },
  {
    "name": "process_sort_index", "ph": "M", "ts": 6097771624940.547, "pid": 7, "tid": 0,
    "args": {
      "sort_index": 5000007
    }
  },
  {
    "name": "process_name", "ph": "M", "ts": 6097771624940.547, "pid": 8, "tid": 0,
    "args": {
      "name": "python"
    }
  },
  {
    "name": "process_labels", "ph": "M", "ts": 6097771624940.547, "pid": 8, "tid": 0,
    "args": {
      "labels": "GPU 8"
    }
  },
  {
    "name": "process_sort_index", "ph": "M", "ts": 6097771624940.547, "pid": 8, "tid": 0,
    "args": {
      "sort_index": 5000008
    }
  },
  {
    "name": "process_name", "ph": "M", "ts": 6097771624940.547, "pid": 9, "tid": 0,
    "args": {
      "name": "python"
    }
  },
  {
    "name": "process_labels", "ph": "M", "ts": 6097771624940.547, "pid": 9, "tid": 0,
    "args": {
      "labels": "GPU 9"
    }
  },
  {
    "name": "process_sort_index", "ph": "M", "ts": 6097771624940.547, "pid": 9, "tid": 0,
    "args": {
      "sort_index": 5000009
    }
  },
  {
    "name": "process_name", "ph": "M", "ts": 6097771624940.547, "pid": 10, "tid": 0,
    "args": {
      "name": "python"
    }
  },
  {
    "name": "process_labels", "ph": "M", "ts": 6097771624940.547, "pid": 10, "tid": 0,
    "args": {
      "labels": "GPU 10"
    }
  },
  {
    "name": "process_sort_index", "ph": "M", "ts": 6097771624940.547, "pid": 10, "tid": 0,
    "args": {
      "sort_index": 5000010
    }
  },
  {
    "name": "process_name", "ph": "M", "ts": 6097771624940.547, "pid": 11, "tid": 0,
    "args": {
      "name": "python"
    }
  },
  {
    "name": "process_labels", "ph": "M", "ts": 6097771624940.547, "pid": 11, "tid": 0,
    "args": {
      "labels": "GPU 11"
    }
  },
  {
    "name": "process_sort_index", "ph": "M", "ts": 6097771624940.547, "pid": 11, "tid": 0,
    "args": {
      "sort_index": 5000011
    }
  },
  {
    "name": "process_name", "ph": "M", "ts": 6097771624940.547, "pid": 12, "tid": 0,
    "args": {
      "name": "python"
    }
  },
  {
    "name": "process_labels", "ph": "M", "ts": 6097771624940.547, "pid": 12, "tid": 0,
    "args": {
      "labels": "GPU 12"
    }
  },
  {
    "name": "process_sort_index", "ph": "M", "ts": 6097771624940.547, "pid": 12, "tid": 0,
    "args": {
      "sort_index": 5000012
    }
  },
  {
    "name": "process_name", "ph": "M", "ts": 6097771624940.547, "pid": 13, "tid": 0,
    "args": {
      "name": "python"
    }
  },
  {
    "name": "process_labels", "ph": "M", "ts": 6097771624940.547, "pid": 13, "tid": 0,
    "args": {
      "labels": "GPU 13"
    }
  },
  {
    "name": "process_sort_index", "ph": "M", "ts": 6097771624940.547, "pid": 13, "tid": 0,
    "args": {
      "sort_index": 5000013
    }
  },
  {
    "name": "process_name", "ph": "M", "ts": 6097771624940.547, "pid": 14, "tid": 0,
    "args": {
      "name": "python"
    }
  },
  {
    "name": "process_labels", "ph": "M", "ts": 6097771624940.547, "pid": 14, "tid": 0,
    "args": {
      "labels": "GPU 14"
    }
  },
  {
    "name": "process_sort_index", "ph": "M", "ts": 6097771624940.547, "pid": 14, "tid": 0,
    "args": {
      "sort_index": 5000014
    }
  },
  {
    "name": "process_name", "ph": "M", "ts": 6097771624940.547, "pid": 15, "tid": 0,
    "args": {
      "name": "python"
    }
  },
  {
    "name": "process_labels", "ph": "M", "ts": 6097771624940.547, "pid": 15, "tid": 0,
    "args": {
      "labels": "GPU 15"
    }
  },
  {
    "name": "process_sort_index", "ph": "M", "ts": 6097771624940.547, "pid": 15, "tid": 0,
    "args": {
      "sort_index": 5000015
    }
  },
  {
    "name": "thread_name", "ph": "M", "ts": 6097771624940.547, "pid": 0, "tid": 7,
    "args": {
      "name": "stream 7 "
    }
  },
  {
    "name": "thread_sort_index", "ph": "M", "ts": 6097771624940.547, "pid": 0, "tid": 7,
    "args": {
      "sort_index": 7
    }
  },
  {
    "name": "thread_name", "ph": "M", "ts": 6097771624940.547, "pid": 92830, "tid": 92830,
    "args": {
      "name": "thread 92830 (python)"
    }
  },
  {
    "name": "thread_sort_index", "ph": "M", "ts": 6097771624940.547, "pid": 92830, "tid": 92830,
    "args": {
      "sort_index": 92830
    }
  },
  {
    "name": "thread_name", "ph": "M", "ts": 6097771624940.547, "pid": 92830, "tid": 92830,
    "args": {
      "name": "thread 92830 (python)"
    }
  },
  {
    "name": "thread_sort_index", "ph": "M", "ts": 6097771624940.547, "pid": 92830, "tid": 92830,
    "args": {
      "sort_index": 92830
    }
  },
  {
    "ph": "X", "cat": "Trace", "ts": 6097771624902.505, "dur": 874.998,
    "pid": "Spans", "tid": "PyTorch Profiler",
    "name": "PyTorch Profiler (0)",
    "args": {
      "Op count": 0
    }
  },
  {
    "name": "process_sort_index", "ph": "M", "ts": 6097771624902.505,
    "pid": "Spans", "tid": 0,
    "args": {
      "sort_index": 536870912
    }
  },
  {
    "name": "Iteration Start: PyTorch Profiler", "ph": "i", "s": "g",
    "pid": "Traces", "tid": "Trace PyTorch Profiler", "ts": 6097771624902.505
  },
  {
    "name": "Record Window End", "ph": "i", "s": "g",
    "pid": "", "tid": "", "ts": 6097771626080.171
  }
  ],
  "traceName": "./traces/01_matmul_add/64_bf16_cold_eager.json"
}