01_matmul_add/4096_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": "E5F5CD0DCA3A43EF8CF2729B8A4FE50F",
      "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": 93046, "tid": 93046,
    "ts": 6098062929278.509, "dur": 559.695,
    "args": {
      "External id": 1,"Record function id": 0, "Ev Idx": 0
    }
  },
  {
    "ph": "X", "cat": "user_annotation", "name": "matmul_add", "pid": 93046, "tid": 93046,
    "ts": 6098062929335.047, "dur": 495.392,
    "args": {
      "External id": 2,"Record function id": 0, "Ev Idx": 1
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::matmul", "pid": 93046, "tid": 93046,
    "ts": 6098062929656.777, "dur": 84.942,
    "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": 93046, "tid": 93046,
    "ts": 6098062929677.647, "dur": 61.087,
    "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": 93046, "tid": 93046,
    "ts": 6098062929781.235, "dur": 29.115,
    "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": 93046, "tid": 93046,
    "ts": 6098062929853.433, "dur": 56.037,
    "args": {
      "External id": 6,"Record function id": 0, "Ev Idx": 5
    }
  },
  {
    "ph": "X", "cat": "user_annotation", "name": "matmul_add", "pid": 93046, "tid": 93046,
    "ts": 6098062929862.921, "dur": 43.423,
    "args": {
      "External id": 7,"Record function id": 0, "Ev Idx": 6
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::matmul", "pid": 93046, "tid": 93046,
    "ts": 6098062929867.410, "dur": 25.659,
    "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": 93046, "tid": 93046,
    "ts": 6098062929868.231, "dur": 24.206,
    "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": 93046, "tid": 93046,
    "ts": 6098062929894.932, "dur": 6.783,
    "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": 93046, "tid": 93046,
    "ts": 6098062929916.894, "dur": 26.411,
    "args": {
      "External id": 11,"Record function id": 0, "Ev Idx": 10
    }
  },
  {
    "ph": "X", "cat": "user_annotation", "name": "matmul_add", "pid": 93046, "tid": 93046,
    "ts": 6098062929921.423, "dur": 19.427,
    "args": {
      "External id": 12,"Record function id": 0, "Ev Idx": 11
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::matmul", "pid": 93046, "tid": 93046,
    "ts": 6098062929923.817, "dur": 9.499,
    "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": 93046, "tid": 93046,
    "ts": 6098062929924.128, "dur": 8.937,
    "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": 93046, "tid": 93046,
    "ts": 6098062929934.027, "dur": 4.248,
    "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_tensorop_bf16_s16816gemm_relu_bf16_128x64_64x3_nn_align8>(cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x64_64x3_nn_align8::Params)", "pid": 0, "tid": 7,
    "ts": 6098062930358.292, "dur": 1241.158,
    "args": {
            "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 17,      "registers per thread": 154,      "shared memory": 73728,      "blocks per SM": 24.380953,      "warps per SM": 97.523811,      "grid": [512, 4, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 0
    }
  },
  {
    "ph": "f", "id": 17, "pid": 0, "tid": 7, "ts": 6098062930358.292,
    "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": 6098062931600.314, "dur": 69.432,
    "args": {
            "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 27,      "registers per thread": 40,      "shared memory": 0,      "blocks per SM": 195.047623,      "warps per SM": 780.190491,      "grid": [16384, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 100
    }
  },
  {
    "ph": "f", "id": 27, "pid": 0, "tid": 7, "ts": 6098062931600.314,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "cudaDeviceGetAttribute", "pid": 93046, "tid": 93046,
    "ts": 6098062929721.350, "dur": 0.722,
    "args": {
      "External id": 4,      "cbid": 200, "correlation": 44
    }
  },
  {
    "ph": "f", "id": 44, "pid": 93046, "tid": 93046, "ts": 6098062929721.350,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void cutlass::Kernel2<cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x64_64x3_nn_align8>(cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x64_64x3_nn_align8::Params)", "pid": 0, "tid": 7,
    "ts": 6098062931670.929, "dur": 1242.726,
    "args": {
      "External id": 4,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 47,      "registers per thread": 154,      "shared memory": 73728,      "blocks per SM": 24.380953,      "warps per SM": 97.523811,      "grid": [512, 4, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 0
    }
  },
  {
    "ph": "f", "id": 47, "pid": 0, "tid": 7, "ts": 6098062931670.929,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "cuda_driver", "name": "cuLaunchKernel", "pid": 93046, "tid": 93046,
    "ts": 6098062929724.546, "dur": 12.574,
    "args": {
      "External id": 4,      "cbid": 307, "correlation": 47
    }
  },
  {
    "ph": "s", "id": 47, "pid": 93046, "tid": 93046, "ts": 6098062929724.546,
    "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": 6098062933067.844, "dur": 75.382,
    "args": {
      "External id": 5,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 57,      "registers per thread": 40,      "shared memory": 0,      "blocks per SM": 195.047623,      "warps per SM": 780.190491,      "grid": [16384, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 100
    }
  },
  {
    "ph": "f", "id": 57, "pid": 0, "tid": 7, "ts": 6098062933067.844,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 93046, "tid": 93046,
    "ts": 6098062929797.376, "dur": 11.462,
    "args": {
      "External id": 5,      "cbid": 211, "correlation": 57
    }
  },
  {
    "ph": "s", "id": 57, "pid": 93046, "tid": 93046, "ts": 6098062929797.376,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "cudaDeviceGetAttribute", "pid": 93046, "tid": 93046,
    "ts": 6098062929885.534, "dur": 0.501,
    "args": {
      "External id": 9,      "cbid": 200, "correlation": 74
    }
  },
  {
    "ph": "f", "id": 74, "pid": 93046, "tid": 93046, "ts": 6098062929885.534,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void cutlass::Kernel2<cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x64_64x3_nn_align8>(cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x64_64x3_nn_align8::Params)", "pid": 0, "tid": 7,
    "ts": 6098062933145.306, "dur": 2156.884,
    "args": {
      "External id": 9,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 77,      "registers per thread": 154,      "shared memory": 73728,      "blocks per SM": 24.380953,      "warps per SM": 97.523811,      "grid": [512, 4, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 0
    }
  },
  {
    "ph": "f", "id": 77, "pid": 0, "tid": 7, "ts": 6098062933145.306,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "cuda_driver", "name": "cuLaunchKernel", "pid": 93046, "tid": 93046,
    "ts": 6098062929887.698, "dur": 3.607,
    "args": {
      "External id": 9,      "cbid": 307, "correlation": 77
    }
  },
  {
    "ph": "s", "id": 77, "pid": 93046, "tid": 93046, "ts": 6098062929887.698,
    "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": 6098062935303.630, "dur": 85.654,
    "args": {
      "External id": 10,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 87,      "registers per thread": 40,      "shared memory": 0,      "blocks per SM": 195.047623,      "warps per SM": 780.190491,      "grid": [16384, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 100
    }
  },
  {
    "ph": "f", "id": 87, "pid": 0, "tid": 7, "ts": 6098062935303.630,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 93046, "tid": 93046,
    "ts": 6098062929898.098, "dur": 2.876,
    "args": {
      "External id": 10,      "cbid": 211, "correlation": 87
    }
  },
  {
    "ph": "s", "id": 87, "pid": 93046, "tid": 93046, "ts": 6098062929898.098,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "cudaDeviceGetAttribute", "pid": 93046, "tid": 93046,
    "ts": 6098062929929.618, "dur": 0.181,
    "args": {
      "External id": 14,      "cbid": 200, "correlation": 104
    }
  },
  {
    "ph": "f", "id": 104, "pid": 93046, "tid": 93046, "ts": 6098062929929.618,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void cutlass::Kernel2<cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x64_64x3_nn_align8>(cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x64_64x3_nn_align8::Params)", "pid": 0, "tid": 7,
    "ts": 6098062935391.683, "dur": 1241.062,
    "args": {
      "External id": 14,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 107,      "registers per thread": 154,      "shared memory": 73728,      "blocks per SM": 24.380953,      "warps per SM": 97.523811,      "grid": [512, 4, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 0
    }
  },
  {
    "ph": "f", "id": 107, "pid": 0, "tid": 7, "ts": 6098062935391.683,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "cuda_driver", "name": "cuLaunchKernel", "pid": 93046, "tid": 93046,
    "ts": 6098062929930.440, "dur": 2.084,
    "args": {
      "External id": 14,      "cbid": 307, "correlation": 107
    }
  },
  {
    "ph": "s", "id": 107, "pid": 93046, "tid": 93046, "ts": 6098062929930.440,
    "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": 6098062936640.104, "dur": 69.432,
    "args": {
      "External id": 15,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 117,      "registers per thread": 40,      "shared memory": 0,      "blocks per SM": 195.047623,      "warps per SM": 780.190491,      "grid": [16384, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 100
    }
  },
  {
    "ph": "f", "id": 117, "pid": 0, "tid": 7, "ts": 6098062936640.104,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 93046, "tid": 93046,
    "ts": 6098062929936.061, "dur": 1.683,
    "args": {
      "External id": 15,      "cbid": 211, "correlation": 117
    }
  },
  {
    "ph": "s", "id": 117, "pid": 93046, "tid": 93046, "ts": 6098062929936.061,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "cudaDeviceSynchronize", "pid": 93046, "tid": 93046,
    "ts": 6098062930223.079, "dur": 6479.526,
    "args": {
            "cbid": 165, "correlation": 123
    }
  },
  {
    "ph": "s", "id": 123, "pid": 93046, "tid": 93046, "ts": 6098062930223.079,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "gpu_user_annotation", "name": "matmul_add", "pid": 0, "tid": 7,
    "ts": 6098062935391.682, "dur": 1317.855,
    "args": {
      "External id": 12
    }
  },
  {
    "ph": "X", "cat": "gpu_user_annotation", "name": "matmul_add", "pid": 0, "tid": 7,
    "ts": 6098062933145.305, "dur": 2243.980,
    "args": {
      "External id": 7
    }
  },
  {
    "ph": "X", "cat": "gpu_user_annotation", "name": "matmul_add", "pid": 0, "tid": 7,
    "ts": 6098062931670.928, "dur": 1472.299,
    "args": {
      "External id": 2
    }
  },
  {
    "name": "process_name", "ph": "M", "ts": 6098062929083.366, "pid": 93046, "tid": 0,
    "args": {
      "name": "python"
    }
  },
  {
    "name": "process_labels", "ph": "M", "ts": 6098062929083.366, "pid": 93046, "tid": 0,
    "args": {
      "labels": "CPU"
    }
  },
  {
    "name": "process_sort_index", "ph": "M", "ts": 6098062929083.366, "pid": 93046, "tid": 0,
    "args": {
      "sort_index": 93046
    }
  },
  {
    "name": "process_name", "ph": "M", "ts": 6098062929083.366, "pid": 0, "tid": 0,
    "args": {
      "name": "python"
    }
  },
  {
    "name": "process_labels", "ph": "M", "ts": 6098062929083.366, "pid": 0, "tid": 0,
    "args": {
      "labels": "GPU 0"
    }
  },
  {
    "name": "process_sort_index", "ph": "M", "ts": 6098062929083.366, "pid": 0, "tid": 0,
    "args": {
      "sort_index": 5000000
    }
  },
  {
    "name": "process_name", "ph": "M", "ts": 6098062929083.366, "pid": 1, "tid": 0,
    "args": {
      "name": "python"
    }
  },
  {
    "name": "process_labels", "ph": "M", "ts": 6098062929083.366, "pid": 1, "tid": 0,
    "args": {
      "labels": "GPU 1"
    }
  },
  {
    "name": "process_sort_index", "ph": "M", "ts": 6098062929083.366, "pid": 1, "tid": 0,
    "args": {
      "sort_index": 5000001
    }
  },
  {
    "name": "process_name", "ph": "M", "ts": 6098062929083.366, "pid": 2, "tid": 0,
    "args": {
      "name": "python"
    }
  },
  {
    "name": "process_labels", "ph": "M", "ts": 6098062929083.366, "pid": 2, "tid": 0,
    "args": {
      "labels": "GPU 2"
    }
  },
  {
    "name": "process_sort_index", "ph": "M", "ts": 6098062929083.366, "pid": 2, "tid": 0,
    "args": {
      "sort_index": 5000002
    }
  },
  {
    "name": "process_name", "ph": "M", "ts": 6098062929083.366, "pid": 3, "tid": 0,
    "args": {
      "name": "python"
    }
  },
  {
    "name": "process_labels", "ph": "M", "ts": 6098062929083.366, "pid": 3, "tid": 0,
    "args": {
      "labels": "GPU 3"
    }
  },
  {
    "name": "process_sort_index", "ph": "M", "ts": 6098062929083.366, "pid": 3, "tid": 0,
    "args": {
      "sort_index": 5000003
    }
  },
  {
    "name": "process_name", "ph": "M", "ts": 6098062929083.366, "pid": 4, "tid": 0,
    "args": {
      "name": "python"
    }
  },
  {
    "name": "process_labels", "ph": "M", "ts": 6098062929083.366, "pid": 4, "tid": 0,
    "args": {
      "labels": "GPU 4"
    }
  },
  {
    "name": "process_sort_index", "ph": "M", "ts": 6098062929083.366, "pid": 4, "tid": 0,
    "args": {
      "sort_index": 5000004
    }
  },
  {
    "name": "process_name", "ph": "M", "ts": 6098062929083.366, "pid": 5, "tid": 0,
    "args": {
      "name": "python"
    }
  },
  {
    "name": "process_labels", "ph": "M", "ts": 6098062929083.366, "pid": 5, "tid": 0,
    "args": {
      "labels": "GPU 5"
    }
  },
  {
    "name": "process_sort_index", "ph": "M", "ts": 6098062929083.366, "pid": 5, "tid": 0,
    "args": {
      "sort_index": 5000005
    }
  },
  {
    "name": "process_name", "ph": "M", "ts": 6098062929083.366, "pid": 6, "tid": 0,
    "args": {
      "name": "python"
    }
  },
  {
    "name": "process_labels", "ph": "M", "ts": 6098062929083.366, "pid": 6, "tid": 0,
    "args": {
      "labels": "GPU 6"
    }
  },
  {
    "name": "process_sort_index", "ph": "M", "ts": 6098062929083.366, "pid": 6, "tid": 0,
    "args": {
      "sort_index": 5000006
    }
  },
  {
    "name": "process_name", "ph": "M", "ts": 6098062929083.366, "pid": 7, "tid": 0,
    "args": {
      "name": "python"
    }
  },
  {
    "name": "process_labels", "ph": "M", "ts": 6098062929083.366, "pid": 7, "tid": 0,
    "args": {
      "labels": "GPU 7"
    }
  },
  {
    "name": "process_sort_index", "ph": "M", "ts": 6098062929083.366, "pid": 7, "tid": 0,
    "args": {
      "sort_index": 5000007
    }
  },
  {
    "name": "process_name", "ph": "M", "ts": 6098062929083.366, "pid": 8, "tid": 0,
    "args": {
      "name": "python"
    }
  },
  {
    "name": "process_labels", "ph": "M", "ts": 6098062929083.366, "pid": 8, "tid": 0,
    "args": {
      "labels": "GPU 8"
    }
  },
  {
    "name": "process_sort_index", "ph": "M", "ts": 6098062929083.366, "pid": 8, "tid": 0,
    "args": {
      "sort_index": 5000008
    }
  },
  {
    "name": "process_name", "ph": "M", "ts": 6098062929083.366, "pid": 9, "tid": 0,
    "args": {
      "name": "python"
    }
  },
  {
    "name": "process_labels", "ph": "M", "ts": 6098062929083.366, "pid": 9, "tid": 0,
    "args": {
      "labels": "GPU 9"
    }
  },
  {
    "name": "process_sort_index", "ph": "M", "ts": 6098062929083.366, "pid": 9, "tid": 0,
    "args": {
      "sort_index": 5000009
    }
  },
  {
    "name": "process_name", "ph": "M", "ts": 6098062929083.366, "pid": 10, "tid": 0,
    "args": {
      "name": "python"
    }
  },
  {
    "name": "process_labels", "ph": "M", "ts": 6098062929083.366, "pid": 10, "tid": 0,
    "args": {
      "labels": "GPU 10"
    }
  },
  {
    "name": "process_sort_index", "ph": "M", "ts": 6098062929083.366, "pid": 10, "tid": 0,
    "args": {
      "sort_index": 5000010
    }
  },
  {
    "name": "process_name", "ph": "M", "ts": 6098062929083.366, "pid": 11, "tid": 0,
    "args": {
      "name": "python"
    }
  },
  {
    "name": "process_labels", "ph": "M", "ts": 6098062929083.366, "pid": 11, "tid": 0,
    "args": {
      "labels": "GPU 11"
    }
  },
  {
    "name": "process_sort_index", "ph": "M", "ts": 6098062929083.366, "pid": 11, "tid": 0,
    "args": {
      "sort_index": 5000011
    }
  },
  {
    "name": "process_name", "ph": "M", "ts": 6098062929083.366, "pid": 12, "tid": 0,
    "args": {
      "name": "python"
    }
  },
  {
    "name": "process_labels", "ph": "M", "ts": 6098062929083.366, "pid": 12, "tid": 0,
    "args": {
      "labels": "GPU 12"
    }
  },
  {
    "name": "process_sort_index", "ph": "M", "ts": 6098062929083.366, "pid": 12, "tid": 0,
    "args": {
      "sort_index": 5000012
    }
  },
  {
    "name": "process_name", "ph": "M", "ts": 6098062929083.366, "pid": 13, "tid": 0,
    "args": {
      "name": "python"
    }
  },
  {
    "name": "process_labels", "ph": "M", "ts": 6098062929083.366, "pid": 13, "tid": 0,
    "args": {
      "labels": "GPU 13"
    }
  },
  {
    "name": "process_sort_index", "ph": "M", "ts": 6098062929083.366, "pid": 13, "tid": 0,
    "args": {
      "sort_index": 5000013
    }
  },
  {
    "name": "process_name", "ph": "M", "ts": 6098062929083.366, "pid": 14, "tid": 0,
    "args": {
      "name": "python"
    }
  },
  {
    "name": "process_labels", "ph": "M", "ts": 6098062929083.366, "pid": 14, "tid": 0,
    "args": {
      "labels": "GPU 14"
    }
  },
  {
    "name": "process_sort_index", "ph": "M", "ts": 6098062929083.366, "pid": 14, "tid": 0,
    "args": {
      "sort_index": 5000014
    }
  },
  {
    "name": "process_name", "ph": "M", "ts": 6098062929083.366, "pid": 15, "tid": 0,
    "args": {
      "name": "python"
    }
  },
  {
    "name": "process_labels", "ph": "M", "ts": 6098062929083.366, "pid": 15, "tid": 0,
    "args": {
      "labels": "GPU 15"
    }
  },
  {
    "name": "process_sort_index", "ph": "M", "ts": 6098062929083.366, "pid": 15, "tid": 0,
    "args": {
      "sort_index": 5000015
    }
  },
  {
    "name": "thread_name", "ph": "M", "ts": 6098062929083.366, "pid": 0, "tid": 7,
    "args": {
      "name": "stream 7 "
    }
  },
  {
    "name": "thread_sort_index", "ph": "M", "ts": 6098062929083.366, "pid": 0, "tid": 7,
    "args": {
      "sort_index": 7
    }
  },
  {
    "name": "thread_name", "ph": "M", "ts": 6098062929083.366, "pid": 93046, "tid": 93046,
    "args": {
      "name": "thread 93046 (python)"
    }
  },
  {
    "name": "thread_sort_index", "ph": "M", "ts": 6098062929083.366, "pid": 93046, "tid": 93046,
    "args": {
      "sort_index": 93046
    }
  },
  {
    "name": "thread_name", "ph": "M", "ts": 6098062929083.366, "pid": 93046, "tid": 93046,
    "args": {
      "name": "thread 93046 (python)"
    }
  },
  {
    "name": "thread_sort_index", "ph": "M", "ts": 6098062929083.366, "pid": 93046, "tid": 93046,
    "args": {
      "sort_index": 93046
    }
  },
  {
    "ph": "X", "cat": "Trace", "ts": 6098062928937.797, "dur": 7814.212,
    "pid": "Spans", "tid": "PyTorch Profiler",
    "name": "PyTorch Profiler (0)",
    "args": {
      "Op count": 0
    }
  },
  {
    "name": "process_sort_index", "ph": "M", "ts": 6098062928937.797,
    "pid": "Spans", "tid": 0,
    "args": {
      "sort_index": 536870912
    }
  },
  {
    "name": "Iteration Start: PyTorch Profiler", "ph": "i", "s": "g",
    "pid": "Traces", "tid": "Trace PyTorch Profiler", "ts": 6098062928937.797
  },
  {
    "name": "Record Window End", "ph": "i", "s": "g",
    "pid": "", "tid": "", "ts": 6098062937060.499
  }
  ],
  "traceName": "./traces/01_matmul_add/4096_bf16_cold_eager.json"
}