{
    "schemaVersion": 1,
    "deviceProperties": [
    {
      "id": 0, "name": "NVIDIA RTX 6000D", "totalGlobalMem": 89171361792,
      "computeMajor": 12, "computeMinor": 0,
      "maxThreadsPerBlock": 1024, "maxThreadsPerMultiprocessor": 1536,
      "regsPerBlock": 65536, "warpSize": 32,
      "sharedMemPerBlock": 49152, "numSms": 156
    , "regsPerMultiprocessor": 65536, "sharedMemPerBlockOptin": 101376, "sharedMemPerMultiprocessor": 102400
    }
    ],
      "trace_id": "7E8FE2D42783495C8D16592D686D7C01",
      "record_shapes": 1,
      "cuda_driver_version": 13000,
      "cuda_runtime_version": 13000,
      "cupti_version": 130001,
  "displayTimeUnit": "ms",
  "baseTimeNanoseconds": 1782967788000000000,
  "traceEvents": [
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::scaled_dot_product_attention", "pid": 2815986, "tid": 2815986,
    "ts": 4656938469238.885, "dur": 4696.250,
    "args": {
      "External id": 513,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", "", "", "", "0.", "False", "", "False"], "Input type": ["c10::BFloat16", "c10::BFloat16", "c10::BFloat16", "", "Scalar", "Scalar", "", "Scalar"], "Input Strides": [[149732352, 5347584, 128, 1], [149732352, 5347584, 128, 1], [149732352, 5347584, 128, 1], [], [], [], [], []], "Input Dims": [[1, 28, 41778, 128], [1, 28, 41778, 128], [1, 28, 41778, 128], [], [], [], [], []], "Ev Idx": 0
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::_scaled_dot_product_cudnn_attention", "pid": 2815986, "tid": 2815986,
    "ts": 4656938469274.583, "dur": 4653.455,
    "args": {
      "External id": 514,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", "", "", "", "False", "0.", "False", "False", ""], "Input type": ["c10::BFloat16", "c10::BFloat16", "c10::BFloat16", "", "Scalar", "Scalar", "Scalar", "Scalar", ""], "Input Strides": [[149732352, 5347584, 128, 1], [149732352, 5347584, 128, 1], [149732352, 5347584, 128, 1], [], [], [], [], [], []], "Input Dims": [[1, 28, 41778, 128], [1, 28, 41778, 128], [1, 28, 41778, 128], [], [], [], [], [], []], "Ev Idx": 1
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::_cudnn_attention_forward", "pid": 2815986, "tid": 2815986,
    "ts": 4656938469287.695, "dur": 4636.076,
    "args": {
      "External id": 515,"Record function id": 0, "Concrete Inputs": ["", "", "", "", "", "", "41778", "41778", "False", "0.", "False", "False", ""], "Input type": ["c10::BFloat16", "c10::BFloat16", "c10::BFloat16", "", "", "", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar", ""], "Input Strides": [[149732352, 5347584, 128, 1], [149732352, 5347584, 128, 1], [149732352, 5347584, 128, 1], [], [], [], [], [], [], [], [], [], []], "Input Dims": [[1, 28, 41778, 128], [1, 28, 41778, 128], [1, 28, 41778, 128], [], [], [], [], [], [], [], [], [], []], "Ev Idx": 2
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 2815986, "tid": 2815986,
    "ts": 4656938469301.241, "dur": 3254.801,
    "args": {
      "External id": 516,"Record function id": 0, "Concrete Inputs": ["[]", "4", "", "", "", ""], "Input type": ["ScalarList", "Scalar", "", "", "", ""], "Input Strides": [[], [], [], [], [], []], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 3
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 2815986, "tid": 2815986,
    "ts": 4656938472568.315, "dur": 9.919,
    "args": {
      "External id": 517,"Record function id": 0, "Concrete Inputs": ["[]", "4", "", "", "", ""], "Input type": ["ScalarList", "Scalar", "", "", "", ""], "Input Strides": [[], [], [], [], [], []], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 4
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::empty_like", "pid": 2815986, "tid": 2815986,
    "ts": 4656938472590.978, "dur": 1211.873,
    "args": {
      "External id": 518,"Record function id": 0, "Concrete Inputs": ["", "", "", "", "", ""], "Input type": ["c10::BFloat16", "", "", "", "", ""], "Input Strides": [[149732352, 5347584, 128, 1], [], [], [], [], []], "Input Dims": [[1, 28, 41778, 128], [], [], [], [], []], "Ev Idx": 5
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::empty_strided", "pid": 2815986, "tid": 2815986,
    "ts": 4656938472596.426, "dur": 1204.722,
    "args": {
      "External id": 519,"Record function id": 0, "Concrete Inputs": ["[1, 28, 41778, 128]", "[149732352, 5347584, 128, 1]", "15", "0", "", ""], "Input type": ["ScalarList", "ScalarList", "Scalar", "Scalar", "", ""], "Input Strides": [[], [], [], [], [], []], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 6
    }
  },
  {
    "ph": "X", "cat": "overhead", "name": "Activity Buffer Request", "pid": -1, "tid": 0,
    "ts": 4656938469317.498, "dur": 2227.519
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "cudaStreamIsCapturing", "pid": 2815986, "tid": 2815986,
    "ts": 4656938471555.236, "dur": 4.141,
    "args": {
      "External id": 516,      "cbid": 317, "correlation": 2488
    }
  },
  {
    "ph": "f", "id": 2488, "pid": 2815986, "tid": 2815986, "ts": 4656938471555.236,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "cudaMalloc", "pid": 2815986, "tid": 2815986,
    "ts": 4656938471560.368, "dur": 973.779,
    "args": {
      "External id": 516,      "cbid": 20, "correlation": 2489
    }
  },
  {
    "ph": "f", "id": 2489, "pid": 2815986, "tid": 2815986, "ts": 4656938471560.368,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "cudaStreamIsCapturing", "pid": 2815986, "tid": 2815986,
    "ts": 4656938472581.420, "dur": 1.478,
    "args": {
      "External id": 515,      "cbid": 317, "correlation": 2504
    }
  },
  {
    "ph": "f", "id": 2504, "pid": 2815986, "tid": 2815986, "ts": 4656938472581.420,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "cudaStreamIsCapturing", "pid": 2815986, "tid": 2815986,
    "ts": 4656938472605.319, "dur": 0.350,
    "args": {
      "External id": 519,      "cbid": 317, "correlation": 2510
    }
  },
  {
    "ph": "f", "id": 2510, "pid": 2815986, "tid": 2815986, "ts": 4656938472605.319,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "cudaMalloc", "pid": 2815986, "tid": 2815986,
    "ts": 4656938472606.031, "dur": 1188.139,
    "args": {
      "External id": 519,      "cbid": 20, "correlation": 2511
    }
  },
  {
    "ph": "f", "id": 2511, "pid": 2815986, "tid": 2815986, "ts": 4656938472606.031,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "cudaEventRecord", "pid": 2815986, "tid": 2815986,
    "ts": 4656938473811.384, "dur": 7.258,
    "args": {
      "External id": 515,      "cbid": 135, "correlation": 2519
    }
  },
  {
    "ph": "f", "id": 2519, "pid": 2815986, "tid": 2815986, "ts": 4656938473811.384,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "cudaStreamGetPriority", "pid": 2815986, "tid": 2815986,
    "ts": 4656938473819.380, "dur": 1.144,
    "args": {
      "External id": 515,      "cbid": 203, "correlation": 2520
    }
  },
  {
    "ph": "f", "id": 2520, "pid": 2815986, "tid": 2815986, "ts": 4656938473819.380,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "cudaDeviceGetStreamPriorityRange", "pid": 2815986, "tid": 2815986,
    "ts": 4656938473821.060, "dur": 0.883,
    "args": {
      "External id": 515,      "cbid": 205, "correlation": 2521
    }
  },
  {
    "ph": "f", "id": 2521, "pid": 2815986, "tid": 2815986, "ts": 4656938473821.060,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "cudnn_generated_fort_native_sdpa_sm120_flash_fprop_f16_knob_3_64x64x128_4x1x1_cga1x1x1_kernel0_0", "pid": 0, "tid": 7,
    "ts": 4656938473913.961, "dur": 173474.738,
    "args": {
      "External id": 515,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 2542,      "registers per thread": 168,      "shared memory": 101376,      "blocks per SM": 58.692307,      "warps per SM": 528.230774,      "grid": [327, 28, 1],      "block": [288, 1, 1],      "est. achieved occupancy %": 0
    }
  },
  {
    "ph": "f", "id": 2542, "pid": 0, "tid": 7, "ts": 4656938473913.961,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "cuda_driver", "name": "cuLaunchKernelEx", "pid": 2815986, "tid": 2815986,
    "ts": 4656938473863.549, "dur": 49.501,
    "args": {
      "External id": 515,      "cbid": 652, "correlation": 2542
    }
  },
  {
    "ph": "s", "id": 2542, "pid": 2815986, "tid": 2815986, "ts": 4656938473863.549,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "cudaGetErrorString", "pid": 2815986, "tid": 2815986,
    "ts": 4656938473916.201, "dur": 0.513,
    "args": {
      "External id": 515,      "cbid": 12, "correlation": 2544
    }
  },
  {
    "ph": "f", "id": 2544, "pid": 2815986, "tid": 2815986, "ts": 4656938473916.201,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "cudaDeviceSynchronize", "pid": 2815986, "tid": 2815986,
    "ts": 4656938474036.376, "dur": 173361.000,
    "args": {
            "cbid": 165, "correlation": 2551
    }
  },
  {
    "ph": "s", "id": 2551, "pid": 2815986, "tid": 2815986, "ts": 4656938474036.376,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "cudaDeviceSynchronize", "pid": 2815986, "tid": 2815986,
    "ts": 4656938647433.039, "dur": 4.836,
    "args": {
            "cbid": 165, "correlation": 2555
    }
  },
  {
    "ph": "s", "id": 2555, "pid": 2815986, "tid": 2815986, "ts": 4656938647433.039,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "name": "process_name", "ph": "M", "ts": 4656938468997.603, "pid": 2815986, "tid": 0,
    "args": {
      "name": "python"
    }
  },
  {
    "name": "process_labels", "ph": "M", "ts": 4656938468997.603, "pid": 2815986, "tid": 0,
    "args": {
      "labels": "CPU"
    }
  },
  {
    "name": "process_sort_index", "ph": "M", "ts": 4656938468997.603, "pid": 2815986, "tid": 0,
    "args": {
      "sort_index": 2815986
    }
  },
  {
    "name": "process_name", "ph": "M", "ts": 4656938468997.603, "pid": 0, "tid": 0,
    "args": {
      "name": "python"
    }
  },
  {
    "name": "process_labels", "ph": "M", "ts": 4656938468997.603, "pid": 0, "tid": 0,
    "args": {
      "labels": "GPU 0"
    }
  },
  {
    "name": "process_sort_index", "ph": "M", "ts": 4656938468997.603, "pid": 0, "tid": 0,
    "args": {
      "sort_index": 5000000
    }
  },
  {
    "name": "process_name", "ph": "M", "ts": 4656938468997.603, "pid": 1, "tid": 0,
    "args": {
      "name": "python"
    }
  },
  {
    "name": "process_labels", "ph": "M", "ts": 4656938468997.603, "pid": 1, "tid": 0,
    "args": {
      "labels": "GPU 1"
    }
  },
  {
    "name": "process_sort_index", "ph": "M", "ts": 4656938468997.603, "pid": 1, "tid": 0,
    "args": {
      "sort_index": 5000001
    }
  },
  {
    "name": "process_name", "ph": "M", "ts": 4656938468997.603, "pid": 2, "tid": 0,
    "args": {
      "name": "python"
    }
  },
  {
    "name": "process_labels", "ph": "M", "ts": 4656938468997.603, "pid": 2, "tid": 0,
    "args": {
      "labels": "GPU 2"
    }
  },
  {
    "name": "process_sort_index", "ph": "M", "ts": 4656938468997.603, "pid": 2, "tid": 0,
    "args": {
      "sort_index": 5000002
    }
  },
  {
    "name": "process_name", "ph": "M", "ts": 4656938468997.603, "pid": 3, "tid": 0,
    "args": {
      "name": "python"
    }
  },
  {
    "name": "process_labels", "ph": "M", "ts": 4656938468997.603, "pid": 3, "tid": 0,
    "args": {
      "labels": "GPU 3"
    }
  },
  {
    "name": "process_sort_index", "ph": "M", "ts": 4656938468997.603, "pid": 3, "tid": 0,
    "args": {
      "sort_index": 5000003
    }
  },
  {
    "name": "process_name", "ph": "M", "ts": 4656938468997.603, "pid": 4, "tid": 0,
    "args": {
      "name": "python"
    }
  },
  {
    "name": "process_labels", "ph": "M", "ts": 4656938468997.603, "pid": 4, "tid": 0,
    "args": {
      "labels": "GPU 4"
    }
  },
  {
    "name": "process_sort_index", "ph": "M", "ts": 4656938468997.603, "pid": 4, "tid": 0,
    "args": {
      "sort_index": 5000004
    }
  },
  {
    "name": "process_name", "ph": "M", "ts": 4656938468997.603, "pid": 5, "tid": 0,
    "args": {
      "name": "python"
    }
  },
  {
    "name": "process_labels", "ph": "M", "ts": 4656938468997.603, "pid": 5, "tid": 0,
    "args": {
      "labels": "GPU 5"
    }
  },
  {
    "name": "process_sort_index", "ph": "M", "ts": 4656938468997.603, "pid": 5, "tid": 0,
    "args": {
      "sort_index": 5000005
    }
  },
  {
    "name": "process_name", "ph": "M", "ts": 4656938468997.603, "pid": 6, "tid": 0,
    "args": {
      "name": "python"
    }
  },
  {
    "name": "process_labels", "ph": "M", "ts": 4656938468997.603, "pid": 6, "tid": 0,
    "args": {
      "labels": "GPU 6"
    }
  },
  {
    "name": "process_sort_index", "ph": "M", "ts": 4656938468997.603, "pid": 6, "tid": 0,
    "args": {
      "sort_index": 5000006
    }
  },
  {
    "name": "process_name", "ph": "M", "ts": 4656938468997.603, "pid": 7, "tid": 0,
    "args": {
      "name": "python"
    }
  },
  {
    "name": "process_labels", "ph": "M", "ts": 4656938468997.603, "pid": 7, "tid": 0,
    "args": {
      "labels": "GPU 7"
    }
  },
  {
    "name": "process_sort_index", "ph": "M", "ts": 4656938468997.603, "pid": 7, "tid": 0,
    "args": {
      "sort_index": 5000007
    }
  },
  {
    "name": "process_name", "ph": "M", "ts": 4656938468997.603, "pid": 8, "tid": 0,
    "args": {
      "name": "python"
    }
  },
  {
    "name": "process_labels", "ph": "M", "ts": 4656938468997.603, "pid": 8, "tid": 0,
    "args": {
      "labels": "GPU 8"
    }
  },
  {
    "name": "process_sort_index", "ph": "M", "ts": 4656938468997.603, "pid": 8, "tid": 0,
    "args": {
      "sort_index": 5000008
    }
  },
  {
    "name": "process_name", "ph": "M", "ts": 4656938468997.603, "pid": 9, "tid": 0,
    "args": {
      "name": "python"
    }
  },
  {
    "name": "process_labels", "ph": "M", "ts": 4656938468997.603, "pid": 9, "tid": 0,
    "args": {
      "labels": "GPU 9"
    }
  },
  {
    "name": "process_sort_index", "ph": "M", "ts": 4656938468997.603, "pid": 9, "tid": 0,
    "args": {
      "sort_index": 5000009
    }
  },
  {
    "name": "process_name", "ph": "M", "ts": 4656938468997.603, "pid": 10, "tid": 0,
    "args": {
      "name": "python"
    }
  },
  {
    "name": "process_labels", "ph": "M", "ts": 4656938468997.603, "pid": 10, "tid": 0,
    "args": {
      "labels": "GPU 10"
    }
  },
  {
    "name": "process_sort_index", "ph": "M", "ts": 4656938468997.603, "pid": 10, "tid": 0,
    "args": {
      "sort_index": 5000010
    }
  },
  {
    "name": "process_name", "ph": "M", "ts": 4656938468997.603, "pid": 11, "tid": 0,
    "args": {
      "name": "python"
    }
  },
  {
    "name": "process_labels", "ph": "M", "ts": 4656938468997.603, "pid": 11, "tid": 0,
    "args": {
      "labels": "GPU 11"
    }
  },
  {
    "name": "process_sort_index", "ph": "M", "ts": 4656938468997.603, "pid": 11, "tid": 0,
    "args": {
      "sort_index": 5000011
    }
  },
  {
    "name": "process_name", "ph": "M", "ts": 4656938468997.603, "pid": 12, "tid": 0,
    "args": {
      "name": "python"
    }
  },
  {
    "name": "process_labels", "ph": "M", "ts": 4656938468997.603, "pid": 12, "tid": 0,
    "args": {
      "labels": "GPU 12"
    }
  },
  {
    "name": "process_sort_index", "ph": "M", "ts": 4656938468997.603, "pid": 12, "tid": 0,
    "args": {
      "sort_index": 5000012
    }
  },
  {
    "name": "process_name", "ph": "M", "ts": 4656938468997.603, "pid": 13, "tid": 0,
    "args": {
      "name": "python"
    }
  },
  {
    "name": "process_labels", "ph": "M", "ts": 4656938468997.603, "pid": 13, "tid": 0,
    "args": {
      "labels": "GPU 13"
    }
  },
  {
    "name": "process_sort_index", "ph": "M", "ts": 4656938468997.603, "pid": 13, "tid": 0,
    "args": {
      "sort_index": 5000013
    }
  },
  {
    "name": "process_name", "ph": "M", "ts": 4656938468997.603, "pid": 14, "tid": 0,
    "args": {
      "name": "python"
    }
  },
  {
    "name": "process_labels", "ph": "M", "ts": 4656938468997.603, "pid": 14, "tid": 0,
    "args": {
      "labels": "GPU 14"
    }
  },
  {
    "name": "process_sort_index", "ph": "M", "ts": 4656938468997.603, "pid": 14, "tid": 0,
    "args": {
      "sort_index": 5000014
    }
  },
  {
    "name": "process_name", "ph": "M", "ts": 4656938468997.603, "pid": 15, "tid": 0,
    "args": {
      "name": "python"
    }
  },
  {
    "name": "process_labels", "ph": "M", "ts": 4656938468997.603, "pid": 15, "tid": 0,
    "args": {
      "labels": "GPU 15"
    }
  },
  {
    "name": "process_sort_index", "ph": "M", "ts": 4656938468997.603, "pid": 15, "tid": 0,
    "args": {
      "sort_index": 5000015
    }
  },
  {
    "name": "thread_name", "ph": "M", "ts": 4656938468997.603, "pid": 0, "tid": 7,
    "args": {
      "name": "stream 7 "
    }
  },
  {
    "name": "thread_sort_index", "ph": "M", "ts": 4656938468997.603, "pid": 0, "tid": 7,
    "args": {
      "sort_index": 7
    }
  },
  {
    "name": "thread_name", "ph": "M", "ts": 4656938468997.603, "pid": 2815986, "tid": 2815986,
    "args": {
      "name": "thread 2815986 (python)"
    }
  },
  {
    "name": "thread_sort_index", "ph": "M", "ts": 4656938468997.603, "pid": 2815986, "tid": 2815986,
    "args": {
      "sort_index": 2815986
    }
  },
  {
    "name": "thread_name", "ph": "M", "ts": 4656938468997.603, "pid": 2815986, "tid": 2815986,
    "args": {
      "name": "thread 2815986 (python)"
    }
  },
  {
    "name": "thread_sort_index", "ph": "M", "ts": 4656938468997.603, "pid": 2815986, "tid": 2815986,
    "args": {
      "sort_index": 2815986
    }
  },
  {
    "ph": "X", "cat": "Trace", "ts": 4656938468937.489, "dur": 178513.187,
    "pid": "Spans", "tid": "PyTorch Profiler",
    "name": "PyTorch Profiler (0)",
    "args": {
      "Op count": 0
    }
  },
  {
    "name": "process_sort_index", "ph": "M", "ts": 4656938468937.489,
    "pid": "Spans", "tid": 0,
    "args": {
      "sort_index": 536870912
    }
  },
  {
    "name": "Iteration Start: PyTorch Profiler", "ph": "i", "s": "g",
    "pid": "Traces", "tid": "Trace PyTorch Profiler", "ts": 4656938468937.489
  },
  {
    "name": "Record Window End", "ph": "i", "s": "g",
    "pid": "", "tid": "", "ts": 4656938647719.679
  }
  ],
  "traceName": "/data/wxy/profile_results/h3-sdpa-kernel-matrix-20260825-run1/jobs/gpu1-F3-cudnn-contiguous.trace.json"
}