{
    "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": "3D6E95AF618E435EB76F2FF7F3F3E489",
      "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": 2815990, "tid": 2815990,
    "ts": 4656948310763.802, "dur": 3582.988,
    "args": {
      "External id": 1,"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": [[876793344, 128, 10752, 1], [876793344, 128, 10752, 1], [876793344, 128, 10752, 1], [], [], [], [], []], "Input Dims": [[1, 28, 81547, 128], [1, 28, 81547, 128], [1, 28, 81547, 128], [], [], [], [], []], "Ev Idx": 0
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::_scaled_dot_product_cudnn_attention", "pid": 2815990, "tid": 2815990,
    "ts": 4656948310822.780, "dur": 3519.118,
    "args": {
      "External id": 2,"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": [[876793344, 128, 10752, 1], [876793344, 128, 10752, 1], [876793344, 128, 10752, 1], [], [], [], [], [], []], "Input Dims": [[1, 28, 81547, 128], [1, 28, 81547, 128], [1, 28, 81547, 128], [], [], [], [], [], []], "Ev Idx": 1
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::_cudnn_attention_forward", "pid": 2815990, "tid": 2815990,
    "ts": 4656948310845.398, "dur": 3493.586,
    "args": {
      "External id": 3,"Record function id": 0, "Concrete Inputs": ["", "", "", "", "", "", "81547", "81547", "False", "0.", "False", "False", ""], "Input type": ["c10::BFloat16", "c10::BFloat16", "c10::BFloat16", "", "", "", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar", ""], "Input Strides": [[876793344, 128, 10752, 1], [876793344, 128, 10752, 1], [876793344, 128, 10752, 1], [], [], [], [], [], [], [], [], [], []], "Input Dims": [[1, 28, 81547, 128], [1, 28, 81547, 128], [1, 28, 81547, 128], [], [], [], [], [], [], [], [], [], []], "Ev Idx": 2
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 2815990, "tid": 2815990,
    "ts": 4656948310860.082, "dur": 2816.944,
    "args": {
      "External id": 4,"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": 2815990, "tid": 2815990,
    "ts": 4656948313687.108, "dur": 7.318,
    "args": {
      "External id": 5,"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": 2815990, "tid": 2815990,
    "ts": 4656948313712.869, "dur": 497.090,
    "args": {
      "External id": 6,"Record function id": 0, "Concrete Inputs": ["", "", "", "", "", ""], "Input type": ["c10::BFloat16", "", "", "", "", ""], "Input Strides": [[876793344, 128, 10752, 1], [], [], [], [], []], "Input Dims": [[1, 28, 81547, 128], [], [], [], [], []], "Ev Idx": 5
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::empty_strided", "pid": 2815990, "tid": 2815990,
    "ts": 4656948313723.337, "dur": 484.896,
    "args": {
      "External id": 7,"Record function id": 0, "Concrete Inputs": ["[1, 28, 81547, 128]", "[292264448, 128, 3584, 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": 4656948310875.899, "dur": 1901.558
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "cudaStreamIsCapturing", "pid": 2815990, "tid": 2815990,
    "ts": 4656948312785.612, "dur": 6.722,
    "args": {
      "External id": 4,      "cbid": 317, "correlation": 13
    }
  },
  {
    "ph": "f", "id": 13, "pid": 2815990, "tid": 2815990, "ts": 4656948312785.612,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "cudaMalloc", "pid": 2815990, "tid": 2815990,
    "ts": 4656948312793.014, "dur": 866.625,
    "args": {
      "External id": 4,      "cbid": 20, "correlation": 14
    }
  },
  {
    "ph": "f", "id": 14, "pid": 2815990, "tid": 2815990, "ts": 4656948312793.014,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "cudaStreamIsCapturing", "pid": 2815990, "tid": 2815990,
    "ts": 4656948313696.834, "dur": 0.756,
    "args": {
      "External id": 3,      "cbid": 317, "correlation": 29
    }
  },
  {
    "ph": "f", "id": 29, "pid": 2815990, "tid": 2815990, "ts": 4656948313696.834,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "cudaStreamIsCapturing", "pid": 2815990, "tid": 2815990,
    "ts": 4656948313731.961, "dur": 0.306,
    "args": {
      "External id": 7,      "cbid": 317, "correlation": 35
    }
  },
  {
    "ph": "f", "id": 35, "pid": 2815990, "tid": 2815990, "ts": 4656948313731.961,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "cudaMalloc", "pid": 2815990, "tid": 2815990,
    "ts": 4656948313732.559, "dur": 469.809,
    "args": {
      "External id": 7,      "cbid": 20, "correlation": 36
    }
  },
  {
    "ph": "f", "id": 36, "pid": 2815990, "tid": 2815990, "ts": 4656948313732.559,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "cudaEventRecord", "pid": 2815990, "tid": 2815990,
    "ts": 4656948314223.800, "dur": 5.836,
    "args": {
      "External id": 3,      "cbid": 135, "correlation": 44
    }
  },
  {
    "ph": "f", "id": 44, "pid": 2815990, "tid": 2815990, "ts": 4656948314223.800,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "cudaStreamGetPriority", "pid": 2815990, "tid": 2815990,
    "ts": 4656948314230.237, "dur": 2.133,
    "args": {
      "External id": 3,      "cbid": 203, "correlation": 45
    }
  },
  {
    "ph": "f", "id": 45, "pid": 2815990, "tid": 2815990, "ts": 4656948314230.237,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "cudaDeviceGetStreamPriorityRange", "pid": 2815990, "tid": 2815990,
    "ts": 4656948314232.935, "dur": 1.083,
    "args": {
      "External id": 3,      "cbid": 205, "correlation": 46
    }
  },
  {
    "ph": "f", "id": 46, "pid": 2815990, "tid": 2815990, "ts": 4656948314232.935,
    "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": 4656948314327.459, "dur": 661030.132,
    "args": {
      "External id": 3,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 67,      "registers per thread": 168,      "shared memory": 101376,      "blocks per SM": 114.512817,      "warps per SM": 1030.615356,      "grid": [638, 28, 1],      "block": [288, 1, 1],      "est. achieved occupancy %": 0
    }
  },
  {
    "ph": "f", "id": 67, "pid": 0, "tid": 7, "ts": 4656948314327.459,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "cuda_driver", "name": "cuLaunchKernelEx", "pid": 2815990, "tid": 2815990,
    "ts": 4656948314289.836, "dur": 36.969,
    "args": {
      "External id": 3,      "cbid": 652, "correlation": 67
    }
  },
  {
    "ph": "s", "id": 67, "pid": 2815990, "tid": 2815990, "ts": 4656948314289.836,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "cudaGetErrorString", "pid": 2815990, "tid": 2815990,
    "ts": 4656948314331.075, "dur": 1.607,
    "args": {
      "External id": 3,      "cbid": 12, "correlation": 69
    }
  },
  {
    "ph": "f", "id": 69, "pid": 2815990, "tid": 2815990, "ts": 4656948314331.075,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "cudaDeviceSynchronize", "pid": 2815990, "tid": 2815990,
    "ts": 4656948314414.333, "dur": 660967.534,
    "args": {
            "cbid": 165, "correlation": 76
    }
  },
  {
    "ph": "s", "id": 76, "pid": 2815990, "tid": 2815990, "ts": 4656948314414.333,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "cudaDeviceSynchronize", "pid": 2815990, "tid": 2815990,
    "ts": 4656948975529.670, "dur": 128.626,
    "args": {
            "cbid": 165, "correlation": 80
    }
  },
  {
    "ph": "s", "id": 80, "pid": 2815990, "tid": 2815990, "ts": 4656948975529.670,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "name": "process_name", "ph": "M", "ts": 4656948310642.045, "pid": 2815990, "tid": 0,
    "args": {
      "name": "python"
    }
  },
  {
    "name": "process_labels", "ph": "M", "ts": 4656948310642.045, "pid": 2815990, "tid": 0,
    "args": {
      "labels": "CPU"
    }
  },
  {
    "name": "process_sort_index", "ph": "M", "ts": 4656948310642.045, "pid": 2815990, "tid": 0,
    "args": {
      "sort_index": 2815990
    }
  },
  {
    "name": "process_name", "ph": "M", "ts": 4656948310642.045, "pid": 0, "tid": 0,
    "args": {
      "name": "python"
    }
  },
  {
    "name": "process_labels", "ph": "M", "ts": 4656948310642.045, "pid": 0, "tid": 0,
    "args": {
      "labels": "GPU 0"
    }
  },
  {
    "name": "process_sort_index", "ph": "M", "ts": 4656948310642.045, "pid": 0, "tid": 0,
    "args": {
      "sort_index": 5000000
    }
  },
  {
    "name": "process_name", "ph": "M", "ts": 4656948310642.045, "pid": 1, "tid": 0,
    "args": {
      "name": "python"
    }
  },
  {
    "name": "process_labels", "ph": "M", "ts": 4656948310642.045, "pid": 1, "tid": 0,
    "args": {
      "labels": "GPU 1"
    }
  },
  {
    "name": "process_sort_index", "ph": "M", "ts": 4656948310642.045, "pid": 1, "tid": 0,
    "args": {
      "sort_index": 5000001
    }
  },
  {
    "name": "process_name", "ph": "M", "ts": 4656948310642.045, "pid": 2, "tid": 0,
    "args": {
      "name": "python"
    }
  },
  {
    "name": "process_labels", "ph": "M", "ts": 4656948310642.045, "pid": 2, "tid": 0,
    "args": {
      "labels": "GPU 2"
    }
  },
  {
    "name": "process_sort_index", "ph": "M", "ts": 4656948310642.045, "pid": 2, "tid": 0,
    "args": {
      "sort_index": 5000002
    }
  },
  {
    "name": "process_name", "ph": "M", "ts": 4656948310642.045, "pid": 3, "tid": 0,
    "args": {
      "name": "python"
    }
  },
  {
    "name": "process_labels", "ph": "M", "ts": 4656948310642.045, "pid": 3, "tid": 0,
    "args": {
      "labels": "GPU 3"
    }
  },
  {
    "name": "process_sort_index", "ph": "M", "ts": 4656948310642.045, "pid": 3, "tid": 0,
    "args": {
      "sort_index": 5000003
    }
  },
  {
    "name": "process_name", "ph": "M", "ts": 4656948310642.045, "pid": 4, "tid": 0,
    "args": {
      "name": "python"
    }
  },
  {
    "name": "process_labels", "ph": "M", "ts": 4656948310642.045, "pid": 4, "tid": 0,
    "args": {
      "labels": "GPU 4"
    }
  },
  {
    "name": "process_sort_index", "ph": "M", "ts": 4656948310642.045, "pid": 4, "tid": 0,
    "args": {
      "sort_index": 5000004
    }
  },
  {
    "name": "process_name", "ph": "M", "ts": 4656948310642.045, "pid": 5, "tid": 0,
    "args": {
      "name": "python"
    }
  },
  {
    "name": "process_labels", "ph": "M", "ts": 4656948310642.045, "pid": 5, "tid": 0,
    "args": {
      "labels": "GPU 5"
    }
  },
  {
    "name": "process_sort_index", "ph": "M", "ts": 4656948310642.045, "pid": 5, "tid": 0,
    "args": {
      "sort_index": 5000005
    }
  },
  {
    "name": "process_name", "ph": "M", "ts": 4656948310642.045, "pid": 6, "tid": 0,
    "args": {
      "name": "python"
    }
  },
  {
    "name": "process_labels", "ph": "M", "ts": 4656948310642.045, "pid": 6, "tid": 0,
    "args": {
      "labels": "GPU 6"
    }
  },
  {
    "name": "process_sort_index", "ph": "M", "ts": 4656948310642.045, "pid": 6, "tid": 0,
    "args": {
      "sort_index": 5000006
    }
  },
  {
    "name": "process_name", "ph": "M", "ts": 4656948310642.045, "pid": 7, "tid": 0,
    "args": {
      "name": "python"
    }
  },
  {
    "name": "process_labels", "ph": "M", "ts": 4656948310642.045, "pid": 7, "tid": 0,
    "args": {
      "labels": "GPU 7"
    }
  },
  {
    "name": "process_sort_index", "ph": "M", "ts": 4656948310642.045, "pid": 7, "tid": 0,
    "args": {
      "sort_index": 5000007
    }
  },
  {
    "name": "process_name", "ph": "M", "ts": 4656948310642.045, "pid": 8, "tid": 0,
    "args": {
      "name": "python"
    }
  },
  {
    "name": "process_labels", "ph": "M", "ts": 4656948310642.045, "pid": 8, "tid": 0,
    "args": {
      "labels": "GPU 8"
    }
  },
  {
    "name": "process_sort_index", "ph": "M", "ts": 4656948310642.045, "pid": 8, "tid": 0,
    "args": {
      "sort_index": 5000008
    }
  },
  {
    "name": "process_name", "ph": "M", "ts": 4656948310642.045, "pid": 9, "tid": 0,
    "args": {
      "name": "python"
    }
  },
  {
    "name": "process_labels", "ph": "M", "ts": 4656948310642.045, "pid": 9, "tid": 0,
    "args": {
      "labels": "GPU 9"
    }
  },
  {
    "name": "process_sort_index", "ph": "M", "ts": 4656948310642.045, "pid": 9, "tid": 0,
    "args": {
      "sort_index": 5000009
    }
  },
  {
    "name": "process_name", "ph": "M", "ts": 4656948310642.045, "pid": 10, "tid": 0,
    "args": {
      "name": "python"
    }
  },
  {
    "name": "process_labels", "ph": "M", "ts": 4656948310642.045, "pid": 10, "tid": 0,
    "args": {
      "labels": "GPU 10"
    }
  },
  {
    "name": "process_sort_index", "ph": "M", "ts": 4656948310642.045, "pid": 10, "tid": 0,
    "args": {
      "sort_index": 5000010
    }
  },
  {
    "name": "process_name", "ph": "M", "ts": 4656948310642.045, "pid": 11, "tid": 0,
    "args": {
      "name": "python"
    }
  },
  {
    "name": "process_labels", "ph": "M", "ts": 4656948310642.045, "pid": 11, "tid": 0,
    "args": {
      "labels": "GPU 11"
    }
  },
  {
    "name": "process_sort_index", "ph": "M", "ts": 4656948310642.045, "pid": 11, "tid": 0,
    "args": {
      "sort_index": 5000011
    }
  },
  {
    "name": "process_name", "ph": "M", "ts": 4656948310642.045, "pid": 12, "tid": 0,
    "args": {
      "name": "python"
    }
  },
  {
    "name": "process_labels", "ph": "M", "ts": 4656948310642.045, "pid": 12, "tid": 0,
    "args": {
      "labels": "GPU 12"
    }
  },
  {
    "name": "process_sort_index", "ph": "M", "ts": 4656948310642.045, "pid": 12, "tid": 0,
    "args": {
      "sort_index": 5000012
    }
  },
  {
    "name": "process_name", "ph": "M", "ts": 4656948310642.045, "pid": 13, "tid": 0,
    "args": {
      "name": "python"
    }
  },
  {
    "name": "process_labels", "ph": "M", "ts": 4656948310642.045, "pid": 13, "tid": 0,
    "args": {
      "labels": "GPU 13"
    }
  },
  {
    "name": "process_sort_index", "ph": "M", "ts": 4656948310642.045, "pid": 13, "tid": 0,
    "args": {
      "sort_index": 5000013
    }
  },
  {
    "name": "process_name", "ph": "M", "ts": 4656948310642.045, "pid": 14, "tid": 0,
    "args": {
      "name": "python"
    }
  },
  {
    "name": "process_labels", "ph": "M", "ts": 4656948310642.045, "pid": 14, "tid": 0,
    "args": {
      "labels": "GPU 14"
    }
  },
  {
    "name": "process_sort_index", "ph": "M", "ts": 4656948310642.045, "pid": 14, "tid": 0,
    "args": {
      "sort_index": 5000014
    }
  },
  {
    "name": "process_name", "ph": "M", "ts": 4656948310642.045, "pid": 15, "tid": 0,
    "args": {
      "name": "python"
    }
  },
  {
    "name": "process_labels", "ph": "M", "ts": 4656948310642.045, "pid": 15, "tid": 0,
    "args": {
      "labels": "GPU 15"
    }
  },
  {
    "name": "process_sort_index", "ph": "M", "ts": 4656948310642.045, "pid": 15, "tid": 0,
    "args": {
      "sort_index": 5000015
    }
  },
  {
    "name": "thread_name", "ph": "M", "ts": 4656948310642.045, "pid": 0, "tid": 7,
    "args": {
      "name": "stream 7 "
    }
  },
  {
    "name": "thread_sort_index", "ph": "M", "ts": 4656948310642.045, "pid": 0, "tid": 7,
    "args": {
      "sort_index": 7
    }
  },
  {
    "name": "thread_name", "ph": "M", "ts": 4656948310642.045, "pid": 2815990, "tid": 2815990,
    "args": {
      "name": "thread 2815990 (python)"
    }
  },
  {
    "name": "thread_sort_index", "ph": "M", "ts": 4656948310642.045, "pid": 2815990, "tid": 2815990,
    "args": {
      "sort_index": 2815990
    }
  },
  {
    "name": "thread_name", "ph": "M", "ts": 4656948310642.045, "pid": 2815990, "tid": 2815990,
    "args": {
      "name": "thread 2815990 (python)"
    }
  },
  {
    "name": "thread_sort_index", "ph": "M", "ts": 4656948310642.045, "pid": 2815990, "tid": 2815990,
    "args": {
      "sort_index": 2815990
    }
  },
  {
    "ph": "X", "cat": "Trace", "ts": 4656948310591.373, "dur": 665088.242,
    "pid": "Spans", "tid": "PyTorch Profiler",
    "name": "PyTorch Profiler (0)",
    "args": {
      "Op count": 0
    }
  },
  {
    "name": "process_sort_index", "ph": "M", "ts": 4656948310591.373,
    "pid": "Spans", "tid": 0,
    "args": {
      "sort_index": 536870912
    }
  },
  {
    "name": "Iteration Start: PyTorch Profiler", "ph": "i", "s": "g",
    "pid": "Traces", "tid": "Trace PyTorch Profiler", "ts": 4656948310591.373
  },
  {
    "name": "Record Window End", "ph": "i", "s": "g",
    "pid": "", "tid": "", "ts": 4656948975952.480
  }
  ],
  "traceName": "/data/wxy/profile_results/h3-sdpa-kernel-matrix-20260825-run1/jobs/gpu5-RVA-cudnn-model_strided.trace.json"
}