{
  "checked_utc": "2026-10-18T22:37:48.777833+00:00",
  "command": [
    "/tmp/course-pytorch-cpu-venv/bin/python",
    "Phase 4 - Track C - DL Inference Optimization/02 - Kernel Engineering/FlashAttention Course/labs/sdpa_experiment.py",
    "--dtype",
    "bfloat16"
  ],
  "result": {
    "torch": "2.14.0+cpu",
    "torch_cuda_build": null,
    "device": "cpu",
    "device_name": "CPU",
    "backend_requested": "math",
    "dtype": "bfloat16",
    "shape": [
      1,
      2,
      7,
      11,
      8
    ],
    "alignment": "lower_right",
    "dropout_p": 0,
    "seed": 17,
    "threads": 1,
    "oracle": "CPU math float64 on the same quantized inputs",
    "max_abs_error": 0.003368344611412777,
    "atol": 0.01,
    "rtol": 0.01,
    "passes_fixture_tolerance": true,
    "warmup": 5,
    "forward_ms_samples": [
      0.04007,
      0.04045,
      0.03693,
      0.03586,
      0.04808,
      0.041882,
      0.17899,
      0.06772,
      0.10123,
      0.04938,
      0.04769,
      0.0453,
      0.04506,
      0.04195,
      0.04394,
      0.0447,
      0.0451,
      0.041881,
      0.043221,
      0.043141
    ],
    "forward_ms_median": 0.04432,
    "timing_boundary": "forward SDPA including explicit mask; excludes allocation, transfer, reference and compilation warmup",
    "scope": "No FlashAttention distribution installed by this script; CPU numbers do not predict GPU speed or serving TPS."
  }
}
