{
  "schema_version": 1,
  "repository": "https://github.com/fattorib/hawk-pytorch",
  "commit": "e43239336cf93b6dd79486447150429e472a4b90",
  "source_file": "/root/autodl-tmp/samu-gpu-prior-art/hawk-pytorch/hawk/scan_fused.py",
  "source_sha256": "4dd6e0a1044a19d12b731f96602a8ffc6ea73b5b6e6b4727877bcd7dfe046c3d",
  "source_diff_empty": true,
  "environment": {
    "torch": "2.4.1+cu124",
    "cuda": "12.4",
    "gpu": "NVIDIA H800 PCIe"
  },
  "shape": {
    "batch": 4,
    "length": 2048,
    "width": 2048
  },
  "protocol": {
    "scope": "official block-diagonal gate projections + activation/write/recurrence",
    "dtype": "BF16 autocast, FP32 recurrent carry",
    "order": "AB/BA",
    "compile_excluded": true,
    "semantic_scope": "restricted no-reset recurrence; secondary comparison only",
    "primary_canonical_ranking": "NOT COMPARABLE because public Fattori source lacks official reset/write normalization, arbitrary segment resets and initial state"
  },
  "correctness": {
    "contract": "zero initial state, no resets, length>=2, width multiple of 64",
    "canonical_reset_semantics": "NOT SUPPORTED by the public kernel",
    "output_max_abs": 0.00048828125,
    "gradient_errors": {
      "x": 9.5367431640625e-07,
      "input_gate_logits": 0.0001220703125,
      "recurrence_gate_logits": 0.0009765625,
      "a_param": 0.000534355640411377
    },
    "gradient_relative_errors": {
      "x": 3.372108080110497e-07,
      "input_gate_logits": 0.0001220703125,
      "recurrence_gate_logits": 0.0009765625,
      "a_param": 0.000534355640411377
    },
    "passed_restricted_contract": true
  },
  "rows": [
    {
      "implementation": "fattori_original",
      "forward": {
        "samples_ms": [
          1.9718557596206665,
          1.9807741045951843,
          1.9840523600578308,
          1.9603483378887177,
          1.958109438419342,
          1.9621960818767548,
          1.9450485706329346,
          1.9761733710765839,
          1.951415091753006,
          1.9473806023597717
        ],
        "median_ms": 1.9612722098827362,
        "minimum_ms": 1.9450485706329346,
        "maximum_ms": 1.9840523600578308,
        "mean_ms": 1.9637353718280792
      },
      "forward_backward": {
        "samples_ms": [
          4.543386399745941,
          4.5276060700416565,
          4.718799144029617,
          4.585329443216324,
          4.526209086179733,
          4.523616284132004,
          4.503756761550903,
          4.6143196523189545,
          7.7250562608242035,
          4.520058631896973
        ],
        "median_ms": 4.535496234893799,
        "minimum_ms": 4.503756761550903,
        "maximum_ms": 7.7250562608242035,
        "mean_ms": 4.878813773393631
      },
      "rounds": [
        {
          "round": 0,
          "order_index": 0,
          "forward_ms": [
            1.9718557596206665,
            1.9807741045951843,
            1.9840523600578308,
            1.9603483378887177,
            1.958109438419342
          ],
          "forward_backward_ms": [
            4.543386399745941,
            4.5276060700416565,
            4.718799144029617,
            4.585329443216324,
            4.526209086179733
          ]
        },
        {
          "round": 1,
          "order_index": 1,
          "forward_ms": [
            1.9621960818767548,
            1.9450485706329346,
            1.9761733710765839,
            1.951415091753006,
            1.9473806023597717
          ],
          "forward_backward_ms": [
            4.523616284132004,
            4.503756761550903,
            4.6143196523189545,
            7.7250562608242035,
            4.520058631896973
          ]
        }
      ],
      "peak_allocated_delta_bytes": 335553536
    },
    {
      "implementation": "ours_restricted",
      "forward": {
        "samples_ms": [
          0.9020045399665833,
          0.8786581456661224,
          0.8766762912273407,
          0.8663460612297058,
          0.8712522685527802,
          0.878017395734787,
          0.8702017366886139,
          0.8669905364513397,
          0.868476927280426,
          0.8705258369445801
        ],
        "median_ms": 0.8708890527486801,
        "minimum_ms": 0.8663460612297058,
        "maximum_ms": 0.9020045399665833,
        "mean_ms": 0.8749149739742279
      },
      "forward_backward": {
        "samples_ms": [
          2.280615270137787,
          2.261284738779068,
          2.2507570683956146,
          2.261560410261154,
          2.2525861859321594,
          2.2665373980998993,
          2.257082611322403,
          2.2590458393096924,
          2.247340977191925,
          2.2663213312625885
        ],
        "median_ms": 2.26016528904438,
        "minimum_ms": 2.247340977191925,
        "maximum_ms": 2.280615270137787,
        "mean_ms": 2.260313183069229
      },
      "rounds": [
        {
          "round": 0,
          "order_index": 1,
          "forward_ms": [
            0.9020045399665833,
            0.8786581456661224,
            0.8766762912273407,
            0.8663460612297058,
            0.8712522685527802
          ],
          "forward_backward_ms": [
            2.280615270137787,
            2.261284738779068,
            2.2507570683956146,
            2.261560410261154,
            2.2525861859321594
          ]
        },
        {
          "round": 1,
          "order_index": 0,
          "forward_ms": [
            0.878017395734787,
            0.8702017366886139,
            0.8669905364513397,
            0.868476927280426,
            0.8705258369445801
          ],
          "forward_backward_ms": [
            2.2665373980998993,
            2.257082611322403,
            2.2590458393096924,
            2.247340977191925,
            2.2663213312625885
          ]
        }
      ],
      "peak_allocated_delta_bytes": 385950720
    }
  ]
}