{
  "device": "NVIDIA GB10",
  "torch": "2.9.1+cu130",
  "cutlass_dsl": "4.6.2",
  "metadata": {
    "width": 1344,
    "height": 768,
    "frames": 124,
    "steps": 12,
    "sampler_step": 1,
    "seed": 440420,
    "text_tokens": 100,
    "hidden_shape": [
      37810,
      5376
    ],
    "segments": [
      [
        0,
        100,
        1
      ],
      [
        100,
        514,
        2
      ],
      [
        514,
        37810,
        0
      ]
    ]
  },
  "linear": "attn_out_proj",
  "rows": 128,
  "mnk": [
    128,
    5376,
    7168
  ],
  "tile_shape_mnk": [
    128,
    128,
    128
  ],
  "swap_nibbles": false,
  "reencode_fp4": true,
  "zero_a": false,
  "cute_shapes": {
    "a": "(128, 7168, 1)",
    "b": "(5376, 7168, 1)",
    "sfa": "(128, 448, 1)",
    "sfb": "(5376, 448, 1)",
    "c": "(128, 5376, 1)"
  },
  "tensor_scales": {
    "a": 0.0263671875,
    "b": 0.00072479248046875,
    "alpha": 1.911073923110962e-05
  },
  "raw_output": {
    "dtype": "torch.bfloat16",
    "checksum": 13503609856.0,
    "max_abs": 71827456.0,
    "nonzero": 688128,
    "finite": true
  },
  "raw_blockscaled_parity": {
    "reference_checksum": 13503609856.0,
    "candidate_checksum": 13503609856.0,
    "equal": true,
    "max_abs": 0.0,
    "mean_abs": 0.0
  },
  "reference_checksum": 257945.5,
  "candidate_checksum": 258224.125,
  "equal": false,
  "max_abs": 8.0,
  "mean_abs": 0.07462421804666519,
  "relative_l2": 0.0028051051776856184,
  "numerical_note": "The stock SM121 kernel's validated BF16 epilogue rounds before the external global-scale product. Exact H3 integration requires applying alpha in a custom epilogue before BF16 conversion."
}