{
  "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,
  "fuse_alpha": true,
  "stream_a": true,
  "streamed_activation_materialization": {
    "global_qdata": false,
    "global_sfa": false
  },
  "timing": {
    "warmup": 5,
    "iterations": 20,
    "total_ms": 29.38243293762207,
    "mean_ms": 1.4691216468811035
  },
  "cute_shapes": {
    "a": "(128, 7168, 1)",
    "b": "(5376, 7168, 1)",
    "sfa": "(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": 257945.5,
    "max_abs": 1368.0,
    "nonzero": 688128,
    "finite": true
  },
  "raw_blockscaled_parity": {
    "applicable": false,
    "reference_checksum": 13503609856.0,
    "candidate_checksum": 257945.5,
    "equal": false,
    "max_abs": 71826088.0,
    "mean_abs": 2754388.25
  },
  "reference_checksum": 257945.5,
  "candidate_checksum": 257945.5,
  "equal": true,
  "max_abs": 0.0,
  "mean_abs": 0.0,
  "relative_l2": 0.0,
  "numerical_note": "The experimental epilogue applies the FP32 global-scale product before BF16 conversion."
}