{
  "date": "2026-08-23",
  "device": "NVIDIA GB10",
  "cutlass_dsl": "4.6.2",
  "tile_shape_mnk": [128, 128, 128],
  "producer_threads": 32,
  "rows_per_producer_thread": 4,
  "activation_input": "BF16",
  "global_activation_qdata": false,
  "global_activation_sfa": false,
  "retained_tma_inputs": ["B", "SFB"],
  "streamed_roles": {
    "attn_qkv_proj": {
      "mnk": [128, 21504, 5376],
      "packed_fp4_differences": 0,
      "block_scale_differences": 0,
      "bf16_output_equal": true,
      "max_abs": 0.0,
      "mean_abs": 0.0,
      "relative_l2": 0.0
    },
    "attn_out_proj": {
      "mnk": [128, 5376, 7168],
      "packed_fp4_differences": 0,
      "block_scale_differences": 0,
      "bf16_output_equal": true,
      "max_abs": 0.0,
      "mean_abs": 0.0,
      "relative_l2": 0.0
    },
    "mlp_fc1": {
      "mnk": [128, 28672, 5376],
      "packed_fp4_differences": 0,
      "block_scale_differences": 0,
      "bf16_output_equal": true,
      "max_abs": 0.0,
      "mean_abs": 0.0,
      "relative_l2": 0.0
    }
  },
  "mlp_fc2": {
    "streaming_enabled": false,
    "fallback": "Comfy/cuBLAS",
    "validator_rejects_streaming": true
  },
  "numerical_contract": {
    "block_scale_order": "(amax / 6.0f) / global_scale",
    "block_scale_clamp": "fminf(value, 448.0f)",
    "encode_scale": "fminf(rcp.approx.ftz.f32(rounded_scale * global_scale), FLT_MAX)",
    "fp4_conversion": "CuTe E2M1 round-to-nearest ties-to-even",
    "epilogue": "FP32 global alpha before BF16 conversion"
  },
  "status": "Fixed 128-row P1 producer-consumer checkpoint is exact; not runtime-integrated or performance-approved."
}