{
  "device": "NVIDIA GB10",
  "compute_capability": "SM121",
  "cutlass_dsl": "4.6.2",
  "rows": 128,
  "kernel": "official persistent cooperative SM120/SM121 block-scaled GEMM",
  "results": [
    {
      "linear": "attn_qkv_proj",
      "mnk": [128, 21504, 5376],
      "raw_blockscaled_equal": true,
      "raw_max_abs": 0.0,
      "fused_alpha_equal": true,
      "fused_alpha_max_abs": 0.0
    },
    {
      "linear": "attn_out_proj",
      "mnk": [128, 5376, 7168],
      "raw_blockscaled_equal": true,
      "raw_max_abs": 0.0,
      "fused_alpha_equal": true,
      "fused_alpha_max_abs": 0.0
    },
    {
      "linear": "mlp_fc1",
      "mnk": [128, 28672, 5376],
      "raw_blockscaled_equal": true,
      "raw_max_abs": 0.0,
      "fused_alpha_equal": true,
      "fused_alpha_max_abs": 0.0
    },
    {
      "linear": "mlp_fc2",
      "mnk": [128, 5376, 14336],
      "raw_blockscaled_equal": false,
      "raw_max_abs": 8192.0,
      "raw_mean_abs": 0.01562502235174179,
      "fused_alpha_equal": false,
      "fused_alpha_max_abs": 16.0,
      "fused_alpha_mean_abs": 0.000024163342459360138,
      "tile_k_128_and_256_identical": true
    }
  ],
  "findings": [
    "Comfy E2M1 values and E4M3 scale layouts are interoperable with CUTLASS after logical FP4 re-encoding.",
    "The raw block-scaled operation is bit-exact for QKV, attention output, and FC1.",
    "The custom alpha-before-BF16 epilogue is bit-exact for QKV, attention output, and FC1.",
    "FC2 still requires matching the reference GEMM reduction policy; changing the stock K tile from 128 to 256 does not restore equality.",
    "The official stock Float32 epilogue failed its own GB10 reference check and is not a valid accumulation oracle."
  ],
  "status": "P0 remains active; do not implement the streaming A producer until the exact alpha epilogue and FC2 reduction contract pass.",
  "source_artifacts": [
    "benchmarks/gb10-cute-p0-attn-qkv-128rows.json",
    "benchmarks/gb10-cute-p0-attn-out-128rows.json",
    "benchmarks/gb10-cute-p0-mlp-fc1-128rows.json",
    "benchmarks/gb10-cute-p0-mlp-fc2-128rows.json",
    "benchmarks/gb10-cute-p0-attn-qkv-128rows-alpha.json",
    "benchmarks/gb10-cute-p0-attn-out-128rows-alpha.json",
    "benchmarks/gb10-cute-p0-mlp-fc1-128rows-alpha.json",
    "benchmarks/gb10-cute-p0-mlp-fc2-128rows-alpha.json"
  ]
}