62 lines
2.3 KiB
JSON
62 lines
2.3 KiB
JSON
{
|
|
"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"
|
|
]
|
|
}
|