h3-blackwell-runtime/benchmarks/gb10-sage2-p2-vprep-analysis.json
2026-08-25 20:30:22 +07:00

51 lines
1.7 KiB
JSON

{
"status": "rejected",
"reason": "The exact candidate missed the 6.0 ms isolated go threshold and cannot reliably clear the 1% complete-block gate.",
"device": "NVIDIA GB10",
"compute_capability": "SM121",
"candidate": "Direct strided-NHD BF16 V to Sage2 padded/permuted FP8 V preparation",
"contract": {
"tensor_layout": "NHD",
"head_dimension": 128,
"scale_max": 2.25,
"mainloop": "unchanged SageAttention 2.2.0 SM89 INT8-QK/FP8-PV mainloop"
},
"correctness": {
"sequence_lengths": [1, 31, 32, 33, 63, 64, 65, 127, 128, 129, 37760, 37761, 37810],
"heads": 56,
"fp8_v": "byte-exact",
"fp32_v_scale": "byte-exact"
},
"canonical_isolated_timing": {
"sequence": 37810,
"warmup": 5,
"iterations": 20,
"baseline_p50_ms": 10.56272,
"candidate_p50_ms": 6.392816,
"preparation_improvement_percent": 39.477,
"absolute_saving_ms": 4.169904,
"complete_block_baseline_ms": 458.78,
"projected_complete_block_improvement_percent": 0.909
},
"gate": {
"go_p50_ms": 6.0,
"stop_p50_ms": 6.5,
"result": "reject after the required borderline repeat",
"complete_block_benchmark": "skipped",
"trajectory_validation": "skipped",
"runtime_integration": "not attempted"
},
"implementation": {
"stages": [
"64-row per-channel BF16 absolute maxima",
"Sage-compatible approximate-reciprocal FP32 scale finalization",
"direct E4M3 quantization into Sage2's padded and sequence-permuted layout"
],
"intermediate_bf16_transpose": "removed",
"production_dispatch": "unchanged"
},
"source_artifacts": [
"benchmarks/gb10-sage2-vprep-randomized.json",
"benchmarks/gb10-sage2-vprep-canonical-repeat.json"
]
}