KernelBench hard · RTX PRO 6000

FP8 GEMM Qwen 3.8 Max

manually audited: clean

The scored implementation is genuine FP8 computation. For M > 64 it invokes a CUTLASS 3 SM120 f8f6f4 tensor-core GEMM with FP32 accumulation and a fused per-output-channel scale/bf16 epilogue; if the extension is unavailable it runs a Triton tl.dot GEMM. M <= 64 uses a dedicated Triton tl.dot kernel. Every call allocates a fresh output and launches a GEMM; there is no constant output, output memoization, CUDA graph replay, reference call, forbidden vendor op, or grader/stack sniffing. The only semantic cache is an input transformation cache on the odd-K padding path: _pad_cache reuses padded copies when the source tensor has the same Python identity and _version. A same-buffer overwrite test on the canonical odd-K shape confirmed that both x.copy_ and weight.copy_ advance the tracked versions, invalidate stale padded copies, recompute changed outputs, and remain reference-correct. Therefore the implementation verdict is clean. The subsequent publish-grade sequential isolated regrade passed with correct=true and peak_fraction=0.3269.

20260803_034401_qwen-claude_qwen3.8-max_01_fp8_gemm