KernelBench hard · RTX PRO 6000

FP8 GEMM qwen3.8-max-preview

manually audited: clean

Genuine fp8 x fp8 Triton tl.dot GEMM with fp32 accumulate, per-output-channel scale epilogue, bf16 store. Autotuned config space (BM/BN/BK up to 128x256x128, num_stages 3-4) with an M-bucketed early_config_prune splitting decode-skinny from fat tiles, and grouped L2 swizzle. Off-alignment K (4127) is zero-padded to a multiple of 128 so the K loop is even and unmasked (TMA/cp.async-friendly); x is copied into its pad buffer EVERY call, only the padded weight is cached, invalidated via weight._version. Empirically verified to recompute on in-place input and weight mutation. Transcript contains only its own run id; KBH_NUMERIC_STRESS only ever set to =1 (extra strictness). 0.3729 geomean in-run, honest.

20260719_141338_qwen-claude_qwen3.8-max-preview_01_fp8_gemm