feat: MXFP8 epilogue fusion for SM100 (B200)#1

Open

wanyaworld wants to merge 4 commits into

mainfrom

gemm_mxfp8_out_tma

Commits on Apr 10, 2026

feat: MXFP8 epilogue fusion — gmem direct write

authored and
Your Name
committed
perf: replace gmem direct write with TMA store in MXFP8 epilogue

authored and
Your Name
committed
test: fair comparison with compiled graph baseline (fullgraph=True)

authored and
Your Name
committed
chore: remove dev artifacts, keep only impl + test + yaml

authored and
Your Name
committed