SM90 (Hopper) FP4 MegaMoE fused kernel with swapAB small-batch path#53

Open

qiushixiaoyu wants to merge 7 commits into

sgl-project:devfrom

qiushixiaoyu:fp4_swapAB

Commits on Jul 19, 2026

Add SM90 FP4 MegaMoE implementation
yinding
committed
Avoid L1 N32 swapAB bucket for FP4 MegaMoE
yinding
committed
FP4 swapAB L1: align routing-weight fold order with non-swap path

yinding
and
claude
committed
Fix SM90 FP4 MegaMoE build against post-#364 signatures
yinding
committed
Simplify SM90 FP4 MegaMoE heuristics to FP8 parity

yinding
and
claude
committed
Raise SM90 FP4 MegaMoE decode->prefill boundary to e=80

yinding
and
claude
committed
Use continuous FP32 activation scale in SM90 FP4 MegaMoE

yinding
and
claude
committed