[SM90] Optimize FP8 MegaMoE: fused shared expert, interleaved scheduler, L2 epilogue, 4WG heuristic and decode swap-AB - #69

Open

Helicoplacoidea wants to merge 10 commits into

sgl-project:devfrom

Helicoplacoidea:sgl

Commits on Aug 30, 2026

Add NVFP4 mega MoE support (#80)
zcnrex
authored
b2aece8

Commits on Aug 31, 2026

fix and add cache hint
zcnrex
committed
1693c4e
Merge pull request #81 from zcnrex/megamoe-nvfp4-fix
zcnrex
authored
ef19220
Rebase SM90 MegaMoE producer-warp scheduler onto current dev

authored and
Helicoplacoidea
committed
ea1e1f4
fix(python): remove duplicate num_shared_experts param in get_symm_buffer_for_mega_moe

authored and
Helicoplacoidea
committed
573a089
fix(ptx): drop duplicate ld_volatile(const uint32_t*) in ld_st.cuh

authored and
Helicoplacoidea
committed
1c2570a
refactor(scheduler): unify SM90 onto the producer-warp MegaMoEScheduler

authored and
Helicoplacoidea
committed
b2bf07d
refactor(sm90): query BlockPhase via TaskInfo methods, drop the free helpers

authored and
Helicoplacoidea
committed
9d7c001

Commits on Sep 1, 2026

refactor(scheduler): remove dead LegacyMegaMoEScheduler

authored and
Helicoplacoidea
committed
5c030d3
refactor(sm90): remove unused SwiGLU/clamp/silu template helpers

authored and
Helicoplacoidea
committed
220a105