[CUDA] Fuse MoE finalize into FC2 GEMV and skip expandInputRows - #29887

Merged

Tianlei Wu (tianleiwu) merged 10 commits into

mainfrom

tlwu/20260725/qmoe_sm120_opt

Jul 27, 2026

Commits on Jul 26, 2026

Fix import crash
Tianlei Wu (tianleiwu)
committed
2831ca5
Fuse MoE finalize into the FC2 GEMV epilogue
Tianlei Wu (tianleiwu)
committed
e50dfeb
Skip expandInputRows when the FC1 MoE GEMV runs
Tianlei Wu (tianleiwu)
committed
37d6609
Fuse MoE softmax/top-k into the expert-map prologue at decode
Tianlei Wu (tianleiwu)
committed
2a1d49b
Revert "Fuse MoE softmax/top-k into the expert-map prologue at decode"
Tianlei Wu (tianleiwu)
committed
67e67e1
Revert "Fix import crash"
Tianlei Wu (tianleiwu)
committed
f7261e2
Deduplicate the FC1 MoE GEMV disable predicate and cover multi-token decode
Tianlei Wu (tianleiwu)
committed
d5f06b8
Cover the MoE GEMV in tests and loosen the fused-finalize n alignment
Tianlei Wu (tianleiwu)
committed
0c3aaea
fix test: no zero point for symmetric test case
Tianlei Wu (tianleiwu)
committed
4b83993
lintrunner
Tianlei Wu (tianleiwu)
committed
4ebbc34