feat(moe): add opt-in FlashInfer AlphaMoE W8A8 and NVFP4 backends - #34072

Open

yyihuang wants to merge 22 commits into

sgl-project:mainfrom

yyihuang:codex/flashinfer-alphamoe-integration-b61a

Commits on Aug 8, 2026

feat(moe): integrate FlashInfer AlphaMoE W8A8 backend

averyhNV
authored and
Audit
committed
998b391
fix(moe): gate AlphaMoE to validated Qwen TP4 geometry

averyhNV
authored and
Audit
committed
6c30810
docs(moe): document experimental AlphaMoE backend
Audit
committed
db4a25d
style(moe): format AlphaMoE adapter
Audit
committed
bba28dc
fix(moe): require TP4 for AlphaMoE
Audit
committed
fecb3dc
fix(moe): reject unsupported AlphaMoE execution modes
Audit
committed
2bba2d7
feat(moe): integrate FlashInfer AlphaMoE W8A8 backend
averyhNV
committed
ecc4452
fix(moe): gate AlphaMoE to validated Qwen TP4 geometry
averyhNV
committed
8454b5c
docs(moe): document experimental AlphaMoE backend

Audit
authored and
averyhNV
committed
2ed82ed
style(moe): format AlphaMoE adapter

Audit
authored and
averyhNV
committed
ea0e6f6
fix(moe): require TP4 for AlphaMoE

Audit
authored and
averyhNV
committed
e71d40e
fix(moe): reject unsupported AlphaMoE execution modes

Audit
authored and
averyhNV
committed
7aa8ef7
feat(moe): route ModelOpt FP4 through FlashInfer AlphaMoE
averyhNV
committed
0a552ae
fix(moe): close unsupported AlphaMoE modes
averyhNV
committed
39dc354
test(moe): arm runtime shape tracing after warmup
averyhNV
committed
a46ef9e

Commits on Sep 12, 2026