QMoE: prepack int4/int8 expert weights in PrePack hook (symmetric with MatMulNBits)#28749

Merged

justinchuby merged 13 commits into

microsoft:mainfrom

justinchuby:qmoe-int-prepack

Jun 9, 2026

Commits on Jun 2, 2026

QMoE: prepack int4/int8 expert weights in PrePack hook

justinchuby
and
Copilot
committed
Add TestQMoEIntPrePackParity for the new QMoE.PrePack hook

justinchuby
and
Copilot
committed
Fix import order in test_qmoe_cuda.py
justinchuby
committed
Rewrite QMoE PrePack test as a smoke test; verified passing on H200
justinchuby
committed
QMoE: address review — opt-in weights_prepacked attribute, SM guard, drop redundant assert

justinchuby
and
Copilot
committed
QMoE: cache weight shapes during PrePack so source initializer can be freed

justinchuby
and
Copilot
committed

Commits on Jun 3, 2026

Commits on Jun 4, 2026

Commits on Jun 8, 2026