[Main] Support paged stash with device-initiated GroupedLinear - #6828

Open

lhb8125 wants to merge 12 commits into

NVIDIA:mainfrom

lhb8125:denliu/paged-stash-grouped-tensor

Commits on Aug 25, 2026

feat: support paged stash with GroupedTensor experts
lhb8125
committed
98b860d
refactor: use TE paged stash markers
lhb8125
committed
2601f76
refactor: require fused activations for paged stash
lhb8125
committed
e11a555
refactor: minimize paged stash integration diff
lhb8125
committed
4860c5b
refactor: inline non-fused grouped MLP forward
lhb8125
committed
6c03cfe
Revert "refactor: inline non-fused grouped MLP forward"
lhb8125
committed
16ae301
refactor: narrow paged stash marker propagation
lhb8125
committed
2a54634
refactor: route paged stash markers through TE adapter
lhb8125
committed
cc3cc7c
docs: trim paged stash implementation details
lhb8125
committed
660e7a3
Merge branch 'main' into denliu/paged-stash-grouped-tensor
lhb8125
authored
df35d24
test: gate grouped-tensor paged stash on TE support
lhb8125
committed
d409603

Commits on Aug 27, 2026

fix(moe): exclude recomputed FC2 inputs from paged stash
lhb8125
committed
9ab53a2