[NVIDIA] Add flashinfer MNNVL backend for allreduce only - #30700

Merged

Fridge003 merged 14 commits into

sgl-project:mainfrom

wenscarl:flashinfer-pure-allreduce

Aug 11, 2026

Commits on Aug 5, 2026

[comm] don't fuse the post-experts all-reduce under hybrid EP+TP

wenscarl
and
claude
committed
6e82c20
[comm] add FlashInfer kAllReduce backend for pure TP allreduce

wenscarl
and
claude
committed
f8b9b5b
[comm] auto-enable FlashInfer pure allreduce for DSV3/V4 archs

wenscarl
and
claude
committed
3afd96e
[comm] limit pure allreduce auto-enable to DSV3/V4 archs only

wenscarl
and
claude
committed
866ac0f
[comm] add DeepseekV4ForCausalLM to flashinfer allreduce auto-enable archs

wenscarl
and
claude
committed
26235c1
[comm] mark enable_flashinfer_pure_allreduce as resolvable

wenscarl
and
claude
committed
ad32f0d
Fix lint
wenscarl
committed
bed1d4b
Fold FlashInfer allreduce-only tests into comm fusion
wenscarl
committed
d83d8c2
Fix lint
wenscarl
committed
4a9126e
[comm] gate FlashInfer kAllReduce on the allreduce-fusion flag itself

wenscarl
and
claude
committed
1474749
[comm] bind the kAllReduce workspace to the group that owns it

wenscarl
and
claude
committed
1a43d9b
[comm] wrap kAllReduce as a custom op so it works under piecewise CUDA graph

wenscarl
and
claude
committed
8141a0d
Upd
wenscarl
committed
aead319

Commits on Aug 10, 2026

Merge branch 'main' into flashinfer-pure-allreduce
b8zhong
authored
a2625dd