Integrate the CuTe DSL allreduce fusion from FlashInfer and tune for GLM-5.2 - #34134

Closed

b8zhong wants to merge 9 commits into

mainfrom

cutedsl-allreduce-fusion

Commits on Aug 7, 2026

[Comm] Add the MNNVL CuTe DSL allreduce fusion backend
Brayden Zhong
committed
2f96202
[Comm] Don't drop the all-reduce when a fused RMSNorm backend declines
Brayden Zhong
committed
2c5429d

Commits on Aug 8, 2026

[Comm] Tune the CuTe DSL allreduce fusion for GLM-5.2 on B300
Brayden Zhong
committed
997823e
Revert "[Comm] Tune the CuTe DSL allreduce fusion for GLM-5.2 on B300"
Brayden Zhong
committed
af84e49
[Comm] Restrict the CuTe DSL backend to the MoE output group on request
Brayden Zhong
committed
eab9fd9
Reapply "[Comm] Tune the CuTe DSL allreduce fusion for GLM-5.2 on B300"
Brayden Zhong
committed
bb2bd9d

Commits on Aug 10, 2026

[Comm] Add env overrides for A/B testing the CuTe DSL AR fusion tuning
Brayden Zhong
committed
90cc479

Commits on Aug 11, 2026

[Comm] Always route both fusion groups through the CuTe DSL backend
Brayden Zhong
committed
b4b6047

Commits on Aug 12, 2026

[Comm] Trim CuTe DSL fusion helpers, comments, and the shard-derivation test
Brayden Zhong
committed
f188513