DeepSpeed
MoE dispatch: fix silent Tutel + TP corruption, speed up native and Tutel paths
#8195

Merged

MoE dispatch: fix silent Tutel + TP corruption, speed up native and Tutel paths #8195

tohtana merged 6 commits into deepspeedai:master from jinyouzhi:moe

Speed up Tutel route extraction for k > 2

3ae3b5af

Shard the Tutel dispatch buffer along capacity under tensor parallelism

9157192a

jinyouzhi requested a review from

tohtana 13 days ago

jinyouzhi requested a review from

tjruwase 13 days ago

jinyouzhi requested a review from

loadams 13 days ago

chatgpt-codex-connector commented on 2026-07-29

Dispatch MoE tokens by index instead of a dense one-hot einsum

db44b2a9

jinyouzhi force pushed from 28353d19 to db44b2a9 13 days ago

Merge branch 'master' into moe

0b8e3541

sfc-gh-truwase commented on 2026-08-02

jinyouzhi force pushed from f90bc32f to 36d58037 9 days ago

Merge remote-tracking branch 'origin/master' into HEAD

c7efd5f7

Move MoE unit tests to tests/unit/v1/moe

36d58037

tohtana approved these changes on 2026-08-02

tohtana enabled auto-merge 9 days ago

tohtana merged 619e3199 into master 9 days ago

jinyouzhi deleted the moe branch 6 days ago

Reviewers

tohtana

sfc-gh-truwase

tjruwase

loadams

Assignees

No one assigned

Labels

None yet

Milestone

No milestone