DeepSpeed
Add Triton grouped-GEMM for MoE experts on Ampere/Ada
#8180

Merged

Add Triton grouped-GEMM for MoE experts on Ampere/Ada #8180

hwchen2017 merged 13 commits into master from hongwei/group_gemm

hwchen2017 requested a review from

delock 36 days ago

hwchen2017 requested a review from

sfc-gh-truwase 36 days ago

hwchen2017 requested a review from

tohtana 36 days ago

hwchen2017 requested a review from

tjruwase 36 days ago

hwchen2017 requested a review from

loadams 36 days ago

chatgpt-codex-connector commented on 2026-07-26

delock commented on 2026-07-28

A good working version for group gemm

123c0309

Refactor code

e7b962dd

change autotune config

a815d4cb

Update comments

0699071e

Add fused triton kernel for some elementwise kernel

ab3f2392

Add benchmark code

f09c1da9

Fix format error, and add doc

584236a0

Fix format error, address reviewer comments

ca2090ba

hwchen2017 force pushed from 264b3990 to ca2090ba 33 days ago

sfc-gh-truwase commented on 2026-07-30

Fix review comments

ed93495e

delock approved these changes on 2026-07-31

Replace disable_triton_group_mm env variabel with config

58ddff65

Skip CUDA forward tests when transformer op is unavailable

51aa775a

Revert "Skip CUDA forward tests when transformer op is unavailable"

1cca5431

hwchen2017 force pushed from fd3a792e to 1cca5431 31 days ago

tohtana requested changes on 2026-08-01

Address review comments

5ac1ad3a

hwchen2017 commented on 2026-08-03

tohtana approved these changes on 2026-08-03

hwchen2017 merged df84f6d8 into master 28 days ago

hwchen2017 deleted the hongwei/group_gemm branch 28 days ago

Reviewers

tohtana

delock

sfc-gh-truwase

chatgpt-codex-connector

tjruwase

loadams

Assignees

No one assigned

Labels

None yet

Milestone

No milestone

DeepSpeed Add Triton grouped-GEMM for MoE experts on Ampere/Ada #8180 Merged

Add Triton grouped-GEMM for MoE experts on Ampere/Ada #8180

DeepSpeed
Add Triton grouped-GEMM for MoE experts on Ampere/Ada
#8180

Merged