deepseek-ai/DeepGEMM
1
0
Fork 0
mirror of https://github.com/deepseek-ai/DeepGEMM.git synced 2026-11-05 00:47:26 +00:00
Code Issues 162 Projects Packages Wiki Activity

[GH-ISSUE #366] [Feature] Will DeepGEMM support more advanced epilogue fusion for GEMM kernels? #176

New issue
Closed
opened 2026-10-26 21:14:13 +00:00 by bsakaguchi · 2 comments
bsakaguchi commented 2026-10-26 21:14:13 +00:00
Owner
Copy link

Originally created by @b8zhong on GitHub (Jun 24, 2026).
Original GitHub issue: https://github.com/deepseek-ai/DeepGEMM/issues/366

https://github.com/deepseek-ai/DeepGEMM/pull/347 has support BF16 accumulation, which can help a lot for large problem sizes. However, I'm wondering if DeepGEMM will consider support direct FP8 output GEMMs in the future and potentially other activation functions too?

Originally created by @b8zhong on GitHub (Jun 24, 2026). Original GitHub issue: https://github.com/deepseek-ai/DeepGEMM/issues/366 https://github.com/deepseek-ai/DeepGEMM/pull/347 has support BF16 accumulation, which can help a lot for large problem sizes. However, I'm wondering if DeepGEMM will consider support direct FP8 output GEMMs in the future and potentially other activation functions too?
bsakaguchi commented 2026-10-26 21:14:15 +00:00
Author
Owner
Copy link

@hiSandog commented on GitHub (Jul 2, 2026):

A useful way to frame this feature request may be to separate API surface from kernel support. Direct FP8 output has different accuracy and scaling implications from BF16 accumulation, while activation fusion adds epilogue shape/layout constraints. If the maintainers are open to it, a target matrix such as accumulator dtype -> output dtype -> supported epilogues would make it clearer which combinations are intended to stay in DeepGEMM versus be handled by a follow-up kernel.

<!-- gh-comment-id:4861721862 --> @hiSandog commented on GitHub (Jul 2, 2026): A useful way to frame this feature request may be to separate API surface from kernel support. Direct FP8 output has different accuracy and scaling implications from BF16 accumulation, while activation fusion adds epilogue shape/layout constraints. If the maintainers are open to it, a target matrix such as `accumulator dtype -> output dtype -> supported epilogues` would make it clearer which combinations are intended to stay in DeepGEMM versus be handled by a follow-up kernel.
bsakaguchi commented 2026-10-26 21:14:16 +00:00
Author
Owner
Copy link

@b8zhong commented on GitHub (Sep 16, 2026):

Solved by https://github.com/deepseek-ai/DeepGEMM/pull/432

<!-- gh-comment-id:5700893322 --> @b8zhong commented on GitHub (Sep 16, 2026): Solved by https://github.com/deepseek-ai/DeepGEMM/pull/432
Sign in to join this conversation.
No Branch/Tag specified
Branches Tags
main
public-sync-20261014
0930-release
nv_dev
fix-mega-moe-fence
public-release-2609
public-release-2607
public-release-2606
mega-update
public-release-260416
ko3n1g/ci/build-wheels
nv_dev_f8e8fb5
nv_dev_950d31a
nv_dev_67fc648
nv_dev_c491439
nv_dev_4ff3f54
v2.1.1.post3
v2.1.1.post2
v2.1.1.post1
v2.1.1
v2.1.0
Labels
Clear labels   
pull-request
Mirrored from GitHub Pull Request

No labels
pull-request
Milestone
Clear milestone
No items
No milestone
Projects
Clear projects
No items
No project
Assignees
Clear assignees
No assignees
1 participant
Notifications
Due date
The due date is invalid or out of range. Please use the format "yyyy-mm-dd".

No due date set.

Dependencies

No dependencies set.

Reference
deepseek-ai/DeepGEMM#176
No description provided.