deepseek-ai/DeepEP
1
0
Fork 0
mirror of https://github.com/deepseek-ai/DeepEP.git synced 2026-11-05 00:47:32 +00:00
Code Issues 308 Projects Packages Wiki Activity

[GH-ISSUE #567] question about Hybrid-EP #346

New issue
Open
opened 2026-10-26 09:58:56 +00:00 by bsakaguchi · 1 comment
bsakaguchi commented 2026-10-26 09:58:56 +00:00
Owner
Copy link

Originally created by @Thunderbrook on GitHub (Jan 19, 2026).
Original GitHub issue: https://github.com/deepseek-ai/DeepEP/issues/567

hello~@Autumn1998 I’d like to ask some questions about Hybrid-EP.

  1. i found that there has cudamemcpy from/to user hidden tensor to/from rdma registed buffer in Hybrid-EP. When using the permute version, the cudamemcpy in postprocess can be fused, but the preprocessing still remains. So i wonder if it can use zero copy in Hybrid-EP like this.
  2. There seem to be redundant stores here.
    thx~
Originally created by @Thunderbrook on GitHub (Jan 19, 2026). Original GitHub issue: https://github.com/deepseek-ai/DeepEP/issues/567 hello~@Autumn1998 I’d like to ask some questions about Hybrid-EP. 1. i found that there has cudamemcpy from/to user hidden tensor to/from rdma registed buffer in Hybrid-EP. When using the permute version, the cudamemcpy in postprocess can be fused, but the preprocessing still remains. So i wonder if it can use zero copy in Hybrid-EP like [this]( https://github.com/deepseek-ai/DeepEP/pull/79/files). 2. There seem to be redundant stores [here](https://github.com/deepseek-ai/DeepEP/blob/hybrid-ep/csrc/hybrid_ep/extension/permute.cu#L202). thx~
bsakaguchi commented 2026-10-26 09:58:57 +00:00
Author
Owner
Copy link

@Autumn1998 commented on GitHub (Jan 30, 2026):

@Thunderbrook Sorry for missing this issue:

  1. Theoretically, we could use zero-copy for the input part in the inter-node case. But this would make the op not standalone, that is., we would need to expose a tensor and let the attn side write the results directly into it, which we prefer not to do unless it causes a significant perf regression.

  2. What does “redundant store” refer to here? If you mean token per expert, after we apply padding we will recompute token per expert and write it again, and its meaning is different from the previous one.

<!-- gh-comment-id:3822602907 --> @Autumn1998 commented on GitHub (Jan 30, 2026): @Thunderbrook Sorry for missing this issue: 1. Theoretically, we could use zero-copy for the input part in the inter-node case. But this would make the op not standalone, that is., we would need to expose a tensor and let the attn side write the results directly into it, which we prefer not to do unless it causes a significant perf regression. 2. What does “redundant store” refer to here? If you mean token per expert, after we apply padding we will recompute token per expert and write it again, and its meaning is different from the previous one.
Sign in to join this conversation.
No Branch/Tag specified
Branches Tags
main
hybrid-ep
chore/remove-clang-format-yapf
v2.5
nvDev
try-fix-cu128
mori-ep
antgroup-opt
compat-old-torch
add-linter
eager-rdma
v1.2.1
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/DeepEP#346
No description provided.