[GH-ISSUE #281] performance using NCCL #202

Closed
opened 2026-10-26 09:56:33 +00:00 by bsakaguchi · 3 comments
Owner

Originally created by @jing-4369 on GitHub (Jul 6, 2025).
Original GitHub issue: https://github.com/deepseek-ai/DeepEP/issues/281

I'm curious about how is the performance will be if I use NCCL' alltoall or send/recv to implement EP in pure python? have you ever tried? Thanks!

Originally created by @jing-4369 on GitHub (Jul 6, 2025). Original GitHub issue: https://github.com/deepseek-ai/DeepEP/issues/281 I'm curious about how is the performance will be if I use NCCL' alltoall or send/recv to implement EP in pure python? have you ever tried? Thanks!
Author
Owner

@sphish commented on GitHub (Jul 10, 2025):

https://github.com/ppl-ai/pplx-kernels has compared the performance of PyTorch all-to-all and DeepEP. You can refer to it.

<!-- gh-comment-id:3056285167 --> @sphish commented on GitHub (Jul 10, 2025): https://github.com/ppl-ai/pplx-kernels has compared the performance of PyTorch all-to-all and DeepEP. You can refer to it.
Author
Owner

@jing-4369 commented on GitHub (Jul 10, 2025):

Thanks!

<!-- gh-comment-id:3058157040 --> @jing-4369 commented on GitHub (Jul 10, 2025): Thanks!
Author
Owner

@polarstormx commented on GitHub (Sep 18, 2026):

Closing as answered: the requested performance comparison was provided and acknowledged by the author.

<!-- gh-comment-id:5727891400 --> @polarstormx commented on GitHub (Sep 18, 2026): Closing as answered: the requested performance comparison was provided and [acknowledged by the author](https://github.com/deepseek-ai/DeepEP/issues/281#issuecomment-3058157040).
Sign in to join this conversation.
No labels
pull-request
No milestone
No project
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#202
No description provided.