[GH-ISSUE #449] num_qps_per_rank equal to the number of local experts. #294

Closed
opened 2026-10-26 09:58:07 +00:00 by bsakaguchi · 1 comment
Owner

Originally created by @MaoZiming on GitHub (Oct 10, 2025).
Original GitHub issue: https://github.com/deepseek-ai/DeepEP/issues/449

Hi!
https://github.com/deepseek-ai/DeepEP/blob/main/deep_ep/buffer.py#L48-L49
I am wondering here, why does the num_qps_per_rank need to equal to the number of local experts? Thank you!

Originally created by @MaoZiming on GitHub (Oct 10, 2025). Original GitHub issue: https://github.com/deepseek-ai/DeepEP/issues/449 Hi! https://github.com/deepseek-ai/DeepEP/blob/main/deep_ep/buffer.py#L48-L49 I am wondering here, why does the `num_qps_per_rank` need to equal to the number of local experts? Thank you!
Author
Owner

@sphish commented on GitHub (Oct 11, 2025):

Since the number of local experts on each GPU are identical, this essentially assigns one QP to each remote expert. This is intended to minimize unnecessary data dependencies as much as possible.

<!-- gh-comment-id:3392735843 --> @sphish commented on GitHub (Oct 11, 2025): Since the number of local experts on each GPU are identical, this essentially assigns one QP to each remote expert. This is intended to minimize unnecessary data dependencies as much as possible.
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#294
No description provided.