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 #71] Confues about the Low-latency kernels data issue #64

New issue
Closed
opened 2026-10-26 09:53:37 +00:00 by bsakaguchi · 1 comment
bsakaguchi commented 2026-10-26 09:53:37 +00:00
Owner
Copy link

Originally created by @ysywh on GitHub (Mar 12, 2025).
Original GitHub issue: https://github.com/deepseek-ai/DeepEP/issues/71

Hello DeepEP developers, appreciate the solid works!
I have two questions about the Low-latency kernels with pure RDMA displayed on homepage, thanks :
1, Is the low_latency data displayed the average, maximum or minimum value of all ranks? Which method makes more sense?
2, Why does the bandwidth decrease as the number of #EP increases?

Originally created by @ysywh on GitHub (Mar 12, 2025). Original GitHub issue: https://github.com/deepseek-ai/DeepEP/issues/71 Hello DeepEP developers, appreciate the solid works! I have two questions about the Low-latency kernels with pure RDMA displayed on homepage, thanks : 1, Is the low_latency data displayed the average, maximum or minimum value of all ranks? Which method makes more sense? 2, Why does the bandwidth decrease as the number of #EP increases?
bsakaguchi commented 2026-10-26 09:53:38 +00:00
Author
Owner
Copy link

@sphish commented on GitHub (Mar 13, 2025):

  1. The performance values indicated in the README are averages, and in a stable network environment, these three values generally only vary by a few microseconds.

  2. Here are some potential impacts when the number of #EP increases:
    a. The gap for synchronization becomes larger, and all-to-all communication effectively acts as an implicit synchronization operation.
    b. The NICs needs to handle more sending destinations, which can impact performance.
    c. It becomes more susceptible to network fluctuations and congestion.
    d. There are fewer experts per rank, making load imbalance more apparent.

<!-- gh-comment-id:2719510152 --> @sphish commented on GitHub (Mar 13, 2025): 1. The performance values indicated in the README are averages, and in a stable network environment, these three values generally only vary by a few microseconds. 2. Here are some potential impacts when the number of #EP increases: a. The gap for synchronization becomes larger, and all-to-all communication effectively acts as an implicit synchronization operation. b. The NICs needs to handle more sending destinations, which can impact performance. c. It becomes more susceptible to network fluctuations and congestion. d. There are fewer experts per rank, making load imbalance more apparent.
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#64
No description provided.