[GH-ISSUE #80] About the number of messages chunked in IBGDA #75

Closed
opened 2026-10-26 09:53:48 +00:00 by bsakaguchi · 2 comments
Owner

Originally created by @GitAlice123 on GitHub (Mar 19, 2025).
Original GitHub issue: https://github.com/deepseek-ai/DeepEP/issues/80

Image
Why theoretically 3 for maximum?

Originally created by @GitAlice123 on GitHub (Mar 19, 2025). Original GitHub issue: https://github.com/deepseek-ai/DeepEP/issues/80 ![Image](https://github.com/user-attachments/assets/16f9c42b-fdc2-42eb-ae9c-9fe9b8954909) Why theoretically 3 for maximum?
Author
Owner

@LyricZhao commented on GitHub (Mar 21, 2025):

Assuming the message size (maximum ~KB level) is much smaller than the page size (i.e. NVSHMEM_CUMEM_GRANULARITY, normally very large >100 MB). So the worst case of getting local/remote key is, the message is splitted into two pages at local, two pages at remote, totally 3 pages.

e.g.

| chunk 0 --- | chunk 1 ------ | chunk 2 --------- |
| local page i | local page i + 1 ------------------ |
| remote page j ------------- | remote page j + 1 |

<!-- gh-comment-id:2742663406 --> @LyricZhao commented on GitHub (Mar 21, 2025): Assuming the message size (maximum ~KB level) is much smaller than the page size (i.e. `NVSHMEM_CUMEM_GRANULARITY`, normally very large >100 MB). So the worst case of getting local/remote key is, the message is splitted into two pages at local, two pages at remote, totally 3 pages. e.g. | chunk 0 --- | chunk 1 ------ | chunk 2 --------- | | local page i | local page i + 1 ------------------ | | remote page j ------------- | remote page j + 1 |
Author
Owner

@LyricZhao commented on GitHub (Mar 21, 2025):

You can ignore that note as the while loop can proceed more than 3 chunks. But we tried some code simplication and optimizations here for the theretical maximum, but it didn't work.

<!-- gh-comment-id:2742666789 --> @LyricZhao commented on GitHub (Mar 21, 2025): You can ignore that note as the while loop can proceed more than 3 chunks. But we tried some code simplication and optimizations here for the theretical maximum, but it didn't work.
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#75
No description provided.