mirror of
https://github.com/deepseek-ai/DeepEP.git
synced 2026-11-05 00:47:32 +00:00
[GH-ISSUE #338] Has the currently implemented logfmt quantization been able to optimize the combine operation to reduce time consumption? #237
Labels
No labels
pull-request
No milestone
No project
No assignees
1 participant
Notifications
Due date
No due date set.
Dependencies
No dependencies set.
Reference
deepseek-ai/DeepEP#237
Loading…
Add table
Add a link
Reference in a new issue
No description provided.
Delete branch "%!s()"
Deleting a branch is permanent. Although the deleted branch may continue to exist for a short time before it actually gets removed, it CANNOT be undone in most cases. Continue?
Originally created by @DongYuqi1 on GitHub (Jul 30, 2025).
Original GitHub issue: https://github.com/deepseek-ai/DeepEP/issues/338
Hello, I have implemented a logfmt-8bit kernel based on the logfmt quantization in internode-ll for testing the 8-card environment of intranode H800. The combine bandwidth is 160GB/s, the token*hidden number propagated is 4096 and 7168, and the topk is 8. At this time, the time cost without quantization is 1.9ms. For the quantization operation, if kNumUnrolls is set to 1, the time cost is 0.11ms. So, can we say that a good optimization effect has been achieved?