mirror of
https://github.com/deepseek-ai/DeepEP.git
synced 2026-11-05 00:47:32 +00:00
[GH-ISSUE #109] question about the INT4 of combine_msg in low latency mode #99
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#99
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 @leehongming on GitHub (Apr 3, 2025).
Original GitHub issue: https://github.com/deepseek-ai/DeepEP/issues/109
Hi chenggang,
We are trying to use the zero_copy mode in DeepEP with DeepGemm, but we encounter a problem.
The gemm output buffer of DeepGemm says that it should be continuous, however the return of buffer.get_next_low_latency_combine_buffer() is not continuous.
Then after reading the source code of low_latency_combine, we find that the incontinuous of low_latency_combine_buffer comes from the structure of combine_msg, which contains a INT4 and a token (hidden data).
What puzzles me is the function of the INT4. It seems to be the control part of each token, but I cannot find its assignment and usage.
And when I try to remove this INT4 in both send&recv phase of low_latency_combine, the testbench test_low_latency works well....
@LyricZhao commented on GitHub (Apr 7, 2025):
Yes, it has no usages. We use it as a reserved part (internal impl has something else to do). I will remove it later in the public repo.
@LyricZhao commented on GitHub (Apr 7, 2025):
Fixed in
github.com/deepseek-ai/DeepEP@42494864ba, feel free to reopen this issue. Thanks for your feedback!