deepseek-ai / deepseek-ai/DeepEP
Profiling Results on H20 with DSv3
Open
- Dominant language
- Cuda
- Stars
- 10.1k
- Forks
- 1.4k
- Avg merge
- 4d 1h
- Merged PRs (30d)
- 2
Description
Issue:
During profiling on H20 (DSv3), we observed that the time spent in both the dispatch and combine operations (specifically the notify interface called before these steps) is significantly longer than the actual communication time. These two kernels appear to be causing the total communication time in DeepEP to nearly double.
Question:
Is this an expected behavior, or could it be due to incorrect usage on our side?
Contributor guide
No contributing guide indexed for this repository
Assessment
This issue has not been assessed yet.