deepseek-ai / deepseek-ai/DeepEP
Will the Normal Kernel (dispatch kernel) exclusively occupy the SM of the GPU?
Open
- Dominant language
- Cuda
- Stars
- 10.1k
- Forks
- 1.4k
- Avg merge
- 4d 1h
- Merged PRs (30d)
- 2
Description
In the dispatch definition, each block has 512 threads. During actual execution, can we ensure that each block of the dispatch kernel exclusively occupies one SM through resource constraints such as register limits?
Contributor guide
No contributing guide indexed for this repository
Research direction
Start with the dispatch definition and its 512-thread block configuration, then trace the resource constraints relevant to SM occupancy. Done means documenting whether exclusive one-block-per-SM occupancy can be guaranteed and identifying the constraints or verification needed.
Written by the indexing model from the issue text.
Assessment
- Domain
- hpc, performance
- Issue type
- Feature
- Difficulty
- 5/5
- Estimated time
- Over a week
- Activity status
- Stale
- Clarity
- Needs clarification
- Newbie friendliness
- 20/100