deepspeedai / deepspeedai/DeepSpeed
[PROBLEM] P2p recv waiting for data will cause other threads under the same process to be unable to perform any operations
Nobody has claimed this yet.
- Dominant language
- Python
- Stars
- 43.1k
- Forks
- 5k
- Avg merge
- 4d 15h
- Merged PRs (30d)
- 112
Description
Context: I want to speed up the InferenceSchedule(deepspeed/runtime/pipe/schedule.py) step function. So I launch two threads for SendActivation and RecvActivation for maximum GPU utilization. But I found that when RecvActivation waiting data(hangs in RecvActivation thread), any operations on this GPU card cannot be performed. Any operations hangs. for example, I want to print this variable, but stuck...however, GPU UTL 100% because waiting for recving future data....
Problem: How to have a thread waiting for data to be received in this card and perform corresponding operations at the same time?（for example, forward/backward）
Contributor guide
First steps
- Read the whole issue, then the project's contributing guide.
- Comment on the issue to say you are picking it up — it saves two people doing the same work.
- Fork the repository and make your change on a branch.
- Open a pull request that references the issue number.
Research direction
Start by reading deepspeed/runtime/pipe/schedule.py, especially InferenceSchedule.step and the SendActivation and RecvActivation paths. Reproduce the reported two-thread behavior and determine the supported concurrency limitation or required change. Done means a receive can wait for data while the same GPU process continues the intended forward or backward operation without hanging.
Written by the indexing model from the issue text.
Assessment
- Tech stack
- python
- Domain
- distributed-systems, machine-learning
- Issue type
- Bug
- Difficulty
- 5/5
- Estimated time
- Over a week
- Activity status
- Stale
- Clarity
- Needs clarification
- Newbie friendliness
- 25/100