deepspeedai / deepspeedai/DeepSpeed
[QUESTION] pipe schedule loads a micro-batch at the last stage
Nobody has claimed this yet.
- Dominant language
- Python
- Stars
- 43.1k
- Forks
- 5k
- Avg merge
- 4d 15h
- Merged PRs (30d)
- 112
Description
Dear team, I am using DeepSpeed's pipeline engine to train models. I would be very grateful if you could answer this question for me.
I wonder why DeepSpeed loads a micro-batch in the last stage at here? I understand that we should load a micro-batch before executing the forward part of the first stage, while I cannot understand the reason for loading a micro-batch in the last stage.
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
Read deepspeed/runtime/pipe/schedule.py around line 224 and trace the surrounding pipeline schedule flow. Use the reported question as the acceptance criterion: the completed work should clearly document why the last stage loads a micro-batch, either in the issue response or appropriate documentation.
Written by the indexing model from the issue text.
Assessment
- Tech stack
- python
- Domain
- distributed-systems, machine-learning
- Issue type
- Documentation
- Difficulty
- 4/5
- Estimated time
- 3-5 days
- Activity status
- Stale
- Clarity
- Needs clarification
- Newbie friendliness
- 25/100