deepspeedai / deepspeedai/DeepSpeed
DeepSpeed ZeRO with Ring-Allreduce
Nobody has claimed this yet.
- Dominant language
- Python
- Stars
- 43.1k
- Forks
- 5k
- Avg merge
- 4d 15h
- Merged PRs (30d)
- 112
Description
Hi, dear contributors,
I am trying to figure out whether we can use Ring-Allreduce to lower the communication volume while training a model.
Intuitively seems like that we could apply Ring-Allreduce with ZeRO stage 1 and stage 2, for the reason that both of them own the whole model's parameters on each worker.
So can we use Ring-Allreduce?
Looking forward for any replies, please :D
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
No files, tests, or entry points are mentioned. Start by reading the ZeRO stage 1 and stage 2 implementation and the distributed communication paths, then determine whether Ring-Allreduce is supported and what communication-volume changes would be required.
Written by the indexing model from the issue text.
Assessment
- Tech stack
- python, pytorch
- Domain
- distributed-systems, machine-learning
- Issue type
- Feature
- Difficulty
- 5/5
- Estimated time
- Over a week
- Activity status
- Stale
- Clarity
- Needs clarification
- Newbie friendliness
- 15/100