deepspeedai / deepspeedai/DeepSpeed
[REQUEST] how many partitios will zero stage 3 divide mode
Nobody has claimed this yet.
- Dominant language
- Python
- Stars
- 43.1k
- Forks
- 5k
- Avg merge
- 4d 15h
- Merged PRs (30d)
- 112
Description
I have 640 GPUs on 80 machines to do some pretrain work now, I have read the zero tutorials but didn't see any config about number of partitions.
The discussion of this question has been solved in https://github.com/microsoft/DeepSpeed/discussions/3438, I hope it can be added to zero tutorials
Btw, I think single zero3 is not suitable for large scale cluster, in my tuning test, the work will take 3 days on 80 machines and only 4 days on 40 machines, I suppose additional network overhead was key of performance.
I suggest the MiCS solution can be a default setting to reduce communaction and make deepspeed easier to use.
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
Review the Zero tutorials and discussion #3438, which contains the resolved partition-count explanation. Update the tutorial to cover how ZeRO-3 determines partitions and consider documenting the MiCS suggestion; the documentation should answer the partition question using the established discussion.
Written by the indexing model from the issue text.
Assessment
- Tech stack
- python
- Domain
- distributed-systems, documentation, machine-learning
- Issue type
- Documentation
- Difficulty
- 2/5
- Estimated time
- Half a day
- Activity status
- Stale
- Clarity
- Mostly clear
- Newbie friendliness
- 42/100