deepspeedai / deepspeedai/DeepSpeed
Communication hooks with deepspeed.initialize()
Nobody has claimed this yet.
- Dominant language
- Python
- Stars
- 43.1k
- Forks
- 5k
- Avg merge
- 4d 15h
- Merged PRs (30d)
- 112
Description
I would like to provide custom DDP communication hooks in the arguments to deepspeed.initialize() so that these hooks can be "registered" (in quotes because DeepSpeed doesn't directly use DDP) internally and used alongside ZeRO stages. I currently have the PowerSGD communication hook available with PyTorch 1.8 in mind.
Can this be supported with DeepSpeed?
@samyam @jeffra @tjruwase
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 or tests are identified in the issue. Start by locating deepspeed.initialize() and the ZeRO communication path, then determine how custom DDP communication hooks such as PowerSGD could be accepted and used alongside ZeRO stages; done would mean supported hook registration and use through initialization.
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
- 25/100