deepspeedai / deepspeedai/DeepSpeed
How to offload/move model weights manually ?
Open
Nobody has claimed this yet.
enhancement
- Dominant language
- Python
- Stars
- 43.1k
- Forks
- 5k
- Avg merge
- 4d 15h
- Merged PRs (30d)
- 112
Description
The ZeRO3 model is loaded onto the GPU by default, but can't be manually offloaded. Currently, I want to offload the model weights and Adam optimizer to the CPU during an inference stage in RLHF to improve the efficiency, and then load them back onto the GPU after that stage ends. How can I implement this?
Thanks
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 file, test, or entry point is named. Start by locating DeepSpeed's ZeRO3 implementation and the paths handling model weights and Adam optimizer during RLHF inference. Done should mean a clear, supported way to move both to CPU and restore them to the GPU.
Written by the indexing model from the issue text.
Assessment
- Tech stack
- python
- Domain
- distributed-systems, machine-learning, performance
- Issue type
- Feature
- Difficulty
- 5/5
- Estimated time
- Over a week
- Activity status
- Stale
- Clarity
- Needs clarification
- Newbie friendliness
- 20/100