deepspeedai / deepspeedai/DeepSpeed
[BUG]not a bug, ask for solution：ppo zero3+offload generate too slow
Nobody has claimed this yet.
- Dominant language
- Python
- Stars
- 43.1k
- Forks
- 5k
- Avg merge
- 4d 15h
- Merged PRs (30d)
- 112
Description
Hi, when I ran ppo with llama-13b (prompt_len=256, answer_len = 512) with zero stage 3 and offload (8*V100-32G), it seems the generation time is too slow (average about 20s-300s). Currently, I consider the reason of generating slowly is using zero stage 3 with open synced_gpus, but it consumes a lot of resource if I do not use this. Is there any solution for solving this problem? Thanks a lot!
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
Start with the PPO generation path and the ZeRO stage 3, offload, and synced_gpus configuration described in the report. Reproduce the llama-13b workload with the stated prompt and answer lengths, compare the relevant configurations, and document a verified way to avoid the excessive generation time without exhausting resources.
Written by the indexing model from the issue text.
Assessment
- Tech stack
- python
- Domain
- distributed-systems, machine-learning, performance
- Issue type
- Bug
- Difficulty
- 4/5
- Estimated time
- 3-5 days
- Activity status
- Stale
- Clarity
- Needs clarification
- Newbie friendliness
- 25/100