deepspeedai / deepspeedai/DeepSpeedExamples
zero3 and enable hybrid engine are not suitable for llama2, how to solve it?
Nobody has claimed this yet.
- Dominant language
- Python
- Stars
- 6.8k
- Forks
- 1.1k
- Avg merge
- 2d 16h
- Merged PRs (30d)
- 1
Description
In my experiment, I found that if I use zero3 and enable hybrid engine setting, the Actor will generate repeat token or nothing during stage 3 (PPO) training. Here is an example:
Besides, I took some other experiments:
Experiment 1: use the zero2 and enable hybrid engine, which is fine.
Experiment 2: use the zero3 and disenable hybrid engine, which is fine.
Experiment 3: I tested the OPT model, which is fine in zero3, and enable hybrid engine setting.
I'm confused by this phenomenon and want to know how to fix it. That said, due to time and GPU memory constraints, I want to use zero3 and enable hybrid engine settings.
Contributor guide
No contributing guide indexed for this repository
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 by reproducing the reported stage 3 PPO behavior with Llama2, ZeRO-3, and hybrid engine enabled, then compare it with the working ZeRO-2, hybrid-disabled, and OPT configurations. Trace the Actor generation path and determine the cause of repeated or empty tokens; done means identifying a reproducible fix and confirming normal generation in the failing configuration.
Written by the indexing model from the issue text.
Assessment
- Tech stack
- python
- Domain
- machine-learning
- Issue type
- Bug
- Difficulty
- 4/5
- Estimated time
- 3-5 days
- Activity status
- Stale
- Clarity
- Needs clarification
- Newbie friendliness
- 30/100