Running out of memory with version v0.5.17
Nobody has claimed this yet.
- Dominant language
- Python
- Stars
- 10.8k
- Forks
- 989
- Avg merge
- 6h 29m
- Merged PRs (30d)
- 85
Description
I used v0.5.6 but upgraded to version v0.5.17 since it uses the newer version of unsloth, which is supposed to be much more memory efficient, especially for MOE models and RL. But with the newer version, I get an out-of-memory error for the same model and the same setup with which v0.5.6 worked fine. I'm training Qwen2.5-14B on A100 GPU. Now, even if I try to run Qwen2.5-7B I get out of memory while loading the model, not even when training starts.
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
Reproduce the report on an A100 with v0.5.17, first loading Qwen2.5-7B and then comparing the same setup against v0.5.6. Investigate the memory difference around model loading and the newer unsloth dependency. Done means identifying the regression or missing configuration needed to restore the earlier memory usage.
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
- 25/100