lm-sys / lm-sys/FastChat

NotImplementedError: offload_to_cpu=True and NO_SHARD is not supported yet

Open
#1,963 0 comments 0 reactions 0 assignees View on GitHub

Nobody has claimed this yet.

Dominant language
Python
Stars
39.5k
Forks
4.8k
PR merge metrics
No merged PRs in 30d

Description

I am using a single A100 to train open llama 3b model. The training runs fine but was interrupted with an error below
`NotImplementedError: offload_to_cpu=True and NO_SHARD is not supported yet` when saving the checkpoint. Interestingly, the final model can be saved in the case when the total steps were smaller than the save steps (seems there is problem when saving intermedia checkpoints).
Google it without any results so raise an issue here.

Contributor guide

No contributing guide indexed for this repository

First steps

  1. Read the whole issue, then the project's contributing guide.
  2. Comment on the issue to say you are picking it up — it saves two people doing the same work.
  3. Fork the repository and make your change on a branch.
  4. Open a pull request that references the issue number.

Research direction

Start by reproducing the checkpoint save failure during intermediate saves with a single A100, using offload_to_cpu=True and NO_SHARD. Compare this path with the final-model save path, which reportedly succeeds when total steps are below save steps. Done means intermediate checkpoints save successfully in the reported training setup without the NotImplementedError.

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

Get new issues in your inbox

A short digest of beginner-friendly GitHub issues.