deepspeedai / deepspeedai/DeepSpeed
[REQUEST] Asynchronous Checkpointing
@tjruwase is already working on this.
Since Aug 3, 2024.
- Dominant language
- Python
- Stars
- 43.1k
- Forks
- 5k
- Avg merge
- 4d 15h
- Merged PRs (30d)
- 112
Description
Is your feature request related to a problem? Please describe.
Checkpointing is significantly faster with Torch Distributed's async checkpoint feature: https://pytorch.org/docs/stable/distributed.checkpoint.html#torch.distributed.checkpoint.state_dict_saver.async_save
Blog post: https://pytorch.org/blog/reducing-checkpointing-times/
We want to checkpoint frequently, but it is expensive because it blocks training.
Describe the solution you'd like
Checkpoints should load the params to CPU, then save the checkpoint while training continues.
Describe alternatives you've considered
Nebula is only for Azure users and is also kinda broken. Torch's FSDP appears to perform worse in general (performance and accuracy) compared to DeepSpeed (likely due to differences in your mixed precision implementations I don't quite understand yet).
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.
Assessment
This issue has not been assessed yet.