Skip to content

Does FSDP with Accelerate support async checkpoint saving (e.g. via torch.distributed.checkpoint)? #3746

@LucienXian

Description

@LucienXian

Hi, I'm using FSDP through Accelerate and would like to improve checkpoint saving performance. I noticed that PyTorch's torch.distributed.checkpoint (DCP) supports asynchronous saving with dcp.async_save, which can help overlap checkpoint I/O with training.

My question is: Does Accelerate's FSDP integration currently support async checkpointing via DCP or any similar mechanism?

If not directly exposed, is it safe to manually use torch.distributed.checkpoint.async_save with models and optimizers wrapped by Accelerate's FSDP? Are there any known compatibility issues or best practices (e.g., handling sharded state dict, optimizer state reconstruction)?

It would be great if Accelerate could expose async save as an option in save_state or provide utilities to integrate with DCP more seamlessly.

Thanks for your help!

Metadata

Metadata

Assignees

No one assigned

    Labels

    No labels
    No labels

    Type

    No type

    Projects

    No projects

    Milestone

    No milestone

    Relationships

    None yet

    Development

    No branches or pull requests

    Issue actions