Skip to content

[checkpoint] Support fsspec checkpoint paths#3730

Open
YassineYousfi wants to merge 7 commits into
pytorch:mainfrom
commaai:checkpoint-fs
Open

[checkpoint] Support fsspec checkpoint paths#3730
YassineYousfi wants to merge 7 commits into
pytorch:mainfrom
commaai:checkpoint-fs

Conversation

@YassineYousfi

@YassineYousfi YassineYousfi commented Jun 21, 2026

Copy link
Copy Markdown
Contributor

make checkpointing fsspec aware, leave it to DCP to pick the appropriate writer/reader
when fsspec is used, base_folder is not used
also added unit test

@meta-cla meta-cla Bot added the CLA Signed This label is managed by the Meta Open Source bot. label Jun 21, 2026
@pytorch-bot

pytorch-bot Bot commented Jun 21, 2026

Copy link
Copy Markdown

Workflows were awaiting approval. CI has now been triggered for the ciflow labels on this PR.

Comment thread .ci/docker/requirements.txt Outdated
Comment thread pyproject.toml Outdated

@fegin fegin left a comment

Copy link
Copy Markdown
Contributor

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

I don't think it is a good idea to make fsspec as the default (we didn't do this for DCP either). Several reasons, we didn't measure the performance (probably not a big issue), unclear support of mmap, which can be a potential use case in the future.

I understand S3 support is important. A better design is to make fsspec as an option not default.

@YassineYousfi

Copy link
Copy Markdown
Contributor Author

I don't think it is a good idea to make fsspec as the default (we didn't do this for DCP either). Several reasons, we didn't measure the performance (probably not a big issue), unclear support of mmap, which can be a potential use case in the future.

I understand S3 support is important. A better design is to make fsspec as an option not default.

Done. DCP already tries fsspec in _storage_setup, so the only changes are the file management/discovery

https://github.com/pytorch/pytorch/blob/main/torch/distributed/checkpoint/_storage_utils.py#L33

Comment thread torchtitan/components/fs.py Outdated
Comment thread torchtitan/components/checkpoint.py Outdated
Comment thread torchtitan/components/checkpoint.py Outdated
@fegin

fegin commented Jul 8, 2026

Copy link
Copy Markdown
Contributor

The title needs to be changed too.

@tianyu-l

tianyu-l commented Jul 8, 2026

Copy link
Copy Markdown
Contributor

related PR #3887

@YassineYousfi YassineYousfi changed the title [checkpoint] FsspecWriter/FsspecReader [checkpoint] Support fsspec checkpoint paths Jul 10, 2026
Sign up for free to join this conversation on GitHub. Already have an account? Sign in to comment

Labels

ciflow/8gpu CLA Signed This label is managed by the Meta Open Source bot.

Projects

None yet

Development

Successfully merging this pull request may close these issues.

3 participants