Skip to content

Pytorch daos dataloader#327

Open
zhenghh04 wants to merge 6 commits intoargonne-lcf:mainfrom
enaktalabs:pytorch-daos-dataloader
Open

Pytorch daos dataloader#327
zhenghh04 wants to merge 6 commits intoargonne-lcf:mainfrom
enaktalabs:pytorch-daos-dataloader

Conversation

@zhenghh04
Copy link
Copy Markdown
Member

No description provided.

@zhenghh04
Copy link
Copy Markdown
Member Author

@0xE0F Could you fix the conflicts?

0xE0F added 4 commits March 6, 2026 08:55
Signed-off-by: Denis Barakhtanov <dbarahtanov@enakta.com>
DAOS provides ready to use DaosDataset that can be directly pluggted in
into torch.DataLoader.
DLIO TorchDataLoader uses it with support of the custom Dataset that
delegates sample reading to the FormatReader.
For DAOS we can ommit FormatReader and create a wrapper for DaosDataset
that interact with performance tooling and pass it directly to the
torch.DataLoader.
@0xE0F 0xE0F force-pushed the pytorch-daos-dataloader branch from 25359dc to e83d68b Compare March 5, 2026 21:56
@enakta
Copy link
Copy Markdown
Contributor

enakta commented Mar 5, 2026

@zhenghh04 rebased to the latest main brach, should be good now.

Copy link
Copy Markdown
Collaborator

@hariharan-devarajan hariharan-devarajan left a comment

Choose a reason for hiding this comment

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

We are missing test cases. I understand we don't have actual doas but we should explore mocking strategies.

@enakta
Copy link
Copy Markdown
Contributor

enakta commented Mar 24, 2026

We are missing test cases. I understand we don't have actual doas but we should explore mocking strategies.

I think mocking in this case would be mostly testing the mock itself as DaosDataset pretty much hands off samples loading to pytroch.DataLoader.
There should be way to run single node ephemeral DAOS server in the container, let me have a look whether it would work CI.

Sign up for free to join this conversation on GitHub. Already have an account? Sign in to comment

Labels

None yet

Projects

None yet

Development

Successfully merging this pull request may close these issues.

4 participants