Skip to content
New issue

Have a question about this project? Sign up for a free GitHub account to open an issue and contact its maintainers and the community.

Sign up for GitHub

By clicking “Sign up for GitHub”, you agree to our terms of service and privacy statement. We’ll occasionally send you account related emails.

Already on GitHub? Sign in to your account

Jump to bottom

patch for smddp import #18244

Merged
merged 2 commits into from Jul 26, 2022
Merged

patch for smddp import #18244

Changes from all commits
Commits
Show all changes
2 commits
Select commit Hold shift + click to select a range
2bed140
add import
carolynwang Jul 21, 2022
0e86151
format
carolynwang Jul 26, 2022
File filter

Filter by extension

Filter by extension
Conversations
Failed to load comments.
Jump to
Jump to file
Failed to load files.
Diff view
Diff view
2 changes: 2 additions & 0 deletions src/transformers/training_args.py
View file
Open in desktop
Expand Up @@ -1333,6 +1333,8 @@ def _setup_devices(self) -> "torch.device":
device = torch.device("cuda", local_rank)
self._n_gpu = 1
elif is_sagemaker_dp_enabled():
import smdistributed.dataparallel.torch.torch_smddp # noqa: F401

Comment on lines +1336 to +1337
Copy link
Member

Choose a reason for hiding this comment

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

i thought this was not needed anymore?

Copy link
Contributor Author

Choose a reason for hiding this comment

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

this torch_smddp import is required to register smddp as a PT backend - the older version of smddp required a [different] import as well as other points of code changes which is no longer needed

Copy link
Member

Choose a reason for hiding this comment

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

I see i guess this slipped through here: https://github.com/huggingface/transformers/pull/16371/files#diff-bfceaff300c851b8e24fc50dc6638482abaec8f7d2a718e877c3828c166bcf79

dist.init_process_group(backend="smddp")
self.local_rank = int(os.getenv("SMDATAPARALLEL_LOCAL_RANK"))
device = torch.device("cuda", self.local_rank)
Expand Down