Skip to content

Fix Zero3 contiguous grads, reduce scatter false accuracy issue #4321

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

Merged
tjruwase merged 2 commits into from
Sep 14, 2023
Merged

Fix Zero3 contiguous grads, reduce scatter false accuracy issue #4321

Changes from all commits
Commits
Show all changes
2 commits
Select commit Hold shift + click to select a range
1a6aeb3
Fix Zero3 contiguous grads, reduce scatter false accuracy issue
Sep 13, 2023
c02e4d3
Merge branch 'master' into master
tjruwase Sep 13, 2023
File filter

Filter by extension

Filter by extension
Conversations
Failed to load comments.
Loading
Jump to
Jump to file
Failed to load files.
Loading
Diff view
Diff view
2 changes: 1 addition & 1 deletion deepspeed/runtime/zero/stage3.py
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -1116,7 +1116,7 @@ def __add_grad_to_ipg_bucket(self, param: Parameter) -> None:
if not get_accelerator().is_synchronized_device():
self.reduce_and_partition_stream.wait_stream(get_accelerator().default_stream())

if self.contiguous_gradients and self.elements_in_ipg_bucket + param.grad.numel() < self.reduce_bucket_size:
if self.contiguous_gradients and self.elements_in_ipg_bucket + param.grad.numel() <= self.reduce_bucket_size:
# move the gradient to a contiguous buffer
with get_accelerator().stream(self.reduce_and_partition_stream):
# move the parameter's gradient to the contiguous flat buffer
Expand Down