Skip to content

Fix grad accum arbitrary value #36691

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
IlyasMoutawwakil merged 2 commits into from
Mar 14, 2025
Merged

Fix grad accum arbitrary value #36691

Show file tree
Hide file tree
Changes from all commits
Commits
Show all changes
2 commits
Select commit Hold shift + click to select a range
b0113af
fix grad accum arbitrary value
IlyasMoutawwakil Mar 13, 2025
a187e38
Merge branch 'main' into fix-grad-accum-test
SunMarc Mar 13, 2025
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 src/transformers/utils/import_utils.py
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -813,7 +813,7 @@ def is_torch_hpu_available():

def patched_masked_fill_(self, mask, value):
if self.dtype == torch.int64:
logger.warning(
logger.warning_once(
"In-place tensor.masked_fill_(mask, value) is not supported for int64 tensors on Gaudi1. "
"This operation will be performed out-of-place using tensor[mask] = value."
)
Expand Down
5 changes: 2 additions & 3 deletions tests/trainer/test_trainer.py
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -894,9 +894,8 @@ def tokenize_function(examples):

# all diff truth should be quite close
self.assertLess(max(diff_truth), 0.01, f"Difference {max(diff_truth)} is not within 0.01")

# max diff broken should be very off
self.assertGreater(max(diff_broken), 1.3, f"Difference {max(diff_broken)} is not greater than 1.3")
# max diff broken should be very off ("very off" is arbitrary, but as long as it's bigger than 0.1, it's fine)
self.assertGreater(max(diff_broken), 0.7, f"Difference {max(diff_broken)} is not greater than 0.7")

loss_base = sum(base_loss_callback.losses)
loss_broken = sum(broken_loss_callback.losses)
Expand Down