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

Fix condition when GA loss bug fix is not performed #35651

Open
wants to merge 4 commits into
base: main
Choose a base branch
from
+10 −10
Open

Fix condition when GA loss bug fix is not performed #35651

Show file tree
Hide file tree
Changes from all commits
Commits
Show all changes
4 commits
Select commit Hold shift + click to select a range
b95bf89
fix condition when GA loss bug fix is not performed
techkang Jan 13, 2025
91add0d
max loss diff is 2.29
techkang Jan 13, 2025
163cab6
fix typo
techkang Jan 13, 2025
f5040b2
add an extra validation that loss should not vary too much
techkang Jan 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
11 changes: 2 additions & 9 deletions src/transformers/trainer.py
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -3672,10 +3672,7 @@ def training_step(
return loss_mb.reduce_mean().detach().to(self.args.device)

with self.compute_loss_context_manager():
if self.model_accepts_loss_kwargs:
loss = self.compute_loss(model, inputs)
else:
loss = self.compute_loss(model, inputs, num_items_in_batch=num_items_in_batch)
loss = self.compute_loss(model, inputs, num_items_in_batch=num_items_in_batch)

del inputs
if (
Expand Down Expand Up @@ -3709,7 +3706,7 @@ def training_step(
scaled_loss.backward()
else:
# Finally we need to normalize the loss for reporting
if num_items_in_batch is None:
if not self.model_accepts_loss_kwargs and self.compute_loss_func is None:
Copy link
Contributor

Choose a reason for hiding this comment

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

Looks good to me

loss = loss / self.args.gradient_accumulation_steps

self.accelerator.backward(loss, **kwargs)
Expand Down Expand Up @@ -5157,10 +5154,6 @@ def get_batch_samples(self, epoch_iterator, num_batches):
except StopIteration:
break

# Keep default behavior the same
if not self.model_accepts_loss_kwargs:
return batch_samples, None

if len(batch_samples) > 0 and "labels" in batch_samples[0]:
# For now we don't support object detection
try:
Expand Down
9 changes: 8 additions & 1 deletion tests/trainer/test_trainer.py
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -855,7 +855,14 @@ def tokenize_function(examples):
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), 3, f"Difference {max(diff_broken)} is not greater than 3")
self.assertGreater(max(diff_broken), 2, f"Difference {max(diff_broken)} is not greater than 2")

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

# mean/sum loss should not vary too much.
relative_diff = abs(loss_base - loss_broken) / max(loss_base, loss_broken)
self.assertLess(relative_diff, 0.1, f"Relative difference {relative_diff} is not within 0.1")

@slow
def test_gradient_accumulation_loss_alignment_with_loss_func(self):
Expand Down