Skip to content

fix(loss): compensate InfoNCE gradients for DDP averaging #10081

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

Open
taking-lying-flat wants to merge 1 commit into
base: main
Choose a base branch
from
Open

fix(loss): compensate InfoNCE gradients for DDP averaging #10081

Changes from all commits
Commits
Show all changes
1 commit
Select commit
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
6 changes: 6 additions & 0 deletions swift/loss/embedding.py
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -133,6 +133,7 @@ def __call__(self, outputs, labels, **kwargs) -> torch.Tensor:
rank, _, world_size, _ = get_dist_setting()
# repeat of anchor(1)+positive(1)+negatives(n)
sentences = outputs['last_hidden_state']
gradient_scale = 1

if world_size > 1 and use_batch:
if getattr(sequence_parallel, 'dp_group', None) is not None:
Expand Down Expand Up @@ -163,6 +164,8 @@ def __call__(self, outputs, labels, **kwargs) -> torch.Tensor:
labels = gather_object(labels)
# override the gathered one
all_sentences[rank] = sentences
if not self.is_megatron:
gradient_scale = len(all_sentences)
for idx in range(len(all_sentences)):
if idx == rank:
continue
Expand Down Expand Up @@ -321,4 +324,7 @@ def __call__(self, outputs, labels, **kwargs) -> torch.Tensor:
# next positive is neg+1
length += tensor.size(0) - 1
loss /= len(split_tensors)
if gradient_scale > 1:
# DDP averages the local contributions to the global loss graph; preserve the reported loss.
loss = loss.detach() + gradient_scale * (loss - loss.detach())
return loss
Loading