Skip to content

fix(loss): use autograd-aware all_reduce in BarlowTwinsLoss #1980

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
saud5150 wants to merge 3 commits into
base: master
Choose a base branch
from
+5 −2
Open

fix(loss): use autograd-aware all_reduce in BarlowTwinsLoss #1980

Changes from all commits
Commits
Show all changes
3 commits
Select commit Hold shift + click to select a range
617f4ea
fix(loss): use autograd-aware all_reduce in BarlowTwinsLoss
Jul 13, 2026
4bf0050
test(loss): remove mock-based gather_distributed tests
Jul 14, 2026
f6d85b3
Merge branch 'master' into fix/barlow-twins-gather-distributed-gradient
saud5150 Jul 14, 2026
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
7 changes: 5 additions & 2 deletions lightly/loss/barlow_twins_loss.py
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -4,6 +4,7 @@
import torch.distributed as dist
import torch.nn.functional as F
from torch import Tensor
from torch.distributed import nn as dist_nn


class BarlowTwinsLoss(torch.nn.Module):
Expand Down Expand Up @@ -76,8 +77,10 @@ def forward(self, z_a: torch.Tensor, z_b: torch.Tensor) -> torch.Tensor:
if self.gather_distributed and dist.is_initialized():
world_size = dist.get_world_size()
if world_size > 1:
c = c / world_size
dist.all_reduce(c)
# Use the autograd-aware torch.distributed.nn.all_reduce so the
# backward pass scales gradients by world_size, matching
# single-GPU behaviour. ref #1977
c = dist_nn.all_reduce(c / world_size)
Comment thread
saud5150 marked this conversation as resolved.

invariance_loss = torch.diagonal(c).add_(-1).pow_(2).sum()
redundancy_reduction_loss = _off_diagonal(c).pow_(2).sum()
Expand Down
Loading