diff --git a/src/open_clip/loss.py b/src/open_clip/loss.py
index e4f9c650b..06134015b 100644
--- a/src/open_clip/loss.py
+++ b/src/open_clip/loss.py
@@ -324,7 +324,7 @@ def forward(self, image_features, text_features, logit_scale, logit_bias, output
                 loss += neg_loss
                 text_features_to_right = text_features_from_left
 
-            loss /= self.world_size  # not 100% clear if this should be here
+            # loss /= self.world_size  # better without this
 
             # approach #2
             # loss = torch.distributed.nn.all_reduce(loss)