diff --git a/train.py b/train.py
index a745fa3..157e449 100644
--- a/train.py
+++ b/train.py
@@ -122,6 +122,11 @@ def main():
             loss_ce, loss_survival = loss_fn(logits, target_events, target_wait_times)
             loss = loss_ce + loss_survival
 
+            # When using DataParallel, loss is a vector of losses from each GPU.
+            # We need to average them to get a single scalar loss.
+            if isinstance(model, nn.DataParallel):
+                loss = loss.mean()
+
             # Backward pass and optimization
             optimizer.zero_grad()
             loss.backward()