diff --git a/train.py b/train.py
index 6b04e8ff3a6..d4a1b48b5aa 100644
--- a/train.py
+++ b/train.py
@@ -252,9 +252,7 @@ def train(hyp,  # path/to/hyp.yaml or hyp dictionary
 
     # DDP mode
     if cuda and RANK != -1:
-        model = DDP(model, device_ids=[LOCAL_RANK], output_device=LOCAL_RANK,
-                    # nn.MultiheadAttention incompatibility with DDP https://github.com/pytorch/pytorch/issues/26698
-                    find_unused_parameters=any(isinstance(layer, nn.MultiheadAttention) for layer in model.modules()))
+        model = DDP(model, device_ids=[LOCAL_RANK], output_device=LOCAL_RANK)
 
     # Model parameters
     hyp['box'] *= 3. / nl  # scale to layers