diff --git a/turbo_alignment/trainers/ref_model.py b/turbo_alignment/trainers/ref_model.py
index 4dd5d93..e441b75 100755
--- a/turbo_alignment/trainers/ref_model.py
+++ b/turbo_alignment/trainers/ref_model.py
@@ -149,7 +149,7 @@ def compute_loss(
 
         gc.collect()
         torch.cuda.empty_cache()
-        return torch.tensor([0.0], requires_grad=True)
+        return torch.tensor([0.0], requires_grad=True).to(self.accelerator.device)
 
     def prediction_step(
         self,
@@ -159,6 +159,10 @@ def prediction_step(
         ignore_keys: list[str] | None = None,
     ) -> tuple[torch.Tensor, torch.Tensor | None, torch.Tensor | None]:
         if prediction_loss_only:
-            return torch.tensor([0.0]).detach(), None, None
+            return torch.tensor([0.0]).to(self.accelerator.device).detach(), None, None
 
-        return torch.tensor([0.0]).detach(), torch.tensor([0.0]), torch.tensor([0.0])
+        return (
+            torch.tensor([0.0]).to(self.accelerator.device).detach(),
+            torch.tensor([0.0]).to(self.accelerator.device),
+            torch.tensor([0.0]).to(self.accelerator.device),
+        )