diff --git a/src/accelerate/accelerator.py b/src/accelerate/accelerator.py
index 0bc7650f682..a1cbfaa8d68 100755
--- a/src/accelerate/accelerator.py
+++ b/src/accelerate/accelerator.py
@@ -1792,6 +1792,12 @@ def prepare_model(
         if device_placement is None:
             device_placement = self.device_placement and self.distributed_type != DistributedType.FSDP
 
+        # Ensure we can't double wrap a model
+        if getattr(model, "_is_accelerate_prepared", False):
+            if model not in self._models:
+                self._models.append(model)
+            return model
+
         self._models.append(model)
 
         # TODO: Look at enabling native TP training directly with a proper config
@@ -2054,6 +2060,7 @@ def prepare_model(
                 model = compile_regions(model, **self.state.dynamo_plugin.to_kwargs())
             else:
                 model = torch.compile(model, **self.state.dynamo_plugin.to_kwargs())
+        model._is_accelerate_prepared = True
         return model
 
     def _prepare_ao(self, *args):
diff --git a/tests/test_accelerator.py b/tests/test_accelerator.py
index c56b00ec3f5..ebab8a8c057 100644
--- a/tests/test_accelerator.py
+++ b/tests/test_accelerator.py
@@ -464,6 +464,17 @@ def test_is_accelerator_prepared(self):
             "Valid Dataloader is missing `_is_accelerator_prepared` or is set to `False`"
         )
 
+    def test_prepare_model_twice_does_not_double_wrap(self):
+        accelerator = Accelerator()
+        model = torch.nn.Linear(10, 2)
+        prepared_model = accelerator.prepare_model(model)
+        num_models_before = len(accelerator._models)
+        reprepared_model = accelerator.prepare_model(prepared_model)
+        assert len(accelerator._models) == num_models_before, (
+            "prepare_model should not add duplicate entries to _models"
+        )
+        assert reprepared_model is prepared_model, "prepare_model should return the same object when called twice"
+
     @require_cuda_or_xpu
     @slow
     @require_bnb