fix tied_params for meta tensor (#25101)

* fix tied_params for meta tensor

* remove duplicate

src/transformers/modeling_utils.py

@@ -3060,13 +3060,17 @@ class PreTrainedModel(nn.Module, ModuleUtilsMixin, GenerationMixin, PushToHubMix
         unexpected_keys = list(unexpected_keys - model_buffers)
 
         model.tie_weights()
+        if device_map is None:
             ptrs = collections.defaultdict(list)
             for name, tensor in model.state_dict().items():
-            id_tensor = id_tensor_storage(tensor) if tensor.device != torch.device("meta") else id(tensor)
+                id_tensor = id_tensor_storage(tensor)
                 ptrs[id_tensor].append(name)
 
             # These are all the pointers of shared tensors.
             tied_params = [names for _, names in ptrs.items() if len(names) > 1]
+        else:
+            # id function doesn't work for meta tensor so we need this function
+            tied_params = find_tied_parameters(model)
 
         for group in tied_params:
             if remove_prefix_from_model: