Fix _load_state_dict_into_meta_model with device_map=None (#36488)

* Fix _load_state_dict_into_meta_model with device_map=None

* Update src/transformers/modeling_utils.py

src/transformers/modeling_utils.py

@@ -785,7 +785,7 @@ def _load_state_dict_into_meta_model(
     tensor_device = None
     if device_map is not None and device_map.get("", None) is not None:
         tensor_device = device_map[""].index if isinstance(device_map[""], torch.device) else device_map[""]
-
+    if device_map is not None:
         device_map_regex = "|".join(sorted(device_map.keys(), reverse=True))
 
     # we need this later to initialize tensor parallelism