[Lazy init] Force fall back to slow init for composite models

src/transformers/modeling_utils.py

@@ -510,6 +510,12 @@ def get_output_embeddings(self) -> nn.Module:
         """
         return None  # Overwrite for models with output embeddings
 
+    def _init_weights(self, module):
+        """
+        Initialize the weights. This method should be overridden by derived class.
+        """
+        raise NotImplementedError(f"Make sure `_init_weigths` is implemented for {self.__class__}")
+
     def tie_weights(self):
         """
         Tie the weights between the input embeddings and the output embeddings.
@@ -1205,7 +1211,7 @@ def from_pretrained(cls, pretrained_model_name_or_path: Optional[Union[str, os.P
                     )
 
             model, missing_keys, unexpected_keys, error_msgs = cls._load_state_dict_into_model(
-                model, state_dict, pretrained_model_name_or_path
+                model, state_dict, pretrained_model_name_or_path, _fast_init=_fast_init
             )
 
         # make sure token embedding weights are still tied if needed
@@ -1225,7 +1231,7 @@ def from_pretrained(cls, pretrained_model_name_or_path: Optional[Union[str, os.P
         return model
 
     @classmethod
-    def _load_state_dict_into_model(cls, model, state_dict, pretrained_model_name_or_path):
+    def _load_state_dict_into_model(cls, model, state_dict, pretrained_model_name_or_path, _fast_init=True):
 
         # Convert old format to new format if needed from a PyTorch state_dict
         old_keys = []
@@ -1273,12 +1279,14 @@ def _load_state_dict_into_model(cls, model, state_dict, pretrained_model_name_or
             for pat in cls._keys_to_ignore_on_load_unexpected:
                 unexpected_keys = [k for k in unexpected_keys if re.search(pat, k) is None]
 
-        # tie unintialized modules
-        unintialized_modules = model.retrieve_modules_from_names(
-            missing_keys, add_prefix=add_prefix, remove_prefix=remove_prefix
-        )
-        for module in unintialized_modules:
-            model._init_weights(module)
+        if _fast_init:
+            # retrieve unintialized modules and initialize
+            unintialized_modules = model.retrieve_modules_from_names(
+                missing_keys, add_prefix=add_prefix, remove_prefix=remove_prefix
+            )
+            for module in unintialized_modules:
+                model._init_weights(module)
+
         # copy state_dict so _load_from_state_dict can modify it
         metadata = getattr(state_dict, "_metadata", None)
         state_dict = state_dict.copy()

src/transformers/models/encoder_decoder/modeling_encoder_decoder.py

@@ -221,6 +221,13 @@ def get_output_embeddings(self):
     def set_output_embeddings(self, new_embeddings):
         return self.decoder.set_output_embeddings(new_embeddings)
 
+    @classmethod
+    def from_pretrained(cls, *args, **kwargs):
+        # At the moment fast initialization is not supported
+        # for composite models
+        kwargs["_fast_init"] = False
+        return super(EncoderDecoderModel, cls).from_pretrained(*args, **kwargs)
+
     @classmethod
     def from_encoder_decoder_pretrained(
         cls,

src/transformers/models/rag/modeling_rag.py

@@ -232,6 +232,13 @@ class RagPreTrainedModel(PreTrainedModel):
     base_model_prefix = "rag"
     _keys_to_ignore_on_load_missing = [r"position_ids"]
 
+    @classmethod
+    def from_pretrained(cls, *args, **kwargs):
+        # At the moment fast initialization is not supported
+        # for composite models
+        kwargs["_fast_init"] = False
+        return super(RagPreTrainedModel, cls).from_pretrained(*args, **kwargs)
+
     @classmethod
     def from_pretrained_question_encoder_generator(
         cls,