[Longformer, Bert, Roberta, ...] Fix multi gpu training (#7272)

* fix multi-gpu

* fix longformer

* force to delete unnecessary layers

* fix notifications

* fix warning

* fix roberta

* fix tests

* remove hasattr

* fix tests

* fix roberta

* merge and clean authorized keys

src/transformers/modeling_utils.py

@@ -398,6 +398,7 @@ class PreTrainedModel(nn.Module, ModuleUtilsMixin, GenerationMixin):
     config_class = None
     base_model_prefix = ""
     authorized_missing_keys = None
+    authorized_unexpected_keys = None
     keys_to_never_save = None
 
     @property
@@ -1013,6 +1014,10 @@ class PreTrainedModel(nn.Module, ModuleUtilsMixin, GenerationMixin):
                 for pat in cls.authorized_missing_keys:
                     missing_keys = [k for k in missing_keys if re.search(pat, k) is None]
 
+            if cls.authorized_unexpected_keys is not None:
+                for pat in cls.authorized_unexpected_keys:
+                    unexpected_keys = [k for k in unexpected_keys if re.search(pat, k) is None]
+
             if len(unexpected_keys) > 0:
                 logger.warning(
                     f"Some weights of the model checkpoint at {pretrained_model_name_or_path} were not used when "