more comments

tests/test_modeling_common.py

@@ -3726,8 +3726,10 @@ def get_mean_reldiff(failcase, x, ref, atol, rtol):
         for model_class in self.all_model_classes:
             config, inputs_dict = self.model_tester.prepare_config_and_inputs_for_common()
             model = model_class(config)
-            # FIXME: we deactivate boolean mask because pretrained models
-            # will not load the mask token
+            # FIXME: we deactivate boolean mask because models using "use_mask_token"
+            # as input to their constructor will create another `mask_token` attribute
+            # **only** if `use_mask_token` is `True`. For these models we will skip creating the attention
+            # mask. A simple fix would be to include it in the config.
             deactivate_mask = "use_mask_token" in inspect.signature(model_class).parameters
 
             is_encoder_decoder = model.config.is_encoder_decoder