Force use_cache to be False in PyTorch (#15385)

* use_cache = False for PT models if labels is passed

* Fix for BigBirdPegasusForConditionalGeneration

* add warning if users specify use_cache=True

* Use logger.warning instead of warnings.warn

Co-authored-by: ydshieh <ydshieh@users.noreply.github.com>

templates/adding_a_new_model/cookiecutter-template-{{cookiecutter.modelname}}/modeling_{{cookiecutter.lowercase_modelname}}.py

@@ -2832,6 +2832,9 @@ class {{cookiecutter.camelcase_modelname}}ForConditionalGeneration({{cookiecutte
         return_dict = return_dict if return_dict is not None else self.config.use_return_dict
 
         if labels is not None:
+            if use_cache:
+                logger.warning("The `use_cache` argument is changed to `False` since `labels` is provided.")
+            use_cache = False
             if decoder_input_ids is None:
                 decoder_input_ids = shift_tokens_right(labels, self.config.pad_token_id, self.config.decoder_start_token_id)