TF: Merge PT and TF behavior for Bart when no decoder_input_ids are passed (#17593)

* Merge PT and TF behavior

src/transformers/models/bart/modeling_tf_bart.py

@@ -1073,14 +1073,16 @@ class TFBartMainLayer(tf.keras.layers.Layer):
         **kwargs
     ) -> Union[TFSeq2SeqModelOutput, Tuple[tf.Tensor]]:
 
+        # different to other models, Bart automatically creates decoder_input_ids from
+        # input_ids if no decoder_input_ids are provided
         if decoder_input_ids is None and decoder_inputs_embeds is None:
-            use_cache = False
-
-        output_hidden_states = (
-            output_hidden_states if output_hidden_states is not None else self.config.output_hidden_states
+            if input_ids is None:
+                raise ValueError(
+                    "If no `decoder_input_ids` or `decoder_inputs_embeds` are "
+                    "passed, `input_ids` cannot be `None`. Please pass either "
+                    "`input_ids` or `decoder_input_ids` or `decoder_inputs_embeds`."
                 )
 
-        if decoder_input_ids is None and input_ids is not None:
             decoder_input_ids = shift_tokens_right(
                 input_ids, self.config.pad_token_id, self.config.decoder_start_token_id
             )