Fix BLIP and LED

src/transformers/models/blip/modeling_tf_blip.py

@@ -1187,6 +1187,12 @@ def _shift_right(self, input_ids):
 
         return shifted_input_ids
 
+    @property
+    def input_signature(self):
+        base_sig = super().input_signature
+        base_sig["decoder_input_ids"] = base_sig["input_ids"]
+        return base_sig
+
     @unpack_inputs
     @add_start_docstrings_to_model_forward(BLIP_VISION_INPUTS_DOCSTRING)
     @replace_return_docstrings(output_type=TFBlipTextVisionModelOutput, config_class=BlipVisionConfig)
@@ -1239,7 +1245,7 @@ def call(
         ```"""
         if labels is None and decoder_input_ids is None:
             raise ValueError(
-                "Either `decoder_input_ids` or `labels` should be passed when calling `forward` with"
+                "Either `decoder_input_ids` or `labels` should be passed when calling"
                 " `TFBlipForQuestionAnswering`. if you are training the model make sure that `labels` is passed, if you"
                 " are using the model for inference make sure that `decoder_input_ids` is passed or call `generate`"
             )

src/transformers/models/blip/modeling_tf_blip_text.py

@@ -797,6 +797,19 @@ def call(
             cross_attentions=encoder_outputs.cross_attentions,
         )
 
+    def serving_output(
+        self, output: TFBaseModelOutputWithPoolingAndCrossAttentions
+    ) -> TFBaseModelOutputWithPoolingAndCrossAttentions:
+        hs = tf.convert_to_tensor(output.hidden_states) if self.config.output_hidden_states else None
+        attns = tf.convert_to_tensor(output.attentions) if self.config.output_attentions else None
+
+        return TFBaseModelOutputWithPoolingAndCrossAttentions(
+            last_hidden_state=output.last_hidden_state,
+            pooler_output=output.pooler_output,
+            hidden_states=hs,
+            attentions=attns,
+        )
+
 
 # Adapted from https://github.com/salesforce/BLIP/blob/main/models/med.py#L811
 class TFBlipTextLMHeadModel(TFBlipTextPreTrainedModel):

src/transformers/models/led/modeling_tf_led.py

@@ -1327,6 +1327,7 @@ def input_signature(self):
             "attention_mask": tf.TensorSpec((None, None), tf.int32, name="attention_mask"),
             "decoder_input_ids": tf.TensorSpec((None, None), tf.int32, name="decoder_input_ids"),
             "decoder_attention_mask": tf.TensorSpec((None, None), tf.int32, name="decoder_attention_mask"),
+            "global_attention_mask": tf.TensorSpec((None, None), tf.int32, name="global_attention_mask"),
         }