New serving (#9419)

* Add a serving method

* Add albert

* Add serving for BERT and BART

* Add more models

* Finish the serving addition

* Temp fix

* Restore DPR

* Fix funnel attribute

* Fix attributes GPT2

* Fix OpenAIGPT attribute

* Fix T5 attributes

* Fix Bart attributes

* Fix TransfoXL attributes

* Add versioning

* better test

* Update template

* Fix Flaubert

* Fix T5

* Apply style

* Remove unused imports

* Deactivate extra parameters

* Remove too long test + saved_model default to False

* Ignore the saved model test for some models

* Fix some inputs

* Fix mpnet serving

* Trigger CI

* Address all comments

src/transformers/models/mpnet/modeling_tf_mpnet.py

@@ -71,6 +71,19 @@ class TFMPNetPreTrainedModel(TFPreTrainedModel):
     config_class = MPNetConfig
     base_model_prefix = "mpnet"
 
+    @tf.function(
+        input_signature=[
+            {
+                "input_ids": tf.TensorSpec((None, None), tf.int32, name="input_ids"),
+                "attention_mask": tf.TensorSpec((None, None), tf.int32, name="attention_mask"),
+            }
+        ]
+    )
+    def serving(self, inputs):
+        output = self.call(inputs)
+
+        return self.serving_output(output)
+
 
 class TFMPNetEmbeddings(tf.keras.layers.Layer):
     """Construct the embeddings from word, position embeddings."""
@@ -792,6 +805,17 @@ class TFMPNetModel(TFMPNetPreTrainedModel):
         )
         return outputs
 
+    def serving_output(self, output):
+        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 TFBaseModelOutputWithPooling(
+            last_hidden_state=output.last_hidden_state,
+            pooler_output=output.pooler_output,
+            hidden_states=hs,
+            attentions=attns,
+        )
+
 
 class TFMPNetLMHead(tf.keras.layers.Layer):
     """MPNet head for masked and permuted language modeling"""
@@ -918,6 +942,16 @@ class TFMPNetForMaskedLM(TFMPNetPreTrainedModel, TFMaskedLanguageModelingLoss):
             attentions=outputs.attentions,
         )
 
+    def serving_output(self, output):
+        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 TFMaskedLMOutput(
+            logits=output.logits,
+            hidden_states=hs,
+            attentions=attns,
+        )
+
 
 class TFMPNetClassificationHead(tf.keras.layers.Layer):
     """Head for sentence-level classification tasks."""
@@ -1035,6 +1069,16 @@ class TFMPNetForSequenceClassification(TFMPNetPreTrainedModel, TFSequenceClassif
             attentions=outputs.attentions,
         )
 
+    def serving_output(self, output):
+        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 TFSequenceClassifierOutput(
+            logits=output.logits,
+            hidden_states=hs,
+            attentions=attns,
+        )
+
 
 @add_start_docstrings(
     """
@@ -1159,6 +1203,29 @@ class TFMPNetForMultipleChoice(TFMPNetPreTrainedModel, TFMultipleChoiceLoss):
             attentions=outputs.attentions,
         )
 
+    @tf.function(
+        input_signature=[
+            {
+                "input_ids": tf.TensorSpec((None, None, None), tf.int32, name="input_ids"),
+                "attention_mask": tf.TensorSpec((None, None, None), tf.int32, name="attention_mask"),
+            }
+        ]
+    )
+    def serving(self, inputs):
+        output = self.call(inputs)
+
+        return self.serving_output(output)
+
+    def serving_output(self, output):
+        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 TFMultipleChoiceModelOutput(
+            logits=output.logits,
+            hidden_states=hs,
+            attentions=attns,
+        )
+
 
 @add_start_docstrings(
     """
@@ -1254,6 +1321,16 @@ class TFMPNetForTokenClassification(TFMPNetPreTrainedModel, TFTokenClassificatio
             attentions=outputs.attentions,
         )
 
+    def serving_output(self, output):
+        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 TFTokenClassifierOutput(
+            logits=output.logits,
+            hidden_states=hs,
+            attentions=attns,
+        )
+
 
 @add_start_docstrings(
     """
@@ -1361,3 +1438,14 @@ class TFMPNetForQuestionAnswering(TFMPNetPreTrainedModel, TFQuestionAnsweringLos
             hidden_states=outputs.hidden_states,
             attentions=outputs.attentions,
         )
+
+    def serving_output(self, output):
+        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 TFQuestionAnsweringModelOutput(
+            start_logits=output.start_logits,
+            end_logits=output.end_logits,
+            hidden_states=hs,
+            attentions=attns,
+        )