Overhaul TF serving signatures + dummy inputs (#23234)

* Let's try autodetecting serving sigs

* Don't clobber existing sigs

* Change shapes for multiplechoice models

* Make default dummy inputs smarter too

* Fix missing f-string

* Let's YOLO a serving output too

* Read __class__.__name__ properly

* Don't just pass naked lists in there and expect it to be okay

* Code cleanup

* Update default serving sig

* Clearer error messages

* Further updates to the default serving output

* make fixup

* Update the serving output a bit more

* Cleanups and renames, raise errors appropriately when we can't infer inputs

* More renames

* we're building in a functional context again, yolo

* import DUMMY_INPUTS from the right place

* import DUMMY_INPUTS from the right place

* Support cross-attention in the dummies

* Support cross-attention in the dummies

* Complete removal of dummy/serving overrides in BERT

* Complete removal of dummy/serving overrides in RoBERTa

* Obliterate lots and lots of serving sig and dummy overrides

* merge type hint changes

* Fix for token_type_ids with vocab_size 1

* Add missing property decorator

* Fix T5 and hopefully some models that take conv inputs

* More signature pruning

* Fix T5's signature

* Fix Wav2Vec2 signature

* Fix LongformerForMultipleChoice input signature

* Fix BLIP and LED

* Better default serving output error handling

* Fix BART dummies

* Fix dummies for cross-attention, esp encoder-decoder models

* Fix visionencoderdecoder signature

* Fix BLIP serving output

* Small tweak to BART dummies

* Cleanup the ugly parameter inspection line that I used in a few places

* committed a breakpoint again

* Move the text_dims check

* Remove blip_text serving_output

* Add decoder_input_ids to the default input sig

* Remove all the manual overrides for encoder-decoder model signatures

* Tweak longformer/led input sigs

* Tweak default serving output

* output.keys() -> output

* make fixup

src/transformers/models/xlm/modeling_tf_xlm.py

@@ -732,13 +732,6 @@ class TFXLMModel(TFXLMPreTrainedModel):
 
         return outputs
 
-    # Copied from transformers.models.distilbert.modeling_tf_distilbert.TFDistilBertModel.serving_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 TFBaseModelOutput(last_hidden_state=output.last_hidden_state, hidden_states=hs, attentions=attns)
-
 
 class TFXLMPredLayer(tf.keras.layers.Layer):
     """
@@ -876,12 +869,6 @@ class TFXLMWithLMHeadModel(TFXLMPreTrainedModel):
             logits=outputs, hidden_states=transformer_outputs.hidden_states, attentions=transformer_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 TFXLMWithLMHeadModelOutput(logits=output.logits, hidden_states=hs, attentions=attns)
-
 
 @add_start_docstrings(
     """
@@ -960,13 +947,6 @@ class TFXLMForSequenceClassification(TFXLMPreTrainedModel, TFSequenceClassificat
             attentions=transformer_outputs.attentions,
         )
 
-    # Copied from transformers.models.bert.modeling_tf_bert.TFBertForSequenceClassification.serving_output
-    def serving_output(self, output: TFSequenceClassifierOutput) -> TFSequenceClassifierOutput:
-        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(
     """
@@ -1086,28 +1066,6 @@ class TFXLMForMultipleChoice(TFXLMPreTrainedModel, TFMultipleChoiceLoss):
             attentions=transformer_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"),
-                "token_type_ids": tf.TensorSpec((None, None, None), tf.int32, name="token_type_ids"),
-            }
-        ]
-    )
-    # Copied from transformers.models.bert.modeling_tf_bert.TFBertForMultipleChoice.serving
-    def serving(self, inputs: Dict[str, tf.Tensor]):
-        output = self.call(input_ids=inputs)
-
-        return self.serving_output(output)
-
-    # Copied from transformers.models.bert.modeling_tf_bert.TFBertForMultipleChoice.serving_output
-    def serving_output(self, output: TFMultipleChoiceModelOutput) -> TFMultipleChoiceModelOutput:
-        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(
     """
@@ -1188,13 +1146,6 @@ class TFXLMForTokenClassification(TFXLMPreTrainedModel, TFTokenClassificationLos
             attentions=transformer_outputs.attentions,
         )
 
-    # Copied from transformers.models.bert.modeling_tf_bert.TFBertForTokenClassification.serving_output
-    def serving_output(self, output: TFTokenClassifierOutput) -> TFTokenClassifierOutput:
-        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(
     """
@@ -1285,12 +1236,3 @@ class TFXLMForQuestionAnsweringSimple(TFXLMPreTrainedModel, TFQuestionAnsweringL
             hidden_states=transformer_outputs.hidden_states,
             attentions=transformer_outputs.attentions,
         )
-
-    # Copied from transformers.models.bert.modeling_tf_bert.TFBertForQuestionAnswering.serving_output
-    def serving_output(self, output: TFQuestionAnsweringModelOutput) -> TFQuestionAnsweringModelOutput:
-        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
-        )

src/transformers/models/xlm_roberta/modeling_tf_xlm_roberta.py

@@ -51,8 +51,6 @@ from ...modeling_tf_utils import (
 )
 from ...tf_utils import check_embeddings_within_bounds, shape_list, stable_softmax
 from ...utils import (
-    DUMMY_INPUTS,
-    MULTIPLE_CHOICE_DUMMY_INPUTS,
     add_code_sample_docstrings,
     add_start_docstrings,
     add_start_docstrings_to_model_forward,
@@ -868,38 +866,6 @@ class TFXLMRobertaPreTrainedModel(TFPreTrainedModel):
     config_class = XLMRobertaConfig
     base_model_prefix = "roberta"
 
-    @property
-    # Copied from transformers.models.bert.modeling_tf_bert.TFBertPreTrainedModel.dummy_inputs
-    def dummy_inputs(self):
-        """
-        Dummy inputs to build the network.
-
-        Returns:
-            `Dict[str, tf.Tensor]`: The dummy inputs.
-        """
-        dummy = {"input_ids": tf.constant(DUMMY_INPUTS, dtype=tf.int32)}
-        # Add `encoder_hidden_states` to make the cross-attention layers' weights initialized
-        if self.config.add_cross_attention:
-            batch_size, seq_len = tf.constant(DUMMY_INPUTS).shape
-            shape = (batch_size, seq_len) + (self.config.hidden_size,)
-            h = tf.random.uniform(shape=shape)
-            dummy["encoder_hidden_states"] = h
-
-        return dummy
-
-    @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)
-
 
 @add_start_docstrings(
     "The bare XLM RoBERTa Model transformer outputting raw hidden-states without any specific head on top.",
@@ -974,27 +940,6 @@ class TFXLMRobertaModel(TFXLMRobertaPreTrainedModel):
 
         return outputs
 
-    # Copied from transformers.models.bert.modeling_tf_bert.TFBertModel.serving_output
-    def serving_output(
-        self, output: TFBaseModelOutputWithPoolingAndCrossAttentions
-    ) -> TFBaseModelOutputWithPoolingAndCrossAttentions:
-        output_cache = self.config.use_cache and self.config.is_decoder
-        pkv = tf.convert_to_tensor(output.past_key_values) if output_cache else None
-        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
-        cross_attns = tf.convert_to_tensor(output.cross_attentions) if output.cross_attentions is not None else None
-        if not (self.config.output_attentions and self.config.add_cross_attention):
-            cross_attns = None
-
-        return TFBaseModelOutputWithPoolingAndCrossAttentions(
-            last_hidden_state=output.last_hidden_state,
-            pooler_output=output.pooler_output,
-            past_key_values=pkv,
-            hidden_states=hs,
-            attentions=attns,
-            cross_attentions=cross_attns,
-        )
-
 
 # Copied from transformers.models.roberta.modeling_tf_roberta.TFRobertaLMHead with Roberta->XLMRoberta
 class TFXLMRobertaLMHead(tf.keras.layers.Layer):
@@ -1127,13 +1072,6 @@ class TFXLMRobertaForMaskedLM(TFXLMRobertaPreTrainedModel, TFMaskedLanguageModel
             attentions=outputs.attentions,
         )
 
-    # Copied from transformers.models.bert.modeling_tf_bert.TFBertForMaskedLM.serving_output
-    def serving_output(self, output: TFMaskedLMOutput) -> TFMaskedLMOutput:
-        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)
-
 
 @add_start_docstrings(
     "XLM-RoBERTa Model with a `language modeling` head on top for CLM fine-tuning.",
@@ -1261,20 +1199,6 @@ class TFXLMRobertaForCausalLM(TFXLMRobertaPreTrainedModel, TFCausalLanguageModel
             cross_attentions=outputs.cross_attentions,
         )
 
-    # Copied from transformers.models.bert.modeling_tf_bert.TFBertLMHeadModel.serving_output
-    def serving_output(self, output: TFCausalLMOutputWithCrossAttentions) -> TFCausalLMOutputWithCrossAttentions:
-        output_cache = self.config.use_cache and self.config.is_decoder
-        pkv = tf.convert_to_tensor(output.past_key_values) if output_cache else None
-        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
-        cross_attns = tf.convert_to_tensor(output.cross_attentions) if output.cross_attentions is not None else None
-        if not (self.config.output_attentions and self.config.add_cross_attention):
-            cross_attns = None
-
-        return TFCausalLMOutputWithCrossAttentions(
-            logits=output.logits, past_key_values=pkv, hidden_states=hs, attentions=attns, cross_attentions=cross_attns
-        )
-
 
 # Copied from transformers.models.roberta.modeling_tf_roberta.TFRobertaClassificationHead with Roberta->XLMRoberta
 class TFXLMRobertaClassificationHead(tf.keras.layers.Layer):
@@ -1381,13 +1305,6 @@ class TFXLMRobertaForSequenceClassification(TFXLMRobertaPreTrainedModel, TFSeque
             attentions=outputs.attentions,
         )
 
-    # Copied from transformers.models.bert.modeling_tf_bert.TFBertForSequenceClassification.serving_output
-    def serving_output(self, output: TFSequenceClassifierOutput) -> TFSequenceClassifierOutput:
-        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(
     """
@@ -1411,16 +1328,6 @@ class TFXLMRobertaForMultipleChoice(TFXLMRobertaPreTrainedModel, TFMultipleChoic
             1, kernel_initializer=get_initializer(config.initializer_range), name="classifier"
         )
 
-    @property
-    def dummy_inputs(self):
-        """
-        Dummy inputs to build the network.
-
-        Returns:
-            tf.Tensor with dummy inputs
-        """
-        return {"input_ids": tf.constant(MULTIPLE_CHOICE_DUMMY_INPUTS, dtype=tf.int32)}
-
     @unpack_inputs
     @add_start_docstrings_to_model_forward(
         XLM_ROBERTA_INPUTS_DOCSTRING.format("batch_size, num_choices, sequence_length")
@@ -1491,26 +1398,6 @@ class TFXLMRobertaForMultipleChoice(TFXLMRobertaPreTrainedModel, TFMultipleChoic
             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)
-
-    # Copied from transformers.models.bert.modeling_tf_bert.TFBertForMultipleChoice.serving_output
-    def serving_output(self, output: TFMultipleChoiceModelOutput) -> TFMultipleChoiceModelOutput:
-        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(
     """
@@ -1595,13 +1482,6 @@ class TFXLMRobertaForTokenClassification(TFXLMRobertaPreTrainedModel, TFTokenCla
             attentions=outputs.attentions,
         )
 
-    # Copied from transformers.models.bert.modeling_tf_bert.TFBertForTokenClassification.serving_output
-    def serving_output(self, output: TFTokenClassifierOutput) -> TFTokenClassifierOutput:
-        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(
     """
@@ -1694,12 +1574,3 @@ class TFXLMRobertaForQuestionAnswering(TFXLMRobertaPreTrainedModel, TFQuestionAn
             hidden_states=outputs.hidden_states,
             attentions=outputs.attentions,
         )
-
-    # Copied from transformers.models.bert.modeling_tf_bert.TFBertForQuestionAnswering.serving_output
-    def serving_output(self, output: TFQuestionAnsweringModelOutput) -> TFQuestionAnsweringModelOutput:
-        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
-        )

src/transformers/models/xlnet/modeling_tf_xlnet.py

@@ -44,7 +44,6 @@ from ...modeling_tf_utils import (
 )
 from ...tf_utils import check_embeddings_within_bounds, shape_list, stable_softmax
 from ...utils import (
-    MULTIPLE_CHOICE_DUMMY_INPUTS,
     ModelOutput,
     add_code_sample_docstrings,
     add_start_docstrings,
@@ -1177,15 +1176,6 @@ class TFXLNetModel(TFXLNetPreTrainedModel):
 
         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
-        mems = tf.convert_to_tensor(output.mems) if output.mems is not None else None
-
-        return TFXLNetModelOutput(
-            last_hidden_state=output.last_hidden_state, mems=mems, hidden_states=hs, attentions=attns
-        )
-
 
 @add_start_docstrings(
     """
@@ -1345,13 +1335,6 @@ class TFXLNetLMHeadModel(TFXLNetPreTrainedModel, TFCausalLanguageModelingLoss):
             attentions=transformer_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
-        mems = tf.convert_to_tensor(output.mems) if output.mems is not None else None
-
-        return TFXLNetLMHeadModelOutput(logits=output.logits, mems=mems, hidden_states=hs, attentions=attns)
-
 
 @add_start_docstrings(
     """
@@ -1439,15 +1422,6 @@ class TFXLNetForSequenceClassification(TFXLNetPreTrainedModel, TFSequenceClassif
             attentions=transformer_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
-        mems = tf.convert_to_tensor(output.mems) if output.mems is not None else None
-
-        return TFXLNetForSequenceClassificationOutput(
-            logits=output.logits, mems=mems, hidden_states=hs, attentions=attns
-        )
-
 
 @add_start_docstrings(
     """
@@ -1468,16 +1442,6 @@ class TFXLNetForMultipleChoice(TFXLNetPreTrainedModel, TFMultipleChoiceLoss):
             1, kernel_initializer=get_initializer(config.initializer_range), name="logits_proj"
         )
 
-    @property
-    def dummy_inputs(self):
-        """
-        Dummy inputs to build the network.
-
-        Returns:
-            tf.Tensor with dummy inputs
-        """
-        return {"input_ids": tf.constant(MULTIPLE_CHOICE_DUMMY_INPUTS, dtype=tf.int32)}
-
     @unpack_inputs
     @add_start_docstrings_to_model_forward(XLNET_INPUTS_DOCSTRING.format("batch_size, num_choices, sequence_length"))
     @add_code_sample_docstrings(
@@ -1559,27 +1523,6 @@ class TFXLNetForMultipleChoice(TFXLNetPreTrainedModel, TFMultipleChoiceLoss):
             attentions=transformer_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"),
-                "token_type_ids": tf.TensorSpec((None, None, None), tf.int32, name="token_type_ids"),
-            }
-        ]
-    )
-    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
-        mems = tf.convert_to_tensor(output.mems) if output.mems is not None else None
-
-        return TFXLNetForMultipleChoiceOutput(logits=output.logits, mems=mems, hidden_states=hs, attentions=attns)
-
 
 @add_start_docstrings(
     """
@@ -1660,13 +1603,6 @@ class TFXLNetForTokenClassification(TFXLNetPreTrainedModel, TFTokenClassificatio
             attentions=transformer_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
-        mems = tf.convert_to_tensor(output.mems) if output.mems is not None else None
-
-        return TFXLNetForTokenClassificationOutput(logits=output.logits, mems=mems, hidden_states=hs, attentions=attns)
-
 
 @add_start_docstrings(
     """
@@ -1760,16 +1696,3 @@ class TFXLNetForQuestionAnsweringSimple(TFXLNetPreTrainedModel, TFQuestionAnswer
             hidden_states=transformer_outputs.hidden_states,
             attentions=transformer_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
-        mems = tf.convert_to_tensor(output.mems) if output.mems is not None else None
-
-        return TFXLNetForQuestionAnsweringSimpleOutput(
-            start_logits=output.start_logits,
-            end_logits=output.end_logits,
-            mems=mems,
-            hidden_states=hs,
-            attentions=attns,
-        )

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

@@ -803,23 +803,6 @@ class TF{{cookiecutter.camelcase_modelname}}PreTrainedModel(TFPreTrainedModel):
     config_class = {{cookiecutter.camelcase_modelname}}Config
     base_model_prefix = "{{cookiecutter.lowercase_modelname}}"
 
-    @property
-    def dummy_inputs(self):
-        """
-        Dummy inputs to build the network.
-
-        Returns:
-            `Dict[str, tf.Tensor]`: The dummy inputs.
-        """
-        dummy = {"input_ids": tf.constant(DUMMY_INPUTS, dtype=tf.int64)}
-        # Add `encoder_hidden_states` to make the cross-attention layers' weights initialized
-        if self.config.add_cross_attention:
-            batch_size, seq_len = tf.constant(DUMMY_INPUTS).shape
-            shape = (batch_size, seq_len) + (self.config.hidden_size,)
-            h = tf.random.uniform(shape=shape)
-            dummy["encoder_hidden_states"] = h
-
-        return dummy
 
 
 {{cookiecutter.uppercase_modelname}}_START_DOCSTRING = r"""
@@ -991,24 +974,6 @@ class TF{{cookiecutter.camelcase_modelname}}Model(TF{{cookiecutter.camelcase_mod
 
         return outputs
 
-    def serving_output(
-        self, output: TFBaseModelOutputWithPastAndCrossAttentions
-    ) -> TFBaseModelOutputWithPastAndCrossAttentions:
-        output_cache = self.config.use_cache and self.config.is_decoder
-        pkv = tf.convert_to_tensor(output.past_key_values) if output_cache else None
-        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
-        cross_attns = tf.convert_to_tensor(output.cross_attentions) if output.cross_attentions is not None else None
-        if not (self.config.output_attentions and self.config.add_cross_attention):
-            cross_attns = None
-
-        return TFBaseModelOutputWithPastAndCrossAttentions(
-            last_hidden_state=output.last_hidden_state,
-            past_key_values=pkv,
-            hidden_states=hs,
-            attentions=attns,
-            cross_attentions=cross_attns,
-        )
 
 
 @add_start_docstrings("""{{cookiecutter.modelname}} Model with a `language modeling` head on top. """, {{cookiecutter.uppercase_modelname}}_START_DOCSTRING)
@@ -1084,13 +1049,6 @@ class TF{{cookiecutter.camelcase_modelname}}ForMaskedLM(TF{{cookiecutter.camelca
             attentions=outputs.attentions,
         )
 
-    # Copied from transformers.models.bert.modeling_tf_bert.TFBertForMaskedLM.serving_output
-    def serving_output(self, output: TFMaskedLMOutput) -> TFMaskedLMOutput:
-        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)
-
 
 @add_start_docstrings(
     """{{cookiecutter.modelname}} Model with a `language modeling` head on top for CLM fine-tuning. """, {{cookiecutter.uppercase_modelname}}_START_DOCSTRING
@@ -1206,19 +1164,6 @@ class TF{{cookiecutter.camelcase_modelname}}ForCausalLM(TF{{cookiecutter.camelca
             cross_attentions=outputs.cross_attentions,
         )
 
-    # Copied from transformers.models.bert.modeling_tf_bert.TFBertLMHeadModel.serving_output
-    def serving_output(self, output: TFCausalLMOutputWithCrossAttentions) -> TFCausalLMOutputWithCrossAttentions:
-        output_cache = self.config.use_cache and self.config.is_decoder
-        pkv = tf.convert_to_tensor(output.past_key_values) if output_cache else None
-        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
-        cross_attns = tf.convert_to_tensor(output.cross_attentions) if output.cross_attentions is not None else None
-        if not (self.config.output_attentions and self.config.add_cross_attention):
-            cross_attns = None
-
-        return TFCausalLMOutputWithCrossAttentions(
-            logits=output.logits, past_key_values=pkv, hidden_states=hs, attentions=attns, cross_attentions=cross_attns
-        )
 
 
 class TF{{cookiecutter.camelcase_modelname}}ClassificationHead(tf.keras.layers.Layer):
@@ -1318,13 +1263,6 @@ class TF{{cookiecutter.camelcase_modelname}}ForSequenceClassification(TF{{cookie
             attentions=outputs.attentions,
         )
 
-    # Copied from transformers.models.bert.modeling_tf_bert.TFBertForSequenceClassification.serving_output
-    def serving_output(self, output: TFSequenceClassifierOutput) -> TFSequenceClassifierOutput:
-        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(
     """{{cookiecutter.modelname}} Model with a multiple choice classification head on top (a linear layer on top of
@@ -1343,16 +1281,6 @@ class TF{{cookiecutter.camelcase_modelname}}ForMultipleChoice(TF{{cookiecutter.c
             units=1, kernel_initializer=get_initializer(config.initializer_range), name="classifier"
         )
 
-    @property
-    def dummy_inputs(self) -> Dict[str, tf.Tensor]:
-        """
-        Dummy inputs to build the network.
-
-        Returns:
-            tf.Tensor with dummy inputs
-        """
-        return {"input_ids": tf.constant(MULTIPLE_CHOICE_DUMMY_INPUTS, dtype=tf.int64)}
-
     @unpack_inputs
     @add_start_docstrings_to_model_forward({{cookiecutter.uppercase_modelname}}_INPUTS_DOCSTRING.format("batch_size, num_choices, sequence_length"))
     @add_code_sample_docstrings(
@@ -1441,24 +1369,6 @@ class TF{{cookiecutter.camelcase_modelname}}ForMultipleChoice(TF{{cookiecutter.c
             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"),
-        "token_type_ids": tf.TensorSpec((None, None, None), tf.int32, name="token_type_ids"),
-    }])
-    # Copied from transformers.models.bert.modeling_tf_bert.TFBertForMultipleChoice.serving
-    def serving(self, inputs: Dict[str, tf.Tensor]) -> TFMultipleChoiceModelOutput:
-        output = self.call(input_ids=inputs)
-
-        return self.serving_output(output)
-
-    # Copied from transformers.models.bert.modeling_tf_bert.TFBertForMultipleChoice.serving_output
-    def serving_output(self, output: TFMultipleChoiceModelOutput) -> TFMultipleChoiceModelOutput:
-        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(
     """{{cookiecutter.modelname}} Model with a token classification head on top (a linear layer on top of
@@ -1532,13 +1442,6 @@ class TF{{cookiecutter.camelcase_modelname}}ForTokenClassification(TF{{cookiecut
             attentions=outputs.attentions,
         )
 
-    # Copied from transformers.models.bert.modeling_tf_bert.TFBertForTokenClassification.serving_output
-    def serving_output(self, output: TFTokenClassifierOutput) -> TFTokenClassifierOutput:
-        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(
     """{{cookiecutter.modelname}} Model with a span classification head on top for extractive question-answering tasks like SQuAD (a linear
@@ -1625,14 +1528,6 @@ class TF{{cookiecutter.camelcase_modelname}}ForQuestionAnswering(TF{{cookiecutte
             attentions=outputs.attentions,
         )
 
-    # Copied from transformers.models.bert.modeling_tf_bert.TFBertForQuestionAnswering.serving_output
-    def serving_output(self, output: TFQuestionAnsweringModelOutput) -> TFQuestionAnsweringModelOutput:
-        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
-        )
 
 {% else %}
 import random
@@ -2777,26 +2672,6 @@ class TF{{cookiecutter.camelcase_modelname}}Model(TF{{cookiecutter.camelcase_mod
 
         return outputs
 
-    # Copied from transformers.models.bart.modeling_tf_bart.TFBartModel.serving_output
-    def serving_output(self, output):
-        pkv = tf.tuple(output.past_key_values)[1] if self.config.use_cache else None
-        dec_hs = tf.convert_to_tensor(output.decoder_hidden_states) if self.config.output_hidden_states else None
-        dec_attns = tf.convert_to_tensor(output.decoder_attentions) if self.config.output_attentions else None
-        cross_attns = tf.convert_to_tensor(output.cross_attentions) if self.config.output_attentions else None
-        enc_hs = tf.convert_to_tensor(output.encoder_hidden_states) if self.config.output_hidden_states else None
-        enc_attns = tf.convert_to_tensor(output.encoder_attentions) if self.config.output_attentions else None
-
-        return TFSeq2SeqModelOutput(
-            last_hidden_state=output.last_hidden_state,
-            past_key_values=pkv,
-            decoder_hidden_states=dec_hs,
-            decoder_attentions=dec_attns,
-            cross_attentions=cross_attns,
-            encoder_last_hidden_state=output.encoder_last_hidden_state,
-            encoder_hidden_states=enc_hs,
-            encoder_attentions=enc_attns,
-        )
-
 
 # Copied from transformers.models.bart.modeling_tf_bart.BiasLayer
 class BiasLayer(tf.keras.layers.Layer):
@@ -2944,26 +2819,6 @@ class TF{{cookiecutter.camelcase_modelname}}ForConditionalGeneration(TF{{cookiec
             encoder_attentions=outputs.encoder_attentions,  # 2 of e out
         )
 
-    # Copied from transformers.models.bart.modeling_tf_bart.TFBartForConditionalGeneration.serving_output
-    def serving_output(self, output):
-        pkv = tf.tuple(output.past_key_values)[1] if self.config.use_cache else None
-        dec_hs = tf.convert_to_tensor(output.decoder_hidden_states) if self.config.output_hidden_states else None
-        dec_attns = tf.convert_to_tensor(output.decoder_attentions) if self.config.output_attentions else None
-        cross_attns = tf.convert_to_tensor(output.cross_attentions) if self.config.output_attentions else None
-        enc_hs = tf.convert_to_tensor(output.encoder_hidden_states) if self.config.output_hidden_states else None
-        enc_attns = tf.convert_to_tensor(output.encoder_attentions) if self.config.output_attentions else None
-
-        return TFSeq2SeqLMOutput(
-            logits=output.logits,
-            past_key_values=pkv,
-            decoder_hidden_states=dec_hs,
-            decoder_attentions=dec_attns,
-            cross_attentions=cross_attns,
-            encoder_last_hidden_state=output.encoder_last_hidden_state,
-            encoder_hidden_states=enc_hs,
-            encoder_attentions=enc_attns,
-        )
-
     def prepare_inputs_for_generation(
         self,
         decoder_input_ids,

tests/test_modeling_tf_common.py

@@ -1677,7 +1677,10 @@ class TFModelTesterMixin:
 
             # After testing that the model accepts all int inputs, confirm that its dummies are int32
             for key, tensor in model.dummy_inputs.items():
-                self.assertTrue(isinstance(tensor, tf.Tensor), "Dummy inputs should be tf.Tensor!")
+                self.assertTrue(
+                    isinstance(tensor, tf.Tensor) or tf.keras.backend.is_keras_tensor(tensor),
+                    "Dummy inputs should be tf.Tensor!",
+                )
                 if tensor.dtype.is_integer:
                     self.assertTrue(tensor.dtype == tf.int32, "Integer dummy inputs should be tf.int32!")