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/opt/modeling_tf_opt.py

@@ -27,7 +27,6 @@ from ...modeling_tf_outputs import TFBaseModelOutputWithPast, TFCausalLMOutputWi
 
 # Public API
 from ...modeling_tf_utils import (
-    DUMMY_INPUTS,
     TFCausalLanguageModelingLoss,
     TFModelInputType,
     TFPreTrainedModel,
@@ -413,29 +412,6 @@ class TFOPTPreTrainedModel(TFPreTrainedModel):
     config_class = OPTConfig
     base_model_prefix = "model"
 
-    @property
-    def dummy_inputs(self):
-        pad_token = 1
-        input_ids = tf.convert_to_tensor(DUMMY_INPUTS, dtype=tf.int32)
-        dummy_inputs = {
-            "attention_mask": tf.cast(input_ids != pad_token, tf.int32),
-            "input_ids": input_ids,
-        }
-        return dummy_inputs
-
-    @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)
-
 
 OPT_INPUTS_DOCSTRING = r"""
     Args:

src/transformers/models/pegasus/modeling_tf_pegasus.py

@@ -33,7 +33,6 @@ from ...modeling_tf_outputs import (
 
 # Public API
 from ...modeling_tf_utils import (
-    DUMMY_INPUTS,
     TFCausalLanguageModelingLoss,
     TFModelInputType,
     TFPreTrainedModel,
@@ -503,34 +502,6 @@ class TFPegasusPreTrainedModel(TFPreTrainedModel):
     config_class = PegasusConfig
     base_model_prefix = "model"
 
-    @property
-    def dummy_inputs(self):
-        pad_token = 1
-        input_ids = tf.convert_to_tensor(DUMMY_INPUTS, dtype=tf.int32)
-        decoder_input_ids = tf.convert_to_tensor(DUMMY_INPUTS, dtype=tf.int32)
-        dummy_inputs = {
-            "decoder_input_ids": decoder_input_ids,
-            "attention_mask": tf.cast(input_ids != pad_token, tf.int32),
-            "input_ids": input_ids,
-        }
-        return dummy_inputs
-
-    @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"),
-                "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"),
-            }
-        ]
-    )
-    # Copied from transformers.models.bart.modeling_tf_bart.TFBartPretrainedModel.serving
-    def serving(self, inputs):
-        output = self.call(inputs)
-
-        return self.serving_output(output)
-
 
 PEGASUS_START_DOCSTRING = r"""
     This model inherits from [`TFPreTrainedModel`]. Check the superclass documentation for the generic methods the

src/transformers/models/regnet/modeling_tf_regnet.py

@@ -14,7 +14,7 @@
 # limitations under the License.
 """ TensorFlow RegNet model."""
 
-from typing import Dict, Optional, Tuple, Union
+from typing import Optional, Tuple, Union
 
 import tensorflow as tf
 
@@ -345,33 +345,8 @@ class TFRegNetPreTrainedModel(TFPreTrainedModel):
     main_input_name = "pixel_values"
 
     @property
-    def dummy_inputs(self) -> Dict[str, tf.Tensor]:
-        """
-        Dummy inputs to build the network.
-
-        Returns:
-            `Dict[str, tf.Tensor]`: The dummy inputs.
-        """
-        VISION_DUMMY_INPUTS = tf.random.uniform(shape=(3, self.config.num_channels, 224, 224), dtype=tf.float32)
-        return {"pixel_values": tf.constant(VISION_DUMMY_INPUTS)}
-
-    @tf.function(
-        input_signature=[
-            {
-                "pixel_values": tf.TensorSpec((None, None, None, None), tf.float32, name="pixel_values"),
-            }
-        ]
-    )
-    def serving(self, inputs):
-        """
-        Method used for serving the model.
-
-        Args:
-            inputs (`Dict[str, tf.Tensor]`):
-                The input of the saved model as a dictionary of tensors.
-        """
-        output = self.call(inputs)
-        return self.serving_output(output)
+    def input_signature(self):
+        return {"pixel_values": tf.TensorSpec(shape=(None, self.config.num_channels, 224, 224), dtype=tf.float32)}
 
 
 REGNET_START_DOCSTRING = r"""
@@ -443,16 +418,6 @@ class TFRegNetModel(TFRegNetPreTrainedModel):
             hidden_states=outputs.hidden_states,
         )
 
-    def serving_output(
-        self, output: TFBaseModelOutputWithPoolingAndNoAttention
-    ) -> TFBaseModelOutputWithPoolingAndNoAttention:
-        # hidden_states not converted to Tensor with tf.convert_to_tensor as they are all of different dimensions
-        return TFBaseModelOutputWithPoolingAndNoAttention(
-            last_hidden_state=output.last_hidden_state,
-            pooler_output=output.pooler_output,
-            hidden_states=output.hidden_states,
-        )
-
 
 @add_start_docstrings(
     """
@@ -514,7 +479,3 @@ class TFRegNetForImageClassification(TFRegNetPreTrainedModel, TFSequenceClassifi
             return ((loss,) + output) if loss is not None else output
 
         return TFSequenceClassifierOutput(loss=loss, logits=logits, hidden_states=outputs.hidden_states)
-
-    def serving_output(self, output: TFSequenceClassifierOutput) -> TFSequenceClassifierOutput:
-        # hidden_states not converted to Tensor with tf.convert_to_tensor as they are all of different dimensions
-        return TFSequenceClassifierOutput(logits=output.logits, hidden_states=output.hidden_states)

src/transformers/models/rembert/modeling_tf_rembert.py

@@ -49,8 +49,6 @@ 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,
@@ -812,24 +810,6 @@ class TFRemBertPreTrainedModel(TFPreTrainedModel):
     config_class = RemBertConfig
     base_model_prefix = "rembert"
 
-    @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.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
-
 
 REMBERT_START_DOCSTRING = r"""
 
@@ -1002,27 +982,6 @@ class TFRemBertModel(TFRemBertPreTrainedModel):
 
         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,
-        )
-
 
 @add_start_docstrings("""RemBERT Model with a `language modeling` head on top.""", REMBERT_START_DOCSTRING)
 class TFRemBertForMaskedLM(TFRemBertPreTrainedModel, TFMaskedLanguageModelingLoss):
@@ -1095,12 +1054,6 @@ class TFRemBertForMaskedLM(TFRemBertPreTrainedModel, TFMaskedLanguageModelingLos
             attentions=outputs.attentions,
         )
 
-    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(
     """RemBERT Model with a `language modeling` head on top for CLM fine-tuning.""", REMBERT_START_DOCSTRING
@@ -1217,20 +1170,6 @@ class TFRemBertForCausalLM(TFRemBertPreTrainedModel, TFCausalLanguageModelingLos
             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
-        )
-
 
 @add_start_docstrings(
     """
@@ -1307,12 +1246,6 @@ class TFRemBertForSequenceClassification(TFRemBertPreTrainedModel, TFSequenceCla
             attentions=outputs.attentions,
         )
 
-    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(
     """
@@ -1331,16 +1264,6 @@ class TFRemBertForMultipleChoice(TFRemBertPreTrainedModel, TFMultipleChoiceLoss)
             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.int32)}
-
     @unpack_inputs
     @add_start_docstrings_to_model_forward(REMBERT_INPUTS_DOCSTRING.format("batch_size, num_choices, sequence_length"))
     @add_code_sample_docstrings(
@@ -1419,26 +1342,6 @@ class TFRemBertForMultipleChoice(TFRemBertPreTrainedModel, 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"),
-                "token_type_ids": tf.TensorSpec((None, None, None), tf.int32, name="token_type_ids"),
-            }
-        ]
-    )
-    def serving(self, inputs: Dict[str, tf.Tensor]) -> TFMultipleChoiceModelOutput:
-        output = self.call(input_ids=inputs)
-
-        return self.serving_output(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(
     """
@@ -1512,12 +1415,6 @@ class TFRemBertForTokenClassification(TFRemBertPreTrainedModel, TFTokenClassific
             attentions=outputs.attentions,
         )
 
-    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(
     """
@@ -1604,11 +1501,3 @@ class TFRemBertForQuestionAnswering(TFRemBertPreTrainedModel, TFQuestionAnswerin
             hidden_states=outputs.hidden_states,
             attentions=outputs.attentions,
         )
-
-    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/resnet/modeling_tf_resnet.py

@@ -14,7 +14,7 @@
 # limitations under the License.
 """ TensorFlow ResNet model."""
 
-from typing import Dict, Optional, Tuple, Union
+from typing import Optional, Tuple, Union
 
 import tensorflow as tf
 
@@ -276,24 +276,8 @@ class TFResNetPreTrainedModel(TFPreTrainedModel):
     main_input_name = "pixel_values"
 
     @property
-    def dummy_inputs(self) -> Dict[str, tf.Tensor]:
-        """
-        Dummy inputs to build the network. Returns:
-            `Dict[str, tf.Tensor]`: The dummy inputs.
-        """
-        VISION_DUMMY_INPUTS = tf.random.uniform(shape=(3, self.config.num_channels, 224, 224), dtype=tf.float32)
-        return {"pixel_values": tf.constant(VISION_DUMMY_INPUTS)}
-
-    @tf.function(
-        input_signature=[
-            {
-                "pixel_values": tf.TensorSpec((None, None, None, None), tf.float32, name="pixel_values"),
-            }
-        ]
-    )
-    def serving(self, inputs):
-        output = self.call(inputs)
-        return self.serving_output(output)
+    def input_signature(self):
+        return {"pixel_values": tf.TensorSpec(shape=(None, self.config.num_channels, 224, 224), dtype=tf.float32)}
 
 
 RESNET_START_DOCSTRING = r"""
@@ -419,16 +403,6 @@ class TFResNetModel(TFResNetPreTrainedModel):
         )
         return resnet_outputs
 
-    def serving_output(
-        self, output: TFBaseModelOutputWithPoolingAndNoAttention
-    ) -> TFBaseModelOutputWithPoolingAndNoAttention:
-        # hidden_states not converted to Tensor with tf.convert_to_tensor as they are all of different dimensions
-        return TFBaseModelOutputWithPoolingAndNoAttention(
-            last_hidden_state=output.last_hidden_state,
-            pooler_output=output.pooler_output,
-            hidden_states=output.hidden_states,
-        )
-
 
 @add_start_docstrings(
     """
@@ -492,7 +466,3 @@ class TFResNetForImageClassification(TFResNetPreTrainedModel, TFSequenceClassifi
             return (loss,) + output if loss is not None else output
 
         return TFImageClassifierOutputWithNoAttention(loss=loss, logits=logits, hidden_states=outputs.hidden_states)
-
-    def serving_output(self, output: TFImageClassifierOutputWithNoAttention) -> TFImageClassifierOutputWithNoAttention:
-        # hidden_states not converted to Tensor with tf.convert_to_tensor as they are all of different dimensions
-        return TFImageClassifierOutputWithNoAttention(logits=output.logits, hidden_states=output.hidden_states)

src/transformers/models/roberta/modeling_tf_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,
@@ -777,38 +775,6 @@ class TFRobertaPreTrainedModel(TFPreTrainedModel):
     config_class = RobertaConfig
     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)
-
 
 ROBERTA_START_DOCSTRING = r"""
 
@@ -980,27 +946,6 @@ class TFRobertaModel(TFRobertaPreTrainedModel):
 
         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,
-        )
-
 
 class TFRobertaLMHead(tf.keras.layers.Layer):
     """Roberta Head for masked language modeling."""
@@ -1131,13 +1076,6 @@ class TFRobertaForMaskedLM(TFRobertaPreTrainedModel, TFMaskedLanguageModelingLos
             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)
-
 
 class TFRobertaForCausalLM(TFRobertaPreTrainedModel, TFCausalLanguageModelingLoss):
     # names with a '.' represents the authorized unexpected/missing layers when a TF model is loaded from a PT model
@@ -1260,20 +1198,6 @@ class TFRobertaForCausalLM(TFRobertaPreTrainedModel, TFCausalLanguageModelingLos
             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 TFRobertaClassificationHead(tf.keras.layers.Layer):
     """Head for sentence-level classification tasks."""
@@ -1378,13 +1302,6 @@ class TFRobertaForSequenceClassification(TFRobertaPreTrainedModel, TFSequenceCla
             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(
     """
@@ -1407,16 +1324,6 @@ class TFRobertaForMultipleChoice(TFRobertaPreTrainedModel, TFMultipleChoiceLoss)
             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(ROBERTA_INPUTS_DOCSTRING.format("batch_size, num_choices, sequence_length"))
     @add_code_sample_docstrings(
@@ -1485,26 +1392,6 @@ class TFRobertaForMultipleChoice(TFRobertaPreTrainedModel, 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)
-
-    # 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(
     """
@@ -1588,13 +1475,6 @@ class TFRobertaForTokenClassification(TFRobertaPreTrainedModel, TFTokenClassific
             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(
     """
@@ -1686,12 +1566,3 @@ class TFRobertaForQuestionAnswering(TFRobertaPreTrainedModel, TFQuestionAnswerin
             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/roberta_prelayernorm/modeling_tf_roberta_prelayernorm.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,
@@ -778,38 +776,6 @@ class TFRobertaPreLayerNormPreTrainedModel(TFPreTrainedModel):
     config_class = RobertaPreLayerNormConfig
     base_model_prefix = "roberta_prelayernorm"
 
-    @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)
-
 
 ROBERTA_PRELAYERNORM_START_DOCSTRING = r"""
 
@@ -982,27 +948,6 @@ class TFRobertaPreLayerNormModel(TFRobertaPreLayerNormPreTrainedModel):
 
         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->RobertaPreLayerNorm
 class TFRobertaPreLayerNormLMHead(tf.keras.layers.Layer):
@@ -1140,13 +1085,6 @@ class TFRobertaPreLayerNormForMaskedLM(TFRobertaPreLayerNormPreTrainedModel, TFM
             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)
-
 
 # Copied from transformers.models.roberta.modeling_tf_roberta.TFRobertaForCausalLM with ROBERTA->ROBERTA_PRELAYERNORM,Roberta->RobertaPreLayerNorm,roberta->roberta_prelayernorm
 class TFRobertaPreLayerNormForCausalLM(TFRobertaPreLayerNormPreTrainedModel, TFCausalLanguageModelingLoss):
@@ -1276,20 +1214,6 @@ class TFRobertaPreLayerNormForCausalLM(TFRobertaPreLayerNormPreTrainedModel, TFC
             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->RobertaPreLayerNorm
 class TFRobertaPreLayerNormClassificationHead(tf.keras.layers.Layer):
@@ -1398,13 +1322,6 @@ class TFRobertaPreLayerNormForSequenceClassification(
             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(
     """
@@ -1428,16 +1345,6 @@ class TFRobertaPreLayerNormForMultipleChoice(TFRobertaPreLayerNormPreTrainedMode
             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(
         ROBERTA_PRELAYERNORM_INPUTS_DOCSTRING.format("batch_size, num_choices, sequence_length")
@@ -1508,26 +1415,6 @@ class TFRobertaPreLayerNormForMultipleChoice(TFRobertaPreLayerNormPreTrainedMode
             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(
     """
@@ -1612,13 +1499,6 @@ class TFRobertaPreLayerNormForTokenClassification(TFRobertaPreLayerNormPreTraine
             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(
     """
@@ -1711,12 +1591,3 @@ class TFRobertaPreLayerNormForQuestionAnswering(TFRobertaPreLayerNormPreTrainedM
             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/roformer/modeling_tf_roformer.py

@@ -50,7 +50,6 @@ from ...modeling_tf_utils import (
 )
 from ...tf_utils import check_embeddings_within_bounds, shape_list, stable_softmax
 from ...utils import (
-    MULTIPLE_CHOICE_DUMMY_INPUTS,
     add_code_sample_docstrings,
     add_start_docstrings,
     add_start_docstrings_to_model_forward,
@@ -835,12 +834,6 @@ class TFRoFormerModel(TFRoFormerPreTrainedModel):
 
         return outputs
 
-    def serving_output(self, output: TFBaseModelOutput) -> TFBaseModelOutput:
-        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)
-
 
 @add_start_docstrings("""RoFormer Model with a `language modeling` head on top.""", ROFORMER_START_DOCSTRING)
 class TFRoFormerForMaskedLM(TFRoFormerPreTrainedModel, TFMaskedLanguageModelingLoss):
@@ -911,12 +904,6 @@ class TFRoFormerForMaskedLM(TFRoFormerPreTrainedModel, TFMaskedLanguageModelingL
             attentions=outputs.attentions,
         )
 
-    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(
     """RoFormer Model with a `language modeling` head on top for CLM fine-tuning.""", ROFORMER_START_DOCSTRING
@@ -990,12 +977,6 @@ class TFRoFormerForCausalLM(TFRoFormerPreTrainedModel, TFCausalLanguageModelingL
             attentions=outputs.attentions,
         )
 
-    def serving_output(self, output: TFCausalLMOutput) -> TFCausalLMOutput:
-        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 TFCausalLMOutput(logits=output.logits, hidden_states=hs, attentions=attns)
-
 
 class TFRoFormerClassificationHead(tf.keras.layers.Layer):
     """Head for sentence-level classification tasks."""
@@ -1094,12 +1075,6 @@ class TFRoFormerForSequenceClassification(TFRoFormerPreTrainedModel, TFSequenceC
             attentions=outputs.attentions,
         )
 
-    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(
     """
@@ -1118,17 +1093,6 @@ class TFRoFormerForMultipleChoice(TFRoFormerPreTrainedModel, TFMultipleChoiceLos
             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.int32)}
-
     @unpack_inputs
     @add_start_docstrings_to_model_forward(
         ROFORMER_INPUTS_DOCSTRING.format("batch_size, num_choices, sequence_length")
@@ -1203,26 +1167,6 @@ class TFRoFormerForMultipleChoice(TFRoFormerPreTrainedModel, TFMultipleChoiceLos
             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"),
-            }
-        ]
-    )
-    def serving(self, inputs: Dict[str, tf.Tensor]) -> TFMultipleChoiceModelOutput:
-        output = self.call(input_ids=inputs)
-
-        return self.serving_output(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(
     """
@@ -1294,12 +1238,6 @@ class TFRoFormerForTokenClassification(TFRoFormerPreTrainedModel, TFTokenClassif
             attentions=outputs.attentions,
         )
 
-    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(
     """
@@ -1383,11 +1321,3 @@ class TFRoFormerForQuestionAnswering(TFRoFormerPreTrainedModel, TFQuestionAnswer
             hidden_states=outputs.hidden_states,
             attentions=outputs.attentions,
         )
-
-    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/whisper/modeling_tf_whisper.py

@@ -486,18 +486,13 @@ class TFWhisperPreTrainedModel(TFPreTrainedModel):
             "decoder_input_ids": tf.constant([[2, 3]], dtype=tf.int32),
         }
 
-    @tf.function(
-        input_signature=[
-            {
-                "input_features": tf.TensorSpec((None, None, None), tf.float32, name="input_features"),
+    @property
+    def input_signature(self):
+        return {
+            "input_features": tf.TensorSpec((None, self.config.num_mel_bins, None), tf.float32, name="input_features"),
             "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"),
         }
-        ]
-    )
-    def serving(self, inputs):
-        output = self.call(inputs)
-        return self.serving_output(output)
 
 
 WHISPER_START_DOCSTRING = r"""