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

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
814de8fa · Matt · GitHub · 3d7baef1 · 814de8fa · 814de8fa
Unverified Commit 814de8fa authored May 24, 2023 by Matt Committed by GitHub May 24, 2023
20 changed files
--- a/src/transformers/models/opt/modeling_tf_opt.py
+++ b/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:

--- a/src/transformers/models/pegasus/modeling_tf_pegasus.py
+++ b/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

--- a/src/transformers/models/regnet/modeling_tf_regnet.py
+++ b/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]:
+    def input_signature(self):
-        """
+        return {"pixel_values": tf.TensorSpec(shape=(None, self.config.num_channels, 224, 224), dtype=tf.float32)}
-        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)
 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)
--- a/src/transformers/models/rembert/modeling_tf_rembert.py
+++ b/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
-        )
--- a/src/transformers/models/resnet/modeling_tf_resnet.py
+++ b/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]:
+    def input_signature(self):
-        """
+        return {"pixel_values": tf.TensorSpec(shape=(None, self.config.num_channels, 224, 224), dtype=tf.float32)}
-        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)
 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)
--- a/src/transformers/models/roberta/modeling_tf_roberta.py
+++ b/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
-        )
--- a/src/transformers/models/roberta_prelayernorm/modeling_tf_roberta_prelayernorm.py
+++ b/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
-        )
--- a/src/transformers/models/roformer/modeling_tf_roformer.py
+++ b/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
-        )
--- a/src/transformers/models/segformer/modeling_tf_segformer.py
+++ b/src/transformers/models/segformer/modeling_tf_segformer.py
@@ -18,7 +18,7 @@
 from __future__ import annotations
 import math
-from typing import Dict, Optional, Tuple, Union
+from typing import Optional, Tuple, Union
 import tensorflow as tf
@@ -521,34 +521,8 @@ class TFSegformerPreTrainedModel(TFPreTrainedModel):
    main_input_name = "pixel_values"
    @property
-    def dummy_inputs(self) -> Dict[str, tf.Tensor]:
+    def input_signature(self):
-        """
+        return {"pixel_values": tf.TensorSpec(shape=(None, self.config.num_channels, 512, 512), dtype=tf.float32)}
-        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, 512, 512), 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)
 SEGFORMER_START_DOCSTRING = r"""
@@ -631,14 +605,6 @@ class TFSegformerModel(TFSegformerPreTrainedModel):
        )
        return outputs
-    def serving_output(self, output: TFBaseModelOutput) -> TFBaseModelOutput:
-        # hidden_states and attention not converted to Tensor with tf.convert_to_tensor as they are all of different dimensions
-        return TFBaseModelOutput(
-            last_hidden_state=output.last_hidden_state,
-            hidden_states=output.hidden_states,
-            attentions=output.attentions,
-        )
 @add_start_docstrings(
    """
@@ -702,12 +668,6 @@ class TFSegformerForImageClassification(TFSegformerPreTrainedModel, TFSequenceCl
            loss=loss, logits=logits, hidden_states=outputs.hidden_states, attentions=outputs.attentions
        )
-    def serving_output(self, output: TFSequenceClassifierOutput) -> TFSequenceClassifierOutput:
-        # hidden_states and attention 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, attentions=output.attentions
-        )
 class TFSegformerMLP(tf.keras.layers.Layer):
    """
@@ -892,9 +852,3 @@ class TFSegformerForSemanticSegmentation(TFSegformerPreTrainedModel):
            hidden_states=outputs.hidden_states if output_hidden_states else None,
            attentions=outputs.attentions,
        )
-    def serving_output(self, output: TFSemanticSegmenterOutput) -> TFSemanticSegmenterOutput:
-        # hidden_states and attention not converted to Tensor with tf.convert_to_tensor as they are all of different dimensions
-        return TFSemanticSegmenterOutput(
-            logits=output.logits, hidden_states=output.hidden_states, attentions=output.attentions
-        )
--- a/src/transformers/models/speech_to_text/modeling_tf_speech_to_text.py
+++ b/src/transformers/models/speech_to_text/modeling_tf_speech_to_text.py
@@ -18,7 +18,7 @@
 from __future__ import annotations
 import random
-from typing import Dict, Optional, Tuple, Union
+from typing import Optional, Tuple, Union
 import numpy as np
 import tensorflow as tf
@@ -563,26 +563,6 @@ class TFSpeech2TextPreTrainedModel(TFPreTrainedModel):
    base_model_prefix = "model"
    main_input_name = "input_features"
-    # Overwritten property due to different expected input shape and type
-    @property
-    def dummy_inputs(self) -> Dict[str, tf.Tensor]:
-        """
-        Dummy inputs to build the network.
-        Returns:
-            `Dict[str, tf.Tensor]`: The dummy inputs.
-        """
-        return {
-            self.main_input_name: tf.random.uniform(
-                [
-                    1,
-                    random.randint(1, self.config.max_source_positions),  # time
-                    self.config.input_feat_per_channel * self.config.input_channels,  # input channels
-                ]
-            ),
-            "decoder_input_ids": tf.constant([[2, 3]], dtype=tf.int32),
-        }
    def _get_feat_extract_output_lengths(self, input_lengths: tf.Tensor):
        """
        Computes the output length of the convolutional layers
@@ -592,20 +572,18 @@ class TFSpeech2TextPreTrainedModel(TFPreTrainedModel):
        return input_lengths
-    @tf.function(
+    @property
-        input_signature=[
+    def input_signature(self):
-            {
+        return {
-                "input_features": tf.TensorSpec((None, None, None), tf.float32, name="input_features"),
+            "input_features": tf.TensorSpec(
-                "attention_mask": tf.TensorSpec((None, None), tf.int32, name="attention_mask"),
+                (None, None, self.config.input_feat_per_channel * self.config.input_channels),
-                "decoder_input_ids": tf.TensorSpec((None, None), tf.int32, name="decoder_input_ids"),
+                tf.float32,
-                "decoder_attention_mask": tf.TensorSpec((None, None), tf.int32, name="decoder_attention_mask"),
+                name="input_features",
-            }
+            ),
-        ]
+            "attention_mask": tf.TensorSpec((None, None), tf.int32, name="attention_mask"),
-    )
+            "decoder_input_ids": tf.TensorSpec((None, None), tf.int32, name="decoder_input_ids"),
-    def serving(self, inputs):
+            "decoder_attention_mask": tf.TensorSpec((None, None), tf.int32, name="decoder_attention_mask"),
-        output = self.call(inputs)
+        }
-        return self.serving_output(output)
 SPEECH_TO_TEXT_START_DOCSTRING = r"""

--- a/src/transformers/models/swin/modeling_tf_swin.py
+++ b/src/transformers/models/swin/modeling_tf_swin.py
@@ -957,29 +957,6 @@ class TFSwinPreTrainedModel(TFPreTrainedModel):
        if isinstance(module, TFSwinEncoder):
            module.gradient_checkpointing = value
-    @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, self.config.image_size, self.config.image_size),
-            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)
 SWIN_START_DOCSTRING = r"""
    This model is a Tensorflow
@@ -1245,16 +1222,6 @@ class TFSwinModel(TFSwinPreTrainedModel):
        return swin_outputs
-    def serving_output(self, output: TFSwinModelOutput) -> TFSwinModelOutput:
-        # hidden_states and attentions not converted to Tensor with tf.convert_to_tensor as they are all of different dimensions
-        return TFSwinModelOutput(
-            last_hidden_state=output.last_hidden_state,
-            pooler_output=output.pooler_output,
-            hidden_states=output.hidden_states,
-            attentions=output.attentions,
-            reshaped_hidden_states=output.reshaped_hidden_states,
-        )
 class TFSwinPixelShuffle(tf.keras.layers.Layer):
    """TF layer implementation of torch.nn.PixelShuffle"""
@@ -1410,15 +1377,6 @@ class TFSwinForMaskedImageModeling(TFSwinPreTrainedModel):
            reshaped_hidden_states=outputs.reshaped_hidden_states,
        )
-    def serving_output(self, output: TFSwinMaskedImageModelingOutput) -> TFSwinMaskedImageModelingOutput:
-        # hidden_states and attentions not converted to Tensor with tf.convert_to_tensor as they are all of different dimensions
-        return TFSwinMaskedImageModelingOutput(
-            reconstruction=output.reconstruction,
-            hidden_states=output.hidden_states,
-            attentions=output.attentions,
-            reshaped_hidden_states=output.reshaped_hidden_states,
-        )
 @add_start_docstrings(
    """
@@ -1493,12 +1451,3 @@ class TFSwinForImageClassification(TFSwinPreTrainedModel, TFSequenceClassificati
            attentions=outputs.attentions,
            reshaped_hidden_states=outputs.reshaped_hidden_states,
        )
-    def serving_output(self, output: TFSwinImageClassifierOutput) -> TFSwinImageClassifierOutput:
-        # hidden_states and attentions not converted to Tensor with tf.convert_to_tensor as they are all of different dimensions
-        return TFSwinImageClassifierOutput(
-            logits=output.logits,
-            hidden_states=output.hidden_states,
-            attentions=output.attentions,
-            reshaped_hidden_states=output.reshaped_hidden_states,
-        )
--- a/src/transformers/models/t5/modeling_tf_t5.py
+++ b/src/transformers/models/t5/modeling_tf_t5.py
@@ -45,8 +45,6 @@ from ...modeling_tf_utils import (
 )
 from ...tf_utils import check_embeddings_within_bounds, shape_list, stable_softmax
 from ...utils import (
-    DUMMY_INPUTS,
-    DUMMY_MASK,
    ContextManagers,
    add_start_docstrings,
    add_start_docstrings_to_model_forward,
@@ -867,32 +865,6 @@ class TFT5PreTrainedModel(TFPreTrainedModel):
    # names with a '.' represents the authorized unexpected/missing layers when a TF model is loaded from a PT model
    _keys_to_ignore_on_load_unexpected = [r"decoder\Wblock[\W_0]+layer[\W_1]+EncDecAttention\Wrelative_attention_bias"]
-    @property
-    def dummy_inputs(self):
-        inputs = tf.constant(DUMMY_INPUTS, dtype=tf.int32)
-        input_mask = tf.constant(DUMMY_MASK, dtype=tf.int32)
-        dummy_inputs = {
-            "input_ids": inputs,
-            "decoder_input_ids": inputs,
-            "decoder_attention_mask": input_mask,
-        }
-        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"),
-            }
-        ]
-    )
-    def serving(self, inputs):
-        output = self.call(inputs)
-        return self.serving_output(output)
    def get_input_embeddings(self):
        return self.shared
@@ -1249,25 +1221,6 @@ class TFT5Model(TFT5PreTrainedModel):
            encoder_attentions=encoder_outputs.attentions,
        )
-    def serving_output(self, output):
-        pkv = tf.convert_to_tensor(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,
-            encoder_last_hidden_state=output.encoder_last_hidden_state,
-            cross_attentions=cross_attns,
-            encoder_hidden_states=enc_hs,
-            encoder_attentions=enc_attns,
-        )
 @add_start_docstrings("""T5 Model with a `language modeling` head on top.""", T5_START_DOCSTRING)
 class TFT5ForConditionalGeneration(TFT5PreTrainedModel, TFCausalLanguageModelingLoss):
@@ -1539,10 +1492,6 @@ class TFT5EncoderModel(TFT5PreTrainedModel):
        encoder_config.use_cache = False
        self.encoder = TFT5MainLayer(encoder_config, self.shared, name="encoder")
-    @property
-    def dummy_inputs(self):
-        return {"input_ids": tf.constant(DUMMY_INPUTS, dtype=tf.int32)}
    def get_encoder(self):
        return self.encoder
@@ -1600,23 +1549,3 @@ class TFT5EncoderModel(TFT5PreTrainedModel):
            hidden_states=encoder_outputs.hidden_states,
            attentions=encoder_outputs.attentions,
        )
-    @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)
-    # 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)
--- a/src/transformers/models/tapas/modeling_tf_tapas.py
+++ b/src/transformers/models/tapas/modeling_tf_tapas.py
@@ -862,18 +862,13 @@ class TFTapasPreTrainedModel(TFPreTrainedModel):
    config_class = TapasConfig
    base_model_prefix = "tapas"
-    @tf.function(
+    @property
-        input_signature=[
+    def input_signature(self):
-            {
+        return {
-                "input_ids": tf.TensorSpec((None, None), tf.int32, name="input_ids"),
+            "input_ids": tf.TensorSpec((None, None), tf.int32, name="input_ids"),
-                "attention_mask": tf.TensorSpec((None, None), tf.float32, name="attention_mask"),
+            "attention_mask": tf.TensorSpec((None, None), tf.float32, name="attention_mask"),
-                "token_type_ids": tf.TensorSpec((None, None, None), tf.int32, name="token_type_ids"),
+            "token_type_ids": tf.TensorSpec((None, None, 7), tf.int32, name="token_type_ids"),
-            }
+        }
-        ]
-    )
-    def serving(self, inputs):
-        output = self.call(inputs)
-        return self.serving_output(output)
 TAPAS_START_DOCSTRING = r"""
@@ -1038,17 +1033,6 @@ class TFTapasModel(TFTapasPreTrainedModel):
        return outputs
-    def serving_output(self, output: TFBaseModelOutputWithPooling) -> TFBaseModelOutputWithPooling:
-        hidden_states = tf.convert_to_tensor(output.hidden_states) if self.config.output_hidden_states else None
-        attentions = 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=hidden_states,
-            attentions=attentions,
-        )
 @add_start_docstrings("""Tapas Model with a `language modeling` head on top.""", TAPAS_START_DOCSTRING)
 class TFTapasForMaskedLM(TFTapasPreTrainedModel, TFMaskedLanguageModelingLoss):
@@ -1145,12 +1129,6 @@ class TFTapasForMaskedLM(TFTapasPreTrainedModel, TFMaskedLanguageModelingLoss):
            attentions=outputs.attentions,
        )
-    def serving_output(self, output: TFMaskedLMOutput) -> TFMaskedLMOutput:
-        hidden_states = tf.convert_to_tensor(output.hidden_states) if self.config.output_hidden_states else None
-        attentions = tf.convert_to_tensor(output.attentions) if self.config.output_attentions else None
-        return TFMaskedLMOutput(logits=output.logits, hidden_states=hidden_states, attentions=attentions)
 class TFTapasComputeTokenLogits(tf.keras.layers.Layer):
    def __init__(self, config: TapasConfig, **kwargs):
@@ -1574,17 +1552,6 @@ class TFTapasForQuestionAnswering(TFTapasPreTrainedModel):
            attentions=outputs.attentions,
        )
-    def serving_output(self, output: TFTableQuestionAnsweringOutput) -> TFTableQuestionAnsweringOutput:
-        hidden_states = tf.convert_to_tensor(output.hidden_states) if self.config.output_hidden_states else None
-        attentions = tf.convert_to_tensor(output.attentions) if self.config.output_attentions else None
-        return TFTableQuestionAnsweringOutput(
-            logits=output.logits,
-            logits_aggregation=output.logits_aggregation,
-            hidden_states=hidden_states,
-            attentions=attentions,
-        )
 @add_start_docstrings(
    """
@@ -1687,12 +1654,6 @@ class TFTapasForSequenceClassification(TFTapasPreTrainedModel, TFSequenceClassif
            attentions=outputs.attentions,
        )
-    def serving_output(self, output: TFSequenceClassifierOutput) -> TFSequenceClassifierOutput:
-        hidden_states = tf.convert_to_tensor(output.hidden_states) if self.config.output_hidden_states else None
-        attentions = tf.convert_to_tensor(output.attentions) if self.config.output_attentions else None
-        return TFSequenceClassifierOutput(logits=output.logits, hidden_states=hidden_states, attentions=attentions)
 """ TAPAS utilities."""

--- a/src/transformers/models/transfo_xl/modeling_tf_transfo_xl.py
+++ b/src/transformers/models/transfo_xl/modeling_tf_transfo_xl.py
@@ -684,18 +684,6 @@ class TFTransfoXLPreTrainedModel(TFPreTrainedModel):
    config_class = TransfoXLConfig
    base_model_prefix = "transformer"
-    @tf.function(
-        input_signature=[
-            {
-                "input_ids": tf.TensorSpec((None, None), tf.int32, name="input_ids"),
-            }
-        ]
-    )
-    def serving(self, inputs):
-        output = self.call(inputs)
-        return self.serving_output(output)
 @dataclass
 class TFTransfoXLModelOutput(ModelOutput):
@@ -916,17 +904,6 @@ class TFTransfoXLModel(TFTransfoXLPreTrainedModel):
        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 TFTransfoXLModelOutput(
-            last_hidden_state=output.last_hidden_state,
-            mems=tf.convert_to_tensor(output.mems),
-            hidden_states=hs,
-            attentions=attns,
-        )
 @add_start_docstrings(
    """
@@ -1015,17 +992,6 @@ class TFTransfoXLLMHeadModel(TFTransfoXLPreTrainedModel):
            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 TFTransfoXLLMHeadModelOutput(
-            prediction_scores=output.prediction_scores,
-            mems=tf.convert_to_tensor(output.mems),
-            hidden_states=hs,
-            attentions=attns,
-        )
    def prepare_inputs_for_generation(self, input_ids, past_key_values=None, **model_kwargs):
        inputs = {}
@@ -1157,11 +1123,3 @@ class TFTransfoXLForSequenceClassification(TFTransfoXLPreTrainedModel, TFSequenc
            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 TFTransfoXLSequenceClassifierOutputWithPast(
-            logits=output.logits, mems=tf.convert_to_tensor(output.mems), hidden_states=hs, attentions=attns
-        )
--- a/src/transformers/models/vision_encoder_decoder/modeling_tf_vision_encoder_decoder.py
+++ b/src/transformers/models/vision_encoder_decoder/modeling_tf_vision_encoder_decoder.py
@@ -29,7 +29,6 @@ from ...modeling_tf_outputs import TFBaseModelOutput, TFSeq2SeqLMOutput
 from ...modeling_tf_utils import TFCausalLanguageModelingLoss, TFPreTrainedModel, get_initializer, unpack_inputs
 from ...tf_utils import shape_list
 from ...utils import (
-    DUMMY_INPUTS,
    ModelOutput,
    add_start_docstrings,
    add_start_docstrings_to_model_forward,
@@ -254,29 +253,26 @@ class TFVisionEncoderDecoderModel(TFPreTrainedModel, TFCausalLanguageModelingLos
            )
    @property
-    def dummy_inputs(self):
+    def input_signature(self):
-        """
+        vision_config = self.config.encoder
-        Dummy inputs to build the network.
+        if hasattr(vision_config, "vision_config"):
+            vision_config = vision_config.vision_config
-        Returns:
+        if hasattr(vision_config, "image_size"):
-            `Dict[str, tf.Tensor]`: The dummy inputs.
+            image_size = vision_config.image_size
-        """
+        else:
-        decoder_input_ids = tf.constant(DUMMY_INPUTS, dtype=tf.int32)
+            image_size = vision_config.input_size
-        batch_size, seq_len = decoder_input_ids.shape
+        return {
+            "pixel_values": tf.TensorSpec(
-        VISION_DUMMY_INPUTS = tf.random.uniform(
+                shape=(
-            shape=(
+                    None,
-                batch_size,
+                    vision_config.num_channels,
-                self.config.encoder.num_channels,
+                    image_size,
-                self.config.encoder.image_size,
+                    image_size,
-                self.config.encoder.image_size,
+                ),
+                dtype=tf.float32,
            ),
-            dtype=tf.float32,
+            "decoder_input_ids": tf.TensorSpec(shape=(None, None), dtype=tf.int32, name="decoder_input_ids"),
-        )
+        }
-        pixel_values = tf.constant(VISION_DUMMY_INPUTS)
-        # Add `decoder_input_ids` because `self.decoder` requires it.
-        dummy = {"pixel_values": pixel_values, "decoder_input_ids": decoder_input_ids}
-        return dummy
    def get_encoder(self):
        return self.encoder

--- a/src/transformers/models/vit/modeling_tf_vit.py
+++ b/src/transformers/models/vit/modeling_tf_vit.py
@@ -19,7 +19,7 @@ from __future__ import annotations
 import collections.abc
 import math
-from typing import Dict, Optional, Tuple, Union
+from typing import Optional, Tuple, Union
 import numpy as np
 import tensorflow as tf
@@ -550,38 +550,6 @@ class TFViTPreTrainedModel(TFPreTrainedModel):
    base_model_prefix = "vit"
    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, self.config.image_size, self.config.image_size), 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)
 VIT_START_DOCSTRING = r"""
@@ -697,17 +665,6 @@ class TFViTModel(TFViTPreTrainedModel):
        return outputs
-    def serving_output(self, output: TFBaseModelOutputWithPooling) -> TFBaseModelOutputWithPooling:
-        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 TFViTPooler(tf.keras.layers.Layer):
    def __init__(self, config: ViTConfig, **kwargs):
@@ -807,9 +764,3 @@ class TFViTForImageClassification(TFViTPreTrainedModel, TFSequenceClassification
            hidden_states=outputs.hidden_states,
            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)
--- a/src/transformers/models/vit_mae/modeling_tf_vit_mae.py
+++ b/src/transformers/models/vit_mae/modeling_tf_vit_mae.py
@@ -21,7 +21,7 @@ import collections.abc
 import math
 from copy import deepcopy
 from dataclasses import dataclass
-from typing import Dict, Optional, Tuple, Union
+from typing import Optional, Tuple, Union
 import numpy as np
 import tensorflow as tf
@@ -698,36 +698,6 @@ class TFViTMAEPreTrainedModel(TFPreTrainedModel):
    base_model_prefix = "vit"
    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, self.config.image_size, self.config.image_size),
-            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)
 VIT_MAE_START_DOCSTRING = r"""
    This model inherits from [`TFPreTrainedModel`]. Check the superclass documentation for the generic methods the
@@ -859,18 +829,6 @@ class TFViTMAEModel(TFViTMAEPreTrainedModel):
        return outputs
-    def serving_output(self, output: TFViTMAEModelOutput) -> TFViTMAEModelOutput:
-        hidden_states = tf.convert_to_tensor(output.hidden_states) if self.config.output_hidden_states else None
-        attentions = tf.convert_to_tensor(output.attentions) if self.config.output_attentions else None
-        return TFViTMAEModelOutput(
-            last_hidden_state=output.last_hidden_state,
-            mask=output.mask,
-            ids_restore=output.ids_restore,
-            hidden_states=hidden_states,
-            attentions=attentions,
-        )
 class TFViTMAEDecoder(tf.keras.layers.Layer):
    def __init__(self, config, num_patches, **kwargs):
@@ -1173,15 +1131,3 @@ class TFViTMAEForPreTraining(TFViTMAEPreTrainedModel):
            hidden_states=outputs.hidden_states,
            attentions=outputs.attentions,
        )
-    def serving_output(self, output: TFViTMAEForPreTrainingOutput) -> TFViTMAEForPreTrainingOutput:
-        hidden_states = tf.convert_to_tensor(output.hidden_states) if self.config.output_hidden_states else None
-        attentions = tf.convert_to_tensor(output.attentions) if self.config.output_attentions else None
-        return TFViTMAEForPreTrainingOutput(
-            logits=output.logits,
-            mask=output.mask,
-            ids_restore=output.ids_restore,
-            hidden_states=hidden_states,
-            attentions=attentions,
-        )
--- a/src/transformers/models/wav2vec2/modeling_tf_wav2vec2.py
+++ b/src/transformers/models/wav2vec2/modeling_tf_wav2vec2.py
@@ -19,7 +19,7 @@ from __future__ import annotations
 import warnings
 from dataclasses import dataclass
-from typing import Any, Dict, Optional, Tuple, Union
+from typing import Any, Optional, Tuple, Union
 import numpy as np
 import tensorflow as tf
@@ -1185,14 +1185,18 @@ class TFWav2Vec2PreTrainedModel(TFPreTrainedModel):
    main_input_name = "input_values"
    @property
-    def dummy_inputs(self) -> Dict[str, tf.Tensor]:
+    def input_signature(self):
-        pad_token = 0.0
+        return {
-        input_values = tf.convert_to_tensor(np.random.rand(1, 16000), tf.float32)
+            "input_values": tf.TensorSpec((None, None), tf.float32, name="input_values"),
-        dummy_inputs = {
+            "attention_mask": tf.TensorSpec((None, None), tf.float32, name="attention_mask"),
-            "input_values": input_values,
+        }
-            "attention_mask": tf.cast(tf.not_equal(input_values, pad_token), tf.float32),
+    @property
+    def dummy_inputs(self):
+        return {
+            "input_values": tf.random.uniform(shape=(1, 16000), dtype=tf.float32),
+            "attention_mask": tf.ones(shape=(1, 16000), dtype=tf.float32),
        }
-        return dummy_inputs
    def __init__(self, config, *inputs, **kwargs):
        super().__init__(config, *inputs, **kwargs)
@@ -1201,20 +1205,6 @@ class TFWav2Vec2PreTrainedModel(TFPreTrainedModel):
            "to train/fine-tine this model, you need a GPU or a TPU"
        )
-    @tf.function(
-        input_signature=[
-            {
-                "input_values": tf.TensorSpec((None, None), tf.float32, name="input_values"),
-                "attention_mask": tf.TensorSpec((None, None), tf.int32, name="attention_mask"),
-                "token_type_ids": tf.TensorSpec((None, None), tf.int32, name="token_type_ids"),
-            }
-        ]
-    )
-    def serving(self, inputs):
-        output = self.call(input_values=inputs, training=False)
-        return self.serving_output(output)
    def _get_feat_extract_output_lengths(self, input_lengths, add_adapter=None):
        """
        Computes the output length of the convolutional layers
@@ -1427,17 +1417,6 @@ class TFWav2Vec2Model(TFWav2Vec2PreTrainedModel):
        return outputs
-    def serving_output(self, output):
-        hidden_states = tf.convert_to_tensor(output.hidden_states) if self.config.output_hidden_states else None
-        attentions = tf.convert_to_tensor(output.attentions) if self.config.output_attentions else None
-        return TFWav2Vec2BaseModelOutput(
-            last_hidden_state=output.last_hidden_state,
-            extract_features=output.extract_features,
-            hidden_states=hidden_states,
-            attentions=attentions,
-        )
 @add_start_docstrings(
    """TFWav2Vec2 Model with a `language modeling` head on top for Connectionist Temporal Classification (CTC).""",
@@ -1591,11 +1570,6 @@ class TFWav2Vec2ForCTC(TFWav2Vec2PreTrainedModel):
            attentions=outputs.attentions,
        )
-    def serving_output(self, output: TFCausalLMOutput) -> TFCausalLMOutput:
-        hidden_states = tf.convert_to_tensor(output.hidden_states) if self.config.output_hidden_states else None
-        attentions = tf.convert_to_tensor(output.attentions) if self.config.output_attentions else None
-        return TFCausalLMOutput(logits=output.logits, hidden_states=hidden_states, attentions=attentions)
 class TFWav2Vec2ForSequenceClassification(TFWav2Vec2PreTrainedModel):
    def __init__(self, config):
@@ -1693,27 +1667,3 @@ class TFWav2Vec2ForSequenceClassification(TFWav2Vec2PreTrainedModel):
            hidden_states=outputs.hidden_states,
            attentions=outputs.attentions,
        )
-    def serving_output(self, output):
-        hidden_states = tf.convert_to_tensor(output.hidden_states) if self.config.output_hidden_states else None
-        attentions = tf.convert_to_tensor(output.attentions) if self.config.output_attentions else None
-        return TFSequenceClassifierOutput(
-            logits=output.logits,
-            hidden_states=hidden_states,
-            attentions=attentions,
-        )
-    @tf.function(
-        input_signature=[
-            {
-                "input_values": tf.TensorSpec((None, None), tf.float32, name="input_values"),
-                "attention_mask": tf.TensorSpec((None, None), tf.int32, name="attention_mask"),
-                "token_type_ids": tf.TensorSpec((None, None), tf.int32, name="token_type_ids"),
-            }
-        ]
-    )
-    def serving(self, inputs):
-        output = self.call(input_values=inputs)
-        return self.serving_output(output)
--- a/src/transformers/models/whisper/modeling_tf_whisper.py
+++ b/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(
+    @property
-        input_signature=[
+    def input_signature(self):
-            {
+        return {
-                "input_features": tf.TensorSpec((None, None, None), tf.float32, name="input_features"),
+            "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_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"),
+            "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"""

--- a/src/transformers/models/xglm/modeling_tf_xglm.py
+++ b/src/transformers/models/xglm/modeling_tf_xglm.py
@@ -28,7 +28,6 @@ from ...activations_tf import get_tf_activation
 # Public API
 from ...file_utils import (
-    DUMMY_INPUTS,
    add_code_sample_docstrings,
    add_start_docstrings,
    add_start_docstrings_to_model_forward,
@@ -620,29 +619,6 @@ class TFXGLMPreTrainedModel(TFPreTrainedModel):
    config_class = XGLMConfig
    base_model_prefix = "model"
-    @property
-    def dummy_inputs(self):
-        pad_token = 1
-        input_ids = tf.cast(tf.convert_to_tensor(DUMMY_INPUTS), tf.int32)
-        dummy_inputs = {
-            "input_ids": input_ids,
-            "attention_mask": tf.cast(input_ids != pad_token, tf.int32),
-        }
-        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)
 XGLM_START_DOCSTRING = r"""
    This model inherits from [`TFPreTrainedModel`]. Check the superclass documentation for the generic methods the
@@ -821,24 +797,6 @@ class TFXGLMModel(TFXGLMPreTrainedModel):
        return outputs
-    def serving_output(self, output):
-        pkv = tf.convert_to_tensor(output.past_key_values) if self.config.use_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 self.config.output_attentions and self.config.add_cross_attention
-            else None
-        )
-        return TFBaseModelOutputWithPastAndCrossAttentions(
-            last_hidden_state=output.hidden_states,
-            past_key_values=pkv,
-            hidden_states=hs,
-            attentions=attns,
-            cross_attentions=cross_attns,
-        )
 @add_start_docstrings(
    """
@@ -971,22 +929,3 @@ class TFXGLMForCausalLM(TFXGLMPreTrainedModel, TFCausalLanguageModelingLoss):
            attentions=outputs.attentions,
            cross_attentions=outputs.cross_attentions,
        )
-    def serving_output(self, output):
-        pkv = tf.convert_to_tensor(output.past_key_values) if self.config.use_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 self.config.output_attentions and self.config.add_cross_attention
-            else None
-        )
-        return TFCausalLMOutputWithCrossAttentions(
-            loss=output.loss,
-            logits=output.logits,
-            past_key_values=pkv,
-            hidden_states=hs,
-            attentions=attns,
-            cross_attentions=cross_attns,
-        )