TF: Finalize `unpack_inputs`-related changes (#16499)

* Add unpack_inputs to remaining models

* removed kwargs to `call()` in TF models

* fix TF T5 tests

src/transformers/models/layoutlm/modeling_tf_layoutlm.py

@@ -706,7 +706,6 @@ class TFLayoutLMMainLayer(tf.keras.layers.Layer):
         output_hidden_states: Optional[bool] = None,
         return_dict: Optional[bool] = None,
         training: bool = False,
-        **kwargs,
     ) -> Union[TFBaseModelOutputWithPoolingAndCrossAttentions, Tuple[tf.Tensor]]:
 
         if input_ids is not None and inputs_embeds is not None:
@@ -928,7 +927,6 @@ class TFLayoutLMModel(TFLayoutLMPreTrainedModel):
         output_hidden_states: Optional[bool] = None,
         return_dict: Optional[bool] = None,
         training: Optional[bool] = False,
-        **kwargs,
     ) -> Union[TFBaseModelOutputWithPoolingAndCrossAttentions, Tuple[tf.Tensor]]:
         r"""
         Returns:
@@ -1048,7 +1046,6 @@ class TFLayoutLMForMaskedLM(TFLayoutLMPreTrainedModel, TFMaskedLanguageModelingL
         return_dict: Optional[bool] = None,
         labels: Optional[Union[np.ndarray, tf.Tensor]] = None,
         training: Optional[bool] = False,
-        **kwargs,
     ) -> Union[TFMaskedLMOutput, Tuple[tf.Tensor]]:
         r"""
         labels (`tf.Tensor` or `np.ndarray` of shape `(batch_size, sequence_length)`, *optional*):
@@ -1172,7 +1169,6 @@ class TFLayoutLMForSequenceClassification(TFLayoutLMPreTrainedModel, TFSequenceC
         return_dict: Optional[bool] = None,
         labels: Optional[Union[np.ndarray, tf.Tensor]] = None,
         training: Optional[bool] = False,
-        **kwargs,
     ) -> Union[TFSequenceClassifierOutput, Tuple[tf.Tensor]]:
         r"""
         labels (`tf.Tensor` or `np.ndarray` of shape `(batch_size,)`, *optional*):
@@ -1303,7 +1299,6 @@ class TFLayoutLMForTokenClassification(TFLayoutLMPreTrainedModel, TFTokenClassif
         return_dict: Optional[bool] = None,
         labels: Optional[Union[np.ndarray, tf.Tensor]] = None,
         training: Optional[bool] = False,
-        **kwargs,
     ) -> Union[TFTokenClassifierOutput, Tuple[tf.Tensor]]:
         r"""
         labels (`tf.Tensor` or `np.ndarray` of shape `(batch_size, sequence_length)`, *optional*):

src/transformers/models/led/modeling_tf_led.py

@@ -1666,7 +1666,6 @@ class TFLEDEncoder(tf.keras.layers.Layer):
         output_hidden_states=None,
         return_dict=None,
         training=False,
-        **kwargs,
     ):
         """
         Args:
@@ -1911,7 +1910,6 @@ class TFLEDDecoder(tf.keras.layers.Layer):
         output_hidden_states=None,
         return_dict=None,
         training=False,
-        **kwargs,
     ):
         r"""
         Args:
@@ -2333,7 +2331,6 @@ class TFLEDForConditionalGeneration(TFLEDPreTrainedModel):
         return_dict=None,
         labels=None,
         training=False,
-        **kwargs,
     ):
         """
         Returns:
@@ -2429,7 +2426,7 @@ class TFLEDForConditionalGeneration(TFLEDPreTrainedModel):
         decoder_head_mask=None,
         use_cache=None,
         encoder_outputs=None,
-        **kwargs,
+        **kwargs
     ):
         # cut decoder_input_ids if past is used
         if past is not None:

src/transformers/models/longformer/modeling_tf_longformer.py

@@ -1676,7 +1676,6 @@ class TFLongformerMainLayer(tf.keras.layers.Layer):
         output_hidden_states=None,
         return_dict=None,
         training=False,
-        **kwargs,
     ):
 
         if input_ids is not None and inputs_embeds is not None:
@@ -2023,7 +2022,6 @@ class TFLongformerModel(TFLongformerPreTrainedModel):
         output_hidden_states: Optional[bool] = None,
         return_dict: Optional[bool] = None,
         training: Optional[bool] = False,
-        **kwargs,
     ) -> Union[TFLongformerBaseModelOutputWithPooling, Tuple[tf.Tensor]]:
 
         outputs = self.longformer(
@@ -2100,7 +2098,6 @@ class TFLongformerForMaskedLM(TFLongformerPreTrainedModel, TFMaskedLanguageModel
         return_dict: Optional[bool] = None,
         labels: Optional[Union[np.ndarray, tf.Tensor]] = None,
         training: Optional[bool] = False,
-        **kwargs,
     ) -> Union[TFLongformerMaskedLMOutput, Tuple[tf.Tensor]]:
         r"""
         labels (`tf.Tensor` of shape `(batch_size, sequence_length)`, *optional*):
@@ -2194,7 +2191,6 @@ class TFLongformerForQuestionAnswering(TFLongformerPreTrainedModel, TFQuestionAn
         start_positions: Optional[Union[np.ndarray, tf.Tensor]] = None,
         end_positions: Optional[Union[np.ndarray, tf.Tensor]] = None,
         training: Optional[bool] = False,
-        **kwargs,
     ) -> Union[TFLongformerQuestionAnsweringModelOutput, Tuple[tf.Tensor]]:
         r"""
         start_positions (`tf.Tensor` of shape `(batch_size,)`, *optional*):
@@ -2340,7 +2336,6 @@ class TFLongformerForSequenceClassification(TFLongformerPreTrainedModel, TFSeque
         return_dict: Optional[bool] = None,
         labels: Optional[Union[np.ndarray, tf.Tensor]] = None,
         training: Optional[bool] = False,
-        **kwargs,
     ) -> Union[TFLongformerSequenceClassifierOutput, Tuple[tf.Tensor]]:
 
         if global_attention_mask is None and input_ids is not None:
@@ -2450,7 +2445,6 @@ class TFLongformerForMultipleChoice(TFLongformerPreTrainedModel, TFMultipleChoic
         return_dict: Optional[bool] = None,
         labels: Optional[Union[np.ndarray, tf.Tensor]] = None,
         training: Optional[bool] = False,
-        **kwargs,
     ) -> Union[TFLongformerMultipleChoiceModelOutput, Tuple[tf.Tensor]]:
         r"""
         labels (`tf.Tensor` of shape `(batch_size,)`, *optional*):
@@ -2580,7 +2574,6 @@ class TFLongformerForTokenClassification(TFLongformerPreTrainedModel, TFTokenCla
         return_dict: Optional[bool] = None,
         labels: Optional[Union[np.array, tf.Tensor]] = None,
         training: Optional[bool] = False,
-        **kwargs,
     ) -> Union[TFLongformerTokenClassifierOutput, Tuple[tf.Tensor]]:
         r"""
         labels (`tf.Tensor` of shape `(batch_size, sequence_length)`, *optional*):

src/transformers/models/lxmert/modeling_tf_lxmert.py

@@ -685,7 +685,6 @@ class TFLxmertMainLayer(tf.keras.layers.Layer):
         output_hidden_states=None,
         return_dict=None,
         training=False,
-        **kwargs,
     ):
 
         if input_ids is not None and inputs_embeds is not None:
@@ -946,7 +945,6 @@ class TFLxmertModel(TFLxmertPreTrainedModel):
         output_hidden_states=None,
         return_dict=None,
         training=False,
-        **kwargs,
     ):
         outputs = self.lxmert(
             input_ids,
@@ -1282,7 +1280,6 @@ class TFLxmertForPreTraining(TFLxmertPreTrainedModel):
         output_hidden_states=None,
         return_dict=None,
         training=False,
-        **kwargs,
     ):
         r"""
         masked_lm_labels (`tf.Tensor` of shape `(batch_size, sequence_length)`, *optional*):

src/transformers/models/marian/modeling_tf_marian.py

@@ -707,7 +707,6 @@ class TFMarianEncoder(tf.keras.layers.Layer):
         output_hidden_states=None,
         return_dict=None,
         training=False,
-        **kwargs,
     ):
         """
         Args:
@@ -866,7 +865,6 @@ class TFMarianDecoder(tf.keras.layers.Layer):
         output_hidden_states=None,
         return_dict=None,
         training=False,
-        **kwargs,
     ):
         r"""
         Args:
@@ -1296,7 +1294,6 @@ class TFMarianMTModel(TFMarianPreTrainedModel, TFCausalLanguageModelingLoss):
         return_dict=None,
         labels=None,
         training=False,
-        **kwargs,
     ):
         r"""
         labels (`tf.tensor` of shape `(batch_size, sequence_length)`, *optional*):

src/transformers/models/mbart/modeling_tf_mbart.py

@@ -684,7 +684,6 @@ class TFMBartEncoder(tf.keras.layers.Layer):
         output_hidden_states: Optional[bool] = None,
         return_dict: Optional[bool] = None,
         training: Optional[bool] = False,
-        **kwargs,
     ) -> Union[TFBaseModelOutput, Tuple[tf.Tensor]]:
         """
         Args:
@@ -848,7 +847,6 @@ class TFMBartDecoder(tf.keras.layers.Layer):
         output_hidden_states: Optional[bool] = None,
         return_dict: Optional[bool] = None,
         training: Optional[bool] = False,
-        **kwargs,
     ) -> Union[
         TFBaseModelOutputWithPastAndCrossAttentions, Tuple[tf.Tensor, tf.Tensor, tf.Tensor, tf.Tensor, tf.Tensor]
     ]:
@@ -1278,7 +1276,7 @@ class TFMBartForConditionalGeneration(TFMBartPreTrainedModel, TFCausalLanguageMo
         decoder_head_mask: Optional[tf.Tensor] = None,
         cross_attn_head_mask: Optional[tf.Tensor] = None,
         encoder_outputs: Optional[TFBaseModelOutput] = None,
-        past_key_values: [Tuple[Tuple[tf.Tensor]]] = None,
+        past_key_values: Tuple[Tuple[tf.Tensor]] = None,
         inputs_embeds: Optional[tf.Tensor] = None,
         decoder_inputs_embeds: Optional[tf.Tensor] = None,
         use_cache: Optional[bool] = None,
@@ -1287,7 +1285,6 @@ class TFMBartForConditionalGeneration(TFMBartPreTrainedModel, TFCausalLanguageMo
         return_dict: Optional[bool] = None,
         labels: Optional[tf.Tensor] = None,
         training: Optional[bool] = False,
-        **kwargs,
     ) -> Union[TFSeq2SeqLMOutput, Tuple[tf.Tensor]]:
         """
         labels (`tf.Tensor` of shape `(batch_size, sequence_length)`, *optional*):

src/transformers/models/mobilebert/modeling_tf_mobilebert.py

@@ -692,7 +692,6 @@ class TFMobileBertMainLayer(tf.keras.layers.Layer):
         output_hidden_states=None,
         return_dict=None,
         training=False,
-        **kwargs,
     ):
         if input_ids is not None and inputs_embeds is not None:
             raise ValueError("You cannot specify both input_ids and inputs_embeds at the same time")
@@ -928,7 +927,6 @@ class TFMobileBertModel(TFMobileBertPreTrainedModel):
         output_hidden_states=None,
         return_dict=None,
         training=False,
-        **kwargs,
     ):
         outputs = self.mobilebert(
             input_ids=input_ids,
@@ -993,7 +991,6 @@ class TFMobileBertForPreTraining(TFMobileBertPreTrainedModel):
         output_hidden_states=None,
         return_dict=None,
         training=False,
-        **kwargs,
     ):
         r"""
         Return:
@@ -1092,7 +1089,6 @@ class TFMobileBertForMaskedLM(TFMobileBertPreTrainedModel, TFMaskedLanguageModel
         return_dict=None,
         labels=None,
         training=False,
-        **kwargs,
     ):
         r"""
         labels (`tf.Tensor` of shape `(batch_size, sequence_length)`, *optional*):
@@ -1176,7 +1172,6 @@ class TFMobileBertForNextSentencePrediction(TFMobileBertPreTrainedModel, TFNextS
         return_dict=None,
         next_sentence_label=None,
         training=False,
-        **kwargs,
     ):
         r"""
         Return:
@@ -1287,7 +1282,6 @@ class TFMobileBertForSequenceClassification(TFMobileBertPreTrainedModel, TFSeque
         return_dict=None,
         labels=None,
         training=False,
-        **kwargs,
     ):
         r"""
         labels (`tf.Tensor` of shape `(batch_size,)`, *optional*):
@@ -1381,7 +1375,6 @@ class TFMobileBertForQuestionAnswering(TFMobileBertPreTrainedModel, TFQuestionAn
         start_positions=None,
         end_positions=None,
         training=False,
-        **kwargs,
     ):
         r"""
         start_positions (`tf.Tensor` of shape `(batch_size,)`, *optional*):
@@ -1498,7 +1491,6 @@ class TFMobileBertForMultipleChoice(TFMobileBertPreTrainedModel, TFMultipleChoic
         return_dict=None,
         labels=None,
         training=False,
-        **kwargs,
     ):
         r"""
         labels (`tf.Tensor` of shape `(batch_size,)`, *optional*):
@@ -1626,7 +1618,6 @@ class TFMobileBertForTokenClassification(TFMobileBertPreTrainedModel, TFTokenCla
         return_dict=None,
         labels=None,
         training=False,
-        **kwargs,
     ):
         r"""
         labels (`tf.Tensor` of shape `(batch_size, sequence_length)`, *optional*):

src/transformers/models/mpnet/modeling_tf_mpnet.py

@@ -497,7 +497,6 @@ class TFMPNetMainLayer(tf.keras.layers.Layer):
         output_hidden_states=None,
         return_dict=None,
         training=False,
-        **kwargs,
     ):
 
         if input_ids is not None and inputs_embeds is not None:
@@ -686,7 +685,6 @@ class TFMPNetModel(TFMPNetPreTrainedModel):
         output_hidden_states=None,
         return_dict=None,
         training=False,
-        **kwargs,
     ):
         outputs = self.mpnet(
             input_ids=input_ids,
@@ -803,7 +801,6 @@ class TFMPNetForMaskedLM(TFMPNetPreTrainedModel, TFMaskedLanguageModelingLoss):
         return_dict=None,
         labels=None,
         training=False,
-        **kwargs,
     ):
         r"""
         labels (`tf.Tensor` of shape `(batch_size, sequence_length)`, *optional*):
@@ -909,7 +906,6 @@ class TFMPNetForSequenceClassification(TFMPNetPreTrainedModel, TFSequenceClassif
         return_dict=None,
         labels=None,
         training=False,
-        **kwargs,
     ):
         r"""
         labels (`tf.Tensor` of shape `(batch_size,)`, *optional*):
@@ -1000,7 +996,6 @@ class TFMPNetForMultipleChoice(TFMPNetPreTrainedModel, TFMultipleChoiceLoss):
         return_dict=None,
         labels=None,
         training=False,
-        **kwargs,
     ):
         r"""
         labels (`tf.Tensor` of shape `(batch_size,)`, *optional*):
@@ -1112,7 +1107,6 @@ class TFMPNetForTokenClassification(TFMPNetPreTrainedModel, TFTokenClassificatio
         return_dict=None,
         labels=None,
         training=False,
-        **kwargs,
     ):
         r"""
         labels (`tf.Tensor` of shape `(batch_size, sequence_length)`, *optional*):

src/transformers/models/openai/modeling_tf_openai.py

@@ -249,7 +249,6 @@ class TFOpenAIGPTMainLayer(tf.keras.layers.Layer):
         output_hidden_states=None,
         return_dict=None,
         training=False,
-        **kwargs,
     ):
 
         if input_ids is not None and inputs_embeds is not None:
@@ -522,7 +521,6 @@ class TFOpenAIGPTModel(TFOpenAIGPTPreTrainedModel):
         output_hidden_states: Optional[bool] = None,
         return_dict: Optional[bool] = None,
         training: Optional[bool] = False,
-        **kwargs,
     ) -> Union[Tuple, TFBaseModelOutput]:
 
         outputs = self.transformer(
@@ -586,7 +584,6 @@ class TFOpenAIGPTLMHeadModel(TFOpenAIGPTPreTrainedModel, TFCausalLanguageModelin
         return_dict: Optional[bool] = None,
         labels: Optional[Union[np.ndarray, tf.Tensor]] = None,
         training: Optional[bool] = False,
-        **kwargs,
     ) -> Union[Tuple, TFCausalLMOutput]:
         r"""
         labels (`tf.Tensor` of shape `(batch_size, sequence_length)`, *optional*):
@@ -669,7 +666,6 @@ class TFOpenAIGPTDoubleHeadsModel(TFOpenAIGPTPreTrainedModel):
         output_hidden_states: Optional[bool] = None,
         return_dict: Optional[bool] = None,
         training: Optional[bool] = False,
-        **kwargs,
     ) -> Union[Tuple, TFOpenAIGPTDoubleHeadsModelOutput]:
         r"""
         mc_token_ids (`tf.Tensor` or `Numpy array` of shape `(batch_size, num_choices)`, *optional*, default to index of the last token of the input):
@@ -813,7 +809,6 @@ class TFOpenAIGPTForSequenceClassification(TFOpenAIGPTPreTrainedModel, TFSequenc
         return_dict: Optional[bool] = None,
         labels: Optional[Union[np.ndarray, tf.Tensor]] = None,
         training: Optional[bool] = False,
-        **kwargs,
     ) -> Union[Tuple, TFSequenceClassifierOutput]:
         r"""
         labels (`tf.Tensor` of shape `(batch_size, sequence_length)`, *optional*):

src/transformers/models/pegasus/modeling_tf_pegasus.py

@@ -710,7 +710,6 @@ class TFPegasusEncoder(tf.keras.layers.Layer):
         output_hidden_states=None,
         return_dict=None,
         training=False,
-        **kwargs,
     ):
         """
         Args:
@@ -872,7 +871,6 @@ class TFPegasusDecoder(tf.keras.layers.Layer):
         output_hidden_states=None,
         return_dict=None,
         training=False,
-        **kwargs,
     ):
         r"""
         Args:
@@ -1305,7 +1303,6 @@ class TFPegasusForConditionalGeneration(TFPegasusPreTrainedModel, TFCausalLangua
         return_dict=None,
         labels=None,
         training=False,
-        **kwargs,
     ):
         """
         labels (`tf.tensor` of shape `(batch_size, sequence_length)`, *optional*):

src/transformers/models/rembert/modeling_tf_rembert.py

@@ -660,7 +660,6 @@ class TFRemBertMainLayer(tf.keras.layers.Layer):
         output_hidden_states: Optional[bool] = None,
         return_dict: Optional[bool] = None,
         training: bool = False,
-        **kwargs,
     ) -> Union[TFBaseModelOutputWithPoolingAndCrossAttentions, Tuple[tf.Tensor]]:
 
         if not self.config.is_decoder:
@@ -959,7 +958,6 @@ class TFRemBertModel(TFRemBertPreTrainedModel):
         output_hidden_states: Optional[bool] = None,
         return_dict: Optional[bool] = None,
         training: Optional[bool] = False,
-        **kwargs,
     ) -> Union[TFBaseModelOutputWithPoolingAndCrossAttentions, Tuple[tf.Tensor]]:
         r"""
         encoder_hidden_states  (`tf.Tensor` of shape `(batch_size, sequence_length, hidden_size)`, *optional*):
@@ -1060,7 +1058,6 @@ class TFRemBertForMaskedLM(TFRemBertPreTrainedModel, TFMaskedLanguageModelingLos
         return_dict: Optional[bool] = None,
         labels: Optional[Union[np.ndarray, tf.Tensor]] = None,
         training: Optional[bool] = False,
-        **kwargs,
     ) -> Union[TFMaskedLMOutput, Tuple[tf.Tensor]]:
         r"""
         labels (`tf.Tensor` or `np.ndarray` of shape `(batch_size, sequence_length)`, *optional*):
@@ -1155,7 +1152,6 @@ class TFRemBertForCausalLM(TFRemBertPreTrainedModel, TFCausalLanguageModelingLos
         return_dict: Optional[bool] = None,
         labels: Optional[Union[np.ndarray, tf.Tensor]] = None,
         training: Optional[bool] = False,
-        **kwargs,
     ) -> Union[TFCausalLMOutputWithCrossAttentions, Tuple[tf.Tensor]]:
         r"""
         encoder_hidden_states  (`tf.Tensor` of shape `(batch_size, sequence_length, hidden_size)`, *optional*):
@@ -1283,7 +1279,6 @@ class TFRemBertForSequenceClassification(TFRemBertPreTrainedModel, TFSequenceCla
         return_dict: Optional[bool] = None,
         labels: Optional[Union[np.ndarray, tf.Tensor]] = None,
         training: Optional[bool] = False,
-        **kwargs,
     ) -> Union[TFSequenceClassifierOutput, Tuple[tf.Tensor]]:
         r"""
         labels (`tf.Tensor` or `np.ndarray` of shape `(batch_size,)`, *optional*):
@@ -1374,7 +1369,6 @@ class TFRemBertForMultipleChoice(TFRemBertPreTrainedModel, TFMultipleChoiceLoss)
         return_dict: Optional[bool] = None,
         labels: Optional[Union[np.ndarray, tf.Tensor]] = None,
         training: Optional[bool] = False,
-        **kwargs,
     ) -> Union[TFMultipleChoiceModelOutput, Tuple[tf.Tensor]]:
         r"""
         labels (`tf.Tensor` or `np.ndarray` of shape `(batch_size,)`, *optional*):
@@ -1494,7 +1488,6 @@ class TFRemBertForTokenClassification(TFRemBertPreTrainedModel, TFTokenClassific
         return_dict: Optional[bool] = None,
         labels: Optional[Union[np.ndarray, tf.Tensor]] = None,
         training: Optional[bool] = False,
-        **kwargs,
     ) -> Union[TFTokenClassifierOutput, Tuple[tf.Tensor]]:
         r"""
         labels (`tf.Tensor` or `np.ndarray` of shape `(batch_size, sequence_length)`, *optional*):
@@ -1575,7 +1568,6 @@ class TFRemBertForQuestionAnswering(TFRemBertPreTrainedModel, TFQuestionAnswerin
         start_positions: Optional[Union[np.ndarray, tf.Tensor]] = None,
         end_positions: Optional[Union[np.ndarray, tf.Tensor]] = None,
         training: Optional[bool] = False,
-        **kwargs,
     ) -> Union[TFQuestionAnsweringModelOutput, Tuple[tf.Tensor]]:
         r"""
         start_positions (`tf.Tensor` or `np.ndarray` of shape `(batch_size,)`, *optional*):

src/transformers/models/roberta/modeling_tf_roberta.py

@@ -624,7 +624,6 @@ class TFRobertaMainLayer(tf.keras.layers.Layer):
         output_hidden_states: Optional[bool] = None,
         return_dict: Optional[bool] = None,
         training: bool = False,
-        **kwargs,
     ) -> Union[TFBaseModelOutputWithPoolingAndCrossAttentions, Tuple[tf.Tensor]]:
 
         if not self.config.is_decoder:
@@ -936,7 +935,6 @@ class TFRobertaModel(TFRobertaPreTrainedModel):
         output_hidden_states: Optional[bool] = None,
         return_dict: Optional[bool] = None,
         training: Optional[bool] = False,
-        **kwargs,
     ) -> Union[Tuple, TFBaseModelOutputWithPoolingAndCrossAttentions]:
         r"""
         encoder_hidden_states  (`tf.Tensor` of shape `(batch_size, sequence_length, hidden_size)`, *optional*):
@@ -1093,7 +1091,6 @@ class TFRobertaForMaskedLM(TFRobertaPreTrainedModel, TFMaskedLanguageModelingLos
         return_dict: Optional[bool] = None,
         labels: Optional[Union[np.ndarray, tf.Tensor]] = None,
         training: Optional[bool] = False,
-        **kwargs,
     ) -> Union[TFMaskedLMOutput, Tuple[tf.Tensor]]:
         r"""
         labels (`tf.Tensor` of shape `(batch_size, sequence_length)`, *optional*):
@@ -1196,7 +1193,6 @@ class TFRobertaForCausalLM(TFRobertaPreTrainedModel, TFCausalLanguageModelingLos
         return_dict: Optional[bool] = None,
         labels: Optional[Union[np.ndarray, tf.Tensor]] = None,
         training: Optional[bool] = False,
-        **kwargs,
     ) -> Union[TFCausalLMOutputWithCrossAttentions, Tuple[tf.Tensor]]:
         r"""
         encoder_hidden_states  (`tf.Tensor` of shape `(batch_size, sequence_length, hidden_size)`, *optional*):
@@ -1353,7 +1349,6 @@ class TFRobertaForSequenceClassification(TFRobertaPreTrainedModel, TFSequenceCla
         return_dict: Optional[bool] = None,
         labels: Optional[Union[np.ndarray, tf.Tensor]] = None,
         training: Optional[bool] = False,
-        **kwargs,
     ) -> Union[TFSequenceClassifierOutput, Tuple[tf.Tensor]]:
         r"""
         labels (`tf.Tensor` of shape `(batch_size,)`, *optional*):
@@ -1449,7 +1444,6 @@ class TFRobertaForMultipleChoice(TFRobertaPreTrainedModel, TFMultipleChoiceLoss)
         return_dict: Optional[bool] = None,
         labels: Optional[Union[np.ndarray, tf.Tensor]] = None,
         training: Optional[bool] = False,
-        **kwargs,
     ) -> Union[TFMultipleChoiceModelOutput, Tuple[tf.Tensor]]:
         r"""
         labels (`tf.Tensor` of shape `(batch_size,)`, *optional*):
@@ -1567,7 +1561,6 @@ class TFRobertaForTokenClassification(TFRobertaPreTrainedModel, TFTokenClassific
         return_dict: Optional[bool] = None,
         labels: Optional[Union[np.ndarray, tf.Tensor]] = None,
         training: Optional[bool] = False,
-        **kwargs,
     ) -> Union[TFTokenClassifierOutput, Tuple[tf.Tensor]]:
         r"""
         labels (`tf.Tensor` of shape `(batch_size, sequence_length)`, *optional*):
@@ -1655,7 +1648,6 @@ class TFRobertaForQuestionAnswering(TFRobertaPreTrainedModel, TFQuestionAnswerin
         start_positions: Optional[Union[np.ndarray, tf.Tensor]] = None,
         end_positions: Optional[Union[np.ndarray, tf.Tensor]] = None,
         training: Optional[bool] = False,
-        **kwargs,
     ) -> Union[TFQuestionAnsweringModelOutput, Tuple[tf.Tensor]]:
         r"""
         start_positions (`tf.Tensor` of shape `(batch_size,)`, *optional*):

src/transformers/models/roformer/modeling_tf_roformer.py

@@ -614,7 +614,6 @@ class TFRoFormerMainLayer(tf.keras.layers.Layer):
         output_hidden_states: Optional[bool] = None,
         return_dict: Optional[bool] = None,
         training: bool = False,
-        **kwargs,
     ) -> Union[TFBaseModelOutput, Tuple[tf.Tensor]]:
 
         if input_ids is not None and inputs_embeds is not None:
@@ -817,7 +816,6 @@ class TFRoFormerModel(TFRoFormerPreTrainedModel):
         output_hidden_states: Optional[bool] = None,
         return_dict: Optional[bool] = None,
         training: Optional[bool] = False,
-        **kwargs,
     ) -> Union[TFBaseModelOutputWithPooling, Tuple[tf.Tensor]]:
         outputs = self.roformer(
             input_ids=input_ids,
@@ -877,7 +875,6 @@ class TFRoFormerForMaskedLM(TFRoFormerPreTrainedModel, TFMaskedLanguageModelingL
         return_dict: Optional[bool] = None,
         labels: Optional[Union[np.ndarray, tf.Tensor]] = None,
         training: Optional[bool] = False,
-        **kwargs,
     ) -> Union[TFMaskedLMOutput, Tuple[tf.Tensor]]:
         r"""
         labels (`tf.Tensor` or `np.ndarray` of shape `(batch_size, sequence_length)`, *optional*):
@@ -953,7 +950,6 @@ class TFRoFormerForCausalLM(TFRoFormerPreTrainedModel, TFCausalLanguageModelingL
         return_dict: Optional[bool] = None,
         labels: Optional[Union[np.ndarray, tf.Tensor]] = None,
         training: Optional[bool] = False,
-        **kwargs,
     ) -> Union[TFCausalLMOutput, Tuple[tf.Tensor]]:
         r"""
         labels (`tf.Tensor` or `np.ndarray` of shape `(batch_size, sequence_length)`, *optional*):
@@ -1064,7 +1060,6 @@ class TFRoFormerForSequenceClassification(TFRoFormerPreTrainedModel, TFSequenceC
         return_dict: Optional[bool] = None,
         labels: Optional[Union[np.ndarray, tf.Tensor]] = None,
         training: Optional[bool] = False,
-        **kwargs,
     ) -> Union[TFSequenceClassifierOutput, Tuple[tf.Tensor]]:
         r"""
         labels (`tf.Tensor` or `np.ndarray` of shape `(batch_size,)`, *optional*):
@@ -1155,7 +1150,6 @@ class TFRoFormerForMultipleChoice(TFRoFormerPreTrainedModel, TFMultipleChoiceLos
         return_dict: Optional[bool] = None,
         labels: Optional[Union[np.ndarray, tf.Tensor]] = None,
         training: Optional[bool] = False,
-        **kwargs,
     ) -> Union[TFMultipleChoiceModelOutput, Tuple[tf.Tensor]]:
         r"""
         labels (`tf.Tensor` or `np.ndarray` of shape `(batch_size,)`, *optional*):
@@ -1269,7 +1263,6 @@ class TFRoFormerForTokenClassification(TFRoFormerPreTrainedModel, TFTokenClassif
         return_dict: Optional[bool] = None,
         labels: Optional[Union[np.ndarray, tf.Tensor]] = None,
         training: Optional[bool] = False,
-        **kwargs,
     ) -> Union[TFTokenClassifierOutput, Tuple[tf.Tensor]]:
         r"""
         labels (`tf.Tensor` or `np.ndarray` of shape `(batch_size, sequence_length)`, *optional*):
@@ -1348,7 +1341,6 @@ class TFRoFormerForQuestionAnswering(TFRoFormerPreTrainedModel, TFQuestionAnswer
         start_positions: Optional[Union[np.ndarray, tf.Tensor]] = None,
         end_positions: Optional[Union[np.ndarray, tf.Tensor]] = None,
         training: Optional[bool] = False,
-        **kwargs,
     ) -> Union[TFQuestionAnsweringModelOutput, Tuple[tf.Tensor]]:
         r"""
         start_positions (`tf.Tensor` or `np.ndarray` of shape `(batch_size,)`, *optional*):

src/transformers/models/speech_to_text/modeling_tf_speech_to_text.py

@@ -791,7 +791,6 @@ class TFSpeech2TextEncoder(tf.keras.layers.Layer):
         output_hidden_states=None,
         return_dict=None,
         training=False,
-        **kwargs,
     ):
         """
         Args:
@@ -957,7 +956,6 @@ class TFSpeech2TextDecoder(tf.keras.layers.Layer):
         output_hidden_states=None,
         return_dict=None,
         training=False,
-        **kwargs,
     ):
         r"""
         Args:

src/transformers/models/t5/modeling_tf_t5.py

@@ -654,7 +654,6 @@ class TFT5MainLayer(tf.keras.layers.Layer):
         output_hidden_states=None,
         return_dict=None,
         training=False,
-        **kwargs,
     ) -> Tuple:
 
         if input_ids is not None and inputs_embeds is not None:
@@ -1152,7 +1151,6 @@ class TFT5Model(TFT5PreTrainedModel):
         output_hidden_states: Optional[bool] = None,
         return_dict: Optional[bool] = None,
         training: Optional[bool] = False,
-        **kwargs,
     ) -> Union[Tuple, TFSeq2SeqModelOutput]:
         r"""
         Returns:
@@ -1329,7 +1327,6 @@ class TFT5ForConditionalGeneration(TFT5PreTrainedModel, TFCausalLanguageModeling
         output_hidden_states: Optional[bool] = None,
         return_dict: Optional[bool] = None,
         training: Optional[bool] = False,
-        **kwargs,
     ) -> Union[Tuple, TFSeq2SeqLMOutput]:
         r"""
         labels (`tf.Tensor` of shape `(batch_size, sequence_length)`, *optional*):
@@ -1611,6 +1608,10 @@ class TFT5EncoderModel(TFT5PreTrainedModel):
         encoder_config.use_cache = False
         self.encoder = TFT5MainLayer(encoder_config, embed_tokens, name="encoder")
 
+    @property
+    def dummy_inputs(self):
+        return {"input_ids": tf.constant(DUMMY_INPUTS)}
+
     def get_encoder(self):
         return self.encoder
 
@@ -1627,7 +1628,6 @@ class TFT5EncoderModel(TFT5PreTrainedModel):
         output_hidden_states: Optional[bool] = None,
         return_dict: Optional[bool] = None,
         training: Optional[bool] = False,
-        **kwargs,
     ) -> Union[Tuple, TFBaseModelOutput]:
         r"""
         Returns:
@@ -1670,6 +1670,19 @@ class TFT5EncoderModel(TFT5PreTrainedModel):
             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

src/transformers/models/tapas/modeling_tf_tapas.py

@@ -770,7 +770,6 @@ class TFTapasMainLayer(tf.keras.layers.Layer):
         output_hidden_states: Optional[bool] = None,
         return_dict: Optional[bool] = None,
         training: bool = False,
-        **kwargs,
     ) -> Union[TFBaseModelOutputWithPooling, Tuple[tf.Tensor]]:
 
         if input_ids is not None and inputs_embeds is not None:
@@ -980,7 +979,6 @@ class TFTapasModel(TFTapasPreTrainedModel):
         output_hidden_states: Optional[bool] = None,
         return_dict: Optional[bool] = None,
         training: Optional[bool] = False,
-        **kwargs,
     ) -> Union[TFBaseModelOutputWithPooling, Tuple[tf.Tensor]]:
         r"""
         Returns:
@@ -1067,7 +1065,6 @@ class TFTapasForMaskedLM(TFTapasPreTrainedModel, TFMaskedLanguageModelingLoss):
         return_dict: Optional[bool] = None,
         labels: Optional[Union[np.ndarray, tf.Tensor]] = None,
         training: Optional[bool] = False,
-        **kwargs,
     ) -> Union[TFMaskedLMOutput, Tuple[tf.Tensor]]:
         r"""
         labels (`tf.Tensor` or `np.ndarray` of shape `(batch_size, sequence_length)`, *optional*):
@@ -1285,7 +1282,6 @@ class TFTapasForQuestionAnswering(TFTapasPreTrainedModel):
         return_dict: Optional[bool] = None,
         labels: Optional[Union[np.ndarray, tf.Tensor]] = None,
         training: Optional[bool] = False,
-        **kwargs,
     ) -> Union[TFTableQuestionAnsweringOutput, Tuple[tf.Tensor]]:
         r"""
         table_mask (`tf.Tensor` of shape `(batch_size, seq_length)`, *optional*):
@@ -1602,7 +1598,6 @@ class TFTapasForSequenceClassification(TFTapasPreTrainedModel, TFSequenceClassif
         return_dict: Optional[bool] = None,
         labels: Optional[Union[np.ndarray, tf.Tensor]] = None,
         training: Optional[bool] = False,
-        **kwargs,
     ) -> Union[TFSequenceClassifierOutput, Tuple[tf.Tensor]]:
         r"""
         labels (`torch.LongTensor` of shape `(batch_size,)`, *optional*):

src/transformers/models/transfo_xl/modeling_tf_transfo_xl.py

@@ -550,7 +550,6 @@ class TFTransfoXLMainLayer(tf.keras.layers.Layer):
         output_hidden_states=None,
         return_dict=None,
         training=False,
-        **kwargs,
     ):
 
         # the original code for Transformer-XL used shapes [len, bsz] but we want a unified interface in the library
@@ -898,7 +897,6 @@ class TFTransfoXLModel(TFTransfoXLPreTrainedModel):
         output_hidden_states=None,
         return_dict=None,
         training=False,
-        **kwargs,
     ):
         outputs = self.transformer(
             input_ids=input_ids,
@@ -979,7 +977,6 @@ class TFTransfoXLLMHeadModel(TFTransfoXLPreTrainedModel):
         return_dict=None,
         labels=None,
         training=False,
-        **kwargs,
     ):
         if input_ids is not None:
             bsz, tgt_len = shape_list(input_ids)[:2]
@@ -1088,7 +1085,6 @@ class TFTransfoXLForSequenceClassification(TFTransfoXLPreTrainedModel, TFSequenc
         return_dict: Optional[bool] = None,
         labels: Optional[Union[np.ndarray, tf.Tensor]] = None,
         training: Optional[bool] = False,
-        **kwargs,
     ) -> Union[Tuple, TFTransfoXLSequenceClassifierOutputWithPast]:
         r"""
         labels (`tf.Tensor` of shape `(batch_size, sequence_length)`, *optional*):

src/transformers/models/vision_encoder_decoder/modeling_tf_vision_encoder_decoder.py

@@ -23,7 +23,7 @@ import tensorflow as tf
 
 from ...configuration_utils import PretrainedConfig
 from ...modeling_tf_outputs import TFBaseModelOutput, TFSeq2SeqLMOutput
-from ...modeling_tf_utils import TFCausalLanguageModelingLoss, TFPreTrainedModel, get_initializer, input_processing
+from ...modeling_tf_utils import TFCausalLanguageModelingLoss, TFPreTrainedModel, get_initializer, unpack_inputs
 from ...tf_utils import shape_list
 from ...utils import (
     DUMMY_INPUTS,
@@ -510,6 +510,7 @@ class TFVisionEncoderDecoderModel(TFPreTrainedModel, TFCausalLanguageModelingLos
         config = VisionEncoderDecoderConfig.from_encoder_decoder_configs(encoder.config, decoder.config, **kwargs)
         return cls(encoder=encoder, decoder=decoder, config=config)
 
+    @unpack_inputs
     @add_start_docstrings_to_model_forward(
         VISION_ENCODER_DECODER_INPUTS_DOCSTRING.format("batch_size, sequence_length")
     )
@@ -585,21 +586,16 @@ class TFVisionEncoderDecoderModel(TFPreTrainedModel, TFCausalLanguageModelingLos
 
         if encoder_outputs is None:
 
-            encoder_processing_inputs = {
-                "func": self.encoder.call,
-                "config": self.encoder.config,
+            encoder_inputs = {
                 "input_ids": pixel_values,
                 "output_attentions": output_attentions,
                 "output_hidden_states": output_hidden_states,
                 "return_dict": return_dict,
                 "training": training,
-                "kwargs_call": {},
             }
 
             # Add arguments to encoder from `kwargs_encoder`
-            encoder_processing_inputs.update(kwargs_encoder)
-
-            encoder_inputs = input_processing(**encoder_processing_inputs)
+            encoder_inputs.update(kwargs_encoder)
 
             if "input_ids" in encoder_inputs:
                 encoder_inputs["pixel_values"] = encoder_inputs.pop("input_ids")
@@ -639,9 +635,7 @@ class TFVisionEncoderDecoderModel(TFPreTrainedModel, TFCausalLanguageModelingLos
         batch_size, sequence_length = shape_list(encoder_hidden_states)[:2]
         encoder_attention_mask = tf.ones(shape=(batch_size, sequence_length), dtype=tf.int32)
 
-        decoder_processing_inputs = {
-            "func": self.decoder.call,
-            "config": self.decoder.config,
+        decoder_inputs = {
             "input_ids": decoder_input_ids,
             "attention_mask": decoder_attention_mask,
             "encoder_hidden_states": encoder_hidden_states,
@@ -653,13 +647,11 @@ class TFVisionEncoderDecoderModel(TFPreTrainedModel, TFCausalLanguageModelingLos
             "past_key_values": past_key_values,
             "return_dict": return_dict,
             "training": training,
-            "kwargs_call": {},
         }
 
         # Add arguments to decoder from `kwargs_decoder`
-        decoder_processing_inputs.update(kwargs_decoder)
+        decoder_inputs.update(kwargs_decoder)
 
-        decoder_inputs = input_processing(**decoder_processing_inputs)
         decoder_outputs = self.decoder(**decoder_inputs)
 
         logits = decoder_outputs[0]

src/transformers/models/vit/modeling_tf_vit.py

@@ -486,7 +486,6 @@ class TFViTMainLayer(tf.keras.layers.Layer):
         interpolate_pos_encoding: Optional[bool] = None,
         return_dict: Optional[bool] = None,
         training: bool = False,
-        **kwargs,
     ) -> Union[TFBaseModelOutputWithPooling, Tuple[tf.Tensor]]:
 
         if pixel_values is None:
@@ -656,7 +655,6 @@ class TFViTModel(TFViTPreTrainedModel):
         interpolate_pos_encoding: Optional[bool] = None,
         return_dict: Optional[bool] = None,
         training: bool = False,
-        **kwargs,
     ) -> Union[TFBaseModelOutputWithPooling, Tuple[tf.Tensor]]:
         r"""
         Returns:
@@ -757,7 +755,6 @@ class TFViTForImageClassification(TFViTPreTrainedModel, TFSequenceClassification
         return_dict: Optional[bool] = None,
         labels: Optional[Union[np.ndarray, tf.Tensor]] = None,
         training: Optional[bool] = False,
-        **kwargs,
     ) -> Union[TFSequenceClassifierOutput, Tuple[tf.Tensor]]:
         r"""
         labels (`tf.Tensor` or `np.ndarray` of shape `(batch_size,)`, *optional*):

src/transformers/models/vit_mae/modeling_tf_vit_mae.py

@@ -647,7 +647,6 @@ class TFViTMAEMainLayer(tf.keras.layers.Layer):
         output_hidden_states: Optional[bool] = None,
         return_dict: Optional[bool] = None,
         training: bool = False,
-        **kwargs,
     ) -> Union[TFViTMAEModelOutput, Tuple[tf.Tensor]]:
         embedding_output, mask, ids_restore = self.embeddings(
             pixel_values=pixel_values, training=training, noise=noise
@@ -811,7 +810,6 @@ class TFViTMAEModel(TFViTMAEPreTrainedModel):
         output_hidden_states: Optional[bool] = None,
         return_dict: Optional[bool] = None,
         training: bool = False,
-        **kwargs,
     ) -> Union[TFViTMAEModelOutput, Tuple[tf.Tensor]]:
         r"""
         Returns:
@@ -1028,7 +1026,6 @@ class TFViTMAEForPreTraining(TFViTMAEPreTrainedModel):
         output_hidden_states: Optional[bool] = None,
         return_dict: Optional[bool] = None,
         training: bool = False,
-        **kwargs,
     ) -> Union[TFViTMAEForPreTrainingOutput, Tuple[tf.Tensor]]:
         r"""
         Returns: