More TF int dtype fixes (#20384)

* Add a test to ensure int dummy inputs are int64

* Move the test into the existing int64 test and update a lot of existing dummies

* Fix remaining dummies

* Fix remaining dummies

* Test for int64 serving sigs as well

* Update core tests to use tf.int64

* Add better messages to the assertions

* Update all serving sigs to int64

* More sneaky hiding tf.int32s

* Add an optional int32 signature in save_pretrained

* make fixup

* Add Amy's suggestions

* Switch all serving sigs back to tf.int32

* Switch all dummies to tf.int32

* Adjust tests to check for tf.int32 instead of tf.int64

* Fix base dummy_inputs dtype

* Start casting to tf.int32 in input_processing

* Change dtype for unpack_inputs test

* Add proper tf.int32 test

* Make the alternate serving signature int64

src/transformers/modeling_tf_utils.py

@@ -585,10 +585,10 @@ def input_processing(func, config, **kwargs):
 
     cast_output = dict()
     for key, val in output.items():
-        if isinstance(val, tf.Tensor) and val.dtype == tf.int32:
-            cast_output[key] = tf.cast(val, tf.int64)
-        elif isinstance(val, np.ndarray) and val.dtype == np.int32:
-            cast_output[key] = val.astype(np.int64)
+        if isinstance(val, tf.Tensor) and val.dtype == tf.int64:
+            cast_output[key] = tf.cast(val, tf.int32)
+        elif isinstance(val, np.ndarray) and val.dtype == np.int64:
+            cast_output[key] = val.astype(np.int32)
         else:
             cast_output[key] = val
 
@@ -1115,7 +1115,7 @@ class TFPreTrainedModel(tf.keras.Model, TFModelUtilsMixin, TFGenerationMixin, Pu
             `Dict[str, tf.Tensor]`: The dummy inputs.
         """
         return {
-            "input_ids": tf.constant(DUMMY_INPUTS),
+            "input_ids": tf.constant(DUMMY_INPUTS, dtype=tf.int32),
         }
 
     @property
@@ -1155,12 +1155,25 @@ class TFPreTrainedModel(tf.keras.Model, TFModelUtilsMixin, TFGenerationMixin, Pu
         """
         return cls(config, **kwargs)
 
+    def eager_serving(self, inputs):
+        """
+        Method used for serving the model. Intended not to be compiled with a tf.function decorator so that we can use
+        it to generate multiple signatures later.
+
+        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)
+
     @tf.function(
         input_signature=[
             {
-                "input_ids": tf.TensorSpec((None, None), tf.int64, name="input_ids"),
-                "attention_mask": tf.TensorSpec((None, None), tf.int64, name="attention_mask"),
-                "token_type_ids": tf.TensorSpec((None, None), tf.int64, name="token_type_ids"),
+                "input_ids": tf.TensorSpec((None, None), tf.int32, name="input_ids"),
+                "attention_mask": tf.TensorSpec((None, None), tf.int32, name="attention_mask"),
+                "token_type_ids": tf.TensorSpec((None, None), tf.int32, name="token_type_ids"),
             }
         ]
     )
@@ -2253,7 +2266,17 @@ class TFPreTrainedModel(tf.keras.Model, TFModelUtilsMixin, TFGenerationMixin, Pu
 
         if saved_model:
             if signatures is None:
-                signatures = self.serving
+                if any(spec.dtype == tf.int32 for spec in self.serving.input_signature[0].values()):
+                    int64_spec = {
+                        key: tf.TensorSpec(
+                            shape=spec.shape, dtype=tf.int64 if spec.dtype == tf.int32 else spec.dtype, name=spec.name
+                        )
+                        for key, spec in self.serving.input_signature[0].items()
+                    }
+                    int64_serving = tf.function(self.eager_serving, input_signature=[int64_spec])
+                    signatures = {"serving_default": self.serving, "int64_serving": int64_serving}
+                else:
+                    signatures = self.serving
             saved_model_dir = os.path.join(save_directory, "saved_model", str(version))
             self.save(saved_model_dir, include_optimizer=False, signatures=signatures)
             logger.info(f"Saved model created in {saved_model_dir}")

src/transformers/models/albert/modeling_tf_albert.py

@@ -1396,7 +1396,7 @@ class TFAlbertForMultipleChoice(TFAlbertPreTrainedModel, TFMultipleChoiceLoss):
         Returns:
             tf.Tensor with dummy inputs
         """
-        return {"input_ids": tf.constant(MULTIPLE_CHOICE_DUMMY_INPUTS)}
+        return {"input_ids": tf.constant(MULTIPLE_CHOICE_DUMMY_INPUTS, dtype=tf.int32)}
 
     @unpack_inputs
     @add_start_docstrings_to_model_forward(ALBERT_INPUTS_DOCSTRING.format("batch_size, num_choices, sequence_length"))

src/transformers/models/bart/modeling_tf_bart.py

@@ -467,11 +467,11 @@ class TFBartPretrainedModel(TFPreTrainedModel):
     @property
     def dummy_inputs(self):
         pad_token = 1
-        input_ids = tf.cast(tf.convert_to_tensor(DUMMY_INPUTS), tf.int32)
-        decoder_input_ids = tf.cast(tf.convert_to_tensor(DUMMY_INPUTS), tf.int32)
+        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.math.not_equal(input_ids, pad_token),
+            "attention_mask": tf.cast(input_ids != pad_token, tf.int32),
             "input_ids": input_ids,
         }
         return dummy_inputs
@@ -479,10 +479,10 @@ class TFBartPretrainedModel(TFPreTrainedModel):
     @tf.function(
         input_signature=[
             {
-                "input_ids": tf.TensorSpec((None, None), tf.int64, name="input_ids"),
-                "attention_mask": tf.TensorSpec((None, None), tf.int64, name="attention_mask"),
-                "decoder_input_ids": tf.TensorSpec((None, None), tf.int64, name="decoder_input_ids"),
-                "decoder_attention_mask": tf.TensorSpec((None, None), tf.int64, name="decoder_attention_mask"),
+                "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"),
             }
         ]
     )

src/transformers/models/bert/modeling_tf_bert.py

@@ -920,7 +920,7 @@ class TFBertPreTrainedModel(TFPreTrainedModel):
         Returns:
             `Dict[str, tf.Tensor]`: The dummy inputs.
         """
-        dummy = {"input_ids": tf.constant(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
@@ -1726,7 +1726,7 @@ class TFBertForMultipleChoice(TFBertPreTrainedModel, TFMultipleChoiceLoss):
         Returns:
             tf.Tensor with dummy inputs
         """
-        return {"input_ids": tf.constant(MULTIPLE_CHOICE_DUMMY_INPUTS)}
+        return {"input_ids": tf.constant(MULTIPLE_CHOICE_DUMMY_INPUTS, dtype=tf.int32)}
 
     @unpack_inputs
     @add_start_docstrings_to_model_forward(BERT_INPUTS_DOCSTRING.format("batch_size, num_choices, sequence_length"))
@@ -1809,9 +1809,9 @@ class TFBertForMultipleChoice(TFBertPreTrainedModel, TFMultipleChoiceLoss):
     @tf.function(
         input_signature=[
             {
-                "input_ids": tf.TensorSpec((None, None, None), tf.int64, name="input_ids"),
-                "attention_mask": tf.TensorSpec((None, None, None), tf.int64, name="attention_mask"),
-                "token_type_ids": tf.TensorSpec((None, None, None), tf.int64, name="token_type_ids"),
+                "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"),
             }
         ]
     )

src/transformers/models/blenderbot/modeling_tf_blenderbot.py

@@ -466,11 +466,11 @@ class TFBlenderbotPreTrainedModel(TFPreTrainedModel):
     @property
     def dummy_inputs(self):
         pad_token = 1
-        input_ids = tf.cast(tf.convert_to_tensor(DUMMY_INPUTS), tf.int32)
-        decoder_input_ids = tf.cast(tf.convert_to_tensor(DUMMY_INPUTS), tf.int32)
+        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.math.not_equal(input_ids, pad_token),
+            "attention_mask": tf.cast(input_ids != pad_token, tf.int32),
             "input_ids": input_ids,
         }
         return dummy_inputs

src/transformers/models/blenderbot_small/modeling_tf_blenderbot_small.py

@@ -466,11 +466,11 @@ class TFBlenderbotSmallPreTrainedModel(TFPreTrainedModel):
     @property
     def dummy_inputs(self):
         pad_token = 1
-        input_ids = tf.cast(tf.convert_to_tensor(DUMMY_INPUTS), tf.int32)
-        decoder_input_ids = tf.cast(tf.convert_to_tensor(DUMMY_INPUTS), tf.int32)
+        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.math.not_equal(input_ids, pad_token),
+            "attention_mask": tf.cast(input_ids != pad_token, tf.int32),
             "input_ids": input_ids,
         }
         return dummy_inputs

src/transformers/models/camembert/modeling_tf_camembert.py

@@ -891,7 +891,7 @@ class TFCamembertPreTrainedModel(TFPreTrainedModel):
         Returns:
             `Dict[str, tf.Tensor]`: The dummy inputs.
         """
-        dummy = {"input_ids": tf.constant(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
@@ -904,8 +904,8 @@ class TFCamembertPreTrainedModel(TFPreTrainedModel):
     @tf.function(
         input_signature=[
             {
-                "input_ids": tf.TensorSpec((None, None), tf.int64, name="input_ids"),
-                "attention_mask": tf.TensorSpec((None, None), tf.int64, name="attention_mask"),
+                "input_ids": tf.TensorSpec((None, None), tf.int32, name="input_ids"),
+                "attention_mask": tf.TensorSpec((None, None), tf.int32, name="attention_mask"),
             }
         ]
     )
@@ -1390,7 +1390,7 @@ class TFCamembertForMultipleChoice(TFCamembertPreTrainedModel, TFMultipleChoiceL
         Returns:
             tf.Tensor with dummy inputs
         """
-        return {"input_ids": tf.constant(MULTIPLE_CHOICE_DUMMY_INPUTS)}
+        return {"input_ids": tf.constant(MULTIPLE_CHOICE_DUMMY_INPUTS, dtype=tf.int32)}
 
     @unpack_inputs
     @add_start_docstrings_to_model_forward(

src/transformers/models/clip/modeling_tf_clip.py

@@ -1107,8 +1107,8 @@ class TFCLIPTextModel(TFCLIPPreTrainedModel):
     @tf.function(
         input_signature=[
             {
-                "input_ids": tf.TensorSpec((None, None), tf.int64, name="input_ids"),
-                "attention_mask": tf.TensorSpec((None, None), tf.int64, name="attention_mask"),
+                "input_ids": tf.TensorSpec((None, None), tf.int32, name="input_ids"),
+                "attention_mask": tf.TensorSpec((None, None), tf.int32, name="attention_mask"),
             }
         ]
     )

src/transformers/models/convbert/modeling_tf_convbert.py

@@ -1058,7 +1058,7 @@ class TFConvBertForMultipleChoice(TFConvBertPreTrainedModel, TFMultipleChoiceLos
         Returns:
             tf.Tensor with dummy inputs
         """
-        return {"input_ids": tf.convert_to_tensor(MULTIPLE_CHOICE_DUMMY_INPUTS)}
+        return {"input_ids": tf.convert_to_tensor(MULTIPLE_CHOICE_DUMMY_INPUTS, dtype=tf.int32)}
 
     @unpack_inputs
     @add_start_docstrings_to_model_forward(
@@ -1137,9 +1137,9 @@ class TFConvBertForMultipleChoice(TFConvBertPreTrainedModel, TFMultipleChoiceLos
     @tf.function(
         input_signature=[
             {
-                "input_ids": tf.TensorSpec((None, None, None), tf.int64, name="input_ids"),
-                "attention_mask": tf.TensorSpec((None, None, None), tf.int64, name="attention_mask"),
-                "token_type_ids": tf.TensorSpec((None, None, None), tf.int64, name="token_type_ids"),
+                "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"),
             }
         ]
     )

src/transformers/models/distilbert/modeling_tf_distilbert.py

@@ -434,8 +434,8 @@ class TFDistilBertPreTrainedModel(TFPreTrainedModel):
     @tf.function(
         input_signature=[
             {
-                "input_ids": tf.TensorSpec((None, None), tf.int64, name="input_ids"),
-                "attention_mask": tf.TensorSpec((None, None), tf.int64, name="attention_mask"),
+                "input_ids": tf.TensorSpec((None, None), tf.int32, name="input_ids"),
+                "attention_mask": tf.TensorSpec((None, None), tf.int32, name="attention_mask"),
             }
         ]
     )
@@ -898,7 +898,7 @@ class TFDistilBertForMultipleChoice(TFDistilBertPreTrainedModel, TFMultipleChoic
         Returns:
             tf.Tensor with dummy inputs
         """
-        return {"input_ids": tf.constant(MULTIPLE_CHOICE_DUMMY_INPUTS)}
+        return {"input_ids": tf.constant(MULTIPLE_CHOICE_DUMMY_INPUTS, dtype=tf.int32)}
 
     @unpack_inputs
     @add_start_docstrings_to_model_forward(

src/transformers/models/dpr/modeling_tf_dpr.py

@@ -376,8 +376,8 @@ class TFDPRPretrainedReader(TFPreTrainedModel):
     @tf.function(
         input_signature=[
             {
-                "input_ids": tf.TensorSpec((None, None), tf.int64, name="input_ids"),
-                "attention_mask": tf.TensorSpec((None, None), tf.int64, name="attention_mask"),
+                "input_ids": tf.TensorSpec((None, None), tf.int32, name="input_ids"),
+                "attention_mask": tf.TensorSpec((None, None), tf.int32, name="attention_mask"),
             }
         ]
     )

src/transformers/models/electra/modeling_tf_electra.py

@@ -613,7 +613,7 @@ class TFElectraPreTrainedModel(TFPreTrainedModel):
         Returns:
             `Dict[str, tf.Tensor]`: The dummy inputs.
         """
-        dummy = {"input_ids": tf.constant(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
@@ -1375,7 +1375,7 @@ class TFElectraForMultipleChoice(TFElectraPreTrainedModel, TFMultipleChoiceLoss)
         Returns:
             tf.Tensor with dummy inputs
         """
-        return {"input_ids": tf.constant(MULTIPLE_CHOICE_DUMMY_INPUTS)}
+        return {"input_ids": tf.constant(MULTIPLE_CHOICE_DUMMY_INPUTS, dtype=tf.int32)}
 
     @unpack_inputs
     @add_start_docstrings_to_model_forward(ELECTRA_INPUTS_DOCSTRING.format("batch_size, num_choices, sequence_length"))

src/transformers/models/encoder_decoder/modeling_tf_encoder_decoder.py

@@ -271,7 +271,7 @@ class TFEncoderDecoderModel(TFPreTrainedModel, TFCausalLanguageModelingLoss):
             `Dict[str, tf.Tensor]`: The dummy inputs.
         """
         # Add `decoder_input_ids` because `self.decoder` requires it.
-        input_ids = tf.constant(DUMMY_INPUTS)
+        input_ids = tf.constant(DUMMY_INPUTS, dtype=tf.int32)
         dummy = {"input_ids": input_ids, "decoder_input_ids": input_ids}
         return dummy
 

src/transformers/models/esm/modeling_tf_esm.py

@@ -981,8 +981,8 @@ class TFEsmModel(TFEsmPreTrainedModel):
     @tf.function(
         input_signature=[
             {
-                "input_ids": tf.TensorSpec((None, None), tf.int64, name="input_ids"),
-                "attention_mask": tf.TensorSpec((None, None), tf.int64, name="attention_mask"),
+                "input_ids": tf.TensorSpec((None, None), tf.int32, name="input_ids"),
+                "attention_mask": tf.TensorSpec((None, None), tf.int32, name="attention_mask"),
             }
         ]
     )
@@ -1113,8 +1113,8 @@ class TFEsmForMaskedLM(TFEsmPreTrainedModel, TFMaskedLanguageModelingLoss):
     @tf.function(
         input_signature=[
             {
-                "input_ids": tf.TensorSpec((None, None), tf.int64, name="input_ids"),
-                "attention_mask": tf.TensorSpec((None, None), tf.int64, name="attention_mask"),
+                "input_ids": tf.TensorSpec((None, None), tf.int32, name="input_ids"),
+                "attention_mask": tf.TensorSpec((None, None), tf.int32, name="attention_mask"),
             }
         ]
     )
@@ -1247,8 +1247,8 @@ class TFEsmForSequenceClassification(TFEsmPreTrainedModel, TFSequenceClassificat
     @tf.function(
         input_signature=[
             {
-                "input_ids": tf.TensorSpec((None, None, None), tf.int64, name="input_ids"),
-                "attention_mask": tf.TensorSpec((None, None, None), tf.int64, name="attention_mask"),
+                "input_ids": tf.TensorSpec((None, None, None), tf.int32, name="input_ids"),
+                "attention_mask": tf.TensorSpec((None, None, None), tf.int32, name="attention_mask"),
             }
         ]
     )
@@ -1344,8 +1344,8 @@ class TFEsmForTokenClassification(TFEsmPreTrainedModel, TFTokenClassificationLos
     @tf.function(
         input_signature=[
             {
-                "input_ids": tf.TensorSpec((None, None), tf.int64, name="input_ids"),
-                "attention_mask": tf.TensorSpec((None, None), tf.int64, name="attention_mask"),
+                "input_ids": tf.TensorSpec((None, None), tf.int32, name="input_ids"),
+                "attention_mask": tf.TensorSpec((None, None), tf.int32, name="attention_mask"),
             }
         ]
     )

src/transformers/models/flaubert/modeling_tf_flaubert.py

@@ -226,13 +226,13 @@ class TFFlaubertPreTrainedModel(TFPreTrainedModel):
     @property
     def dummy_inputs(self):
         # Sometimes Flaubert has language embeddings so don't forget to build them as well if needed
-        inputs_list = tf.constant([[7, 6, 0, 0, 1], [1, 2, 3, 0, 0], [0, 0, 0, 4, 5]])
-        attns_list = tf.constant([[1, 1, 0, 0, 1], [1, 1, 1, 0, 0], [1, 0, 0, 1, 1]])
+        inputs_list = tf.constant([[7, 6, 0, 0, 1], [1, 2, 3, 0, 0], [0, 0, 0, 4, 5]], dtype=tf.int32)
+        attns_list = tf.constant([[1, 1, 0, 0, 1], [1, 1, 1, 0, 0], [1, 0, 0, 1, 1]], dtype=tf.int32)
         if self.config.use_lang_emb and self.config.n_langs > 1:
             return {
                 "input_ids": inputs_list,
                 "attention_mask": attns_list,
-                "langs": tf.constant([[1, 1, 0, 0, 1], [1, 1, 1, 0, 0], [1, 0, 0, 1, 1]]),
+                "langs": tf.constant([[1, 1, 0, 0, 1], [1, 1, 1, 0, 0], [1, 0, 0, 1, 1]], dtype=tf.int32),
             }
         else:
             return {"input_ids": inputs_list, "attention_mask": attns_list}
@@ -1169,12 +1169,12 @@ class TFFlaubertForMultipleChoice(TFFlaubertPreTrainedModel, TFMultipleChoiceLos
         # Sometimes Flaubert has language embeddings so don't forget to build them as well if needed
         if self.config.use_lang_emb and self.config.n_langs > 1:
             return {
-                "input_ids": tf.constant(MULTIPLE_CHOICE_DUMMY_INPUTS),
-                "langs": tf.constant(MULTIPLE_CHOICE_DUMMY_INPUTS),
+                "input_ids": tf.constant(MULTIPLE_CHOICE_DUMMY_INPUTS, dtype=tf.int32),
+                "langs": tf.constant(MULTIPLE_CHOICE_DUMMY_INPUTS, dtype=tf.int32),
             }
         else:
             return {
-                "input_ids": tf.constant(MULTIPLE_CHOICE_DUMMY_INPUTS),
+                "input_ids": tf.constant(MULTIPLE_CHOICE_DUMMY_INPUTS, dtype=tf.int32),
             }
 
     @unpack_inputs

src/transformers/models/funnel/modeling_tf_funnel.py

@@ -1446,7 +1446,7 @@ class TFFunnelForMultipleChoice(TFFunnelPreTrainedModel, TFMultipleChoiceLoss):
         Returns:
             tf.Tensor with dummy inputs
         """
-        return {"input_ids": tf.constant(MULTIPLE_CHOICE_DUMMY_INPUTS)}
+        return {"input_ids": tf.constant(MULTIPLE_CHOICE_DUMMY_INPUTS, dtype=tf.int32)}
 
     @unpack_inputs
     @add_start_docstrings_to_model_forward(FUNNEL_INPUTS_DOCSTRING.format("batch_size, num_choices, sequence_length"))
@@ -1521,9 +1521,9 @@ class TFFunnelForMultipleChoice(TFFunnelPreTrainedModel, TFMultipleChoiceLoss):
     @tf.function(
         input_signature=[
             {
-                "input_ids": tf.TensorSpec((None, None), tf.int64, 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"),
-                "token_type_ids": tf.TensorSpec((None, None), tf.int64, name="token_type_ids"),
+                "token_type_ids": tf.TensorSpec((None, None), tf.int32, name="token_type_ids"),
             }
         ]
     )

src/transformers/models/gpt2/modeling_tf_gpt2.py

@@ -545,7 +545,7 @@ class TFGPT2PreTrainedModel(TFPreTrainedModel):
         Returns:
             `Dict[str, tf.Tensor]`: The dummy inputs.
         """
-        dummy = {"input_ids": tf.constant(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
@@ -558,8 +558,8 @@ class TFGPT2PreTrainedModel(TFPreTrainedModel):
     @tf.function(
         input_signature=[
             {
-                "input_ids": tf.TensorSpec((None, None), tf.int64, name="input_ids"),
-                "attention_mask": tf.TensorSpec((None, None), tf.int64, name="attention_mask"),
+                "input_ids": tf.TensorSpec((None, None), tf.int32, name="input_ids"),
+                "attention_mask": tf.TensorSpec((None, None), tf.int32, name="attention_mask"),
             }
         ]
     )

src/transformers/models/gptj/modeling_tf_gptj.py

@@ -531,14 +531,14 @@ class TFGPTJPreTrainedModel(TFPreTrainedModel):
         Returns:
             `Dict[str, tf.Tensor]`: The dummy inputs.
         """
-        dummy = {"input_ids": tf.constant(DUMMY_INPUTS)}
+        dummy = {"input_ids": tf.constant(DUMMY_INPUTS, dtype=tf.int32)}
         return dummy
 
     @tf.function(
         input_signature=[
             {
-                "input_ids": tf.TensorSpec((None, None), tf.int64, name="input_ids"),
-                "attention_mask": tf.TensorSpec((None, None), tf.int64, name="attention_mask"),
+                "input_ids": tf.TensorSpec((None, None), tf.int32, name="input_ids"),
+                "attention_mask": tf.TensorSpec((None, None), tf.int32, name="attention_mask"),
             }
         ]
     )

src/transformers/models/groupvit/modeling_tf_groupvit.py

@@ -1640,8 +1640,8 @@ class TFGroupViTTextModel(TFGroupViTPreTrainedModel):
     @tf.function(
         input_signature=[
             {
-                "input_ids": tf.TensorSpec((None, None), tf.int64, name="input_ids"),
-                "attention_mask": tf.TensorSpec((None, None), tf.int64, name="attention_mask"),
+                "input_ids": tf.TensorSpec((None, None), tf.int32, name="input_ids"),
+                "attention_mask": tf.TensorSpec((None, None), tf.int32, name="attention_mask"),
             }
         ]
     )
@@ -1828,9 +1828,9 @@ class TFGroupViTModel(TFGroupViTPreTrainedModel):
     @tf.function(
         input_signature=[
             {
-                "input_ids": tf.TensorSpec((None, None), tf.int64, name="input_ids"),
+                "input_ids": tf.TensorSpec((None, None), tf.int32, name="input_ids"),
                 "pixel_values": tf.TensorSpec((None, None, None, None), tf.float64, name="pixel_values"),
-                "attention_mask": tf.TensorSpec((None, None), tf.int64, name="attention_mask"),
+                "attention_mask": tf.TensorSpec((None, None), tf.int32, name="attention_mask"),
             }
         ]
     )

src/transformers/models/hubert/modeling_tf_hubert.py

@@ -1312,8 +1312,8 @@ class TFHubertPreTrainedModel(TFPreTrainedModel):
         input_signature=[
             {
                 "input_values": tf.TensorSpec((None, None), tf.float32, name="input_values"),
-                "attention_mask": tf.TensorSpec((None, None), tf.int64, name="attention_mask"),
-                "token_type_ids": tf.TensorSpec((None, None), tf.int64, name="token_type_ids"),
+                "attention_mask": tf.TensorSpec((None, None), tf.int32, name="attention_mask"),
+                "token_type_ids": tf.TensorSpec((None, None), tf.int32, name="token_type_ids"),
             }
         ]
     )