Add unpack_inputs decorator for ctrl (#16242)

* add unpack_inputs decorator for ctrl

* replace "past" with "past_key_values"
Co-authored-by: Johannes Kolbe <johannes.kolbe@tech.better.team>

src/transformers/models/ctrl/modeling_tf_ctrl.py

@@ -29,8 +29,8 @@ from ...modeling_tf_utils import (
     TFSequenceClassificationLoss,
     TFSharedEmbeddings,
     get_initializer,
-    input_processing,
     keras_serializable,
+    unpack_inputs,
 )
 from ...tf_utils import shape_list
 from ...utils import logging
@@ -254,10 +254,11 @@ class TFCTRLMainLayer(tf.keras.layers.Layer):
         """
         raise NotImplementedError
 
+    @unpack_inputs
     def call(
         self,
         input_ids=None,
-        past=None,
+        past_key_values=None,
         attention_mask=None,
         token_type_ids=None,
         position_ids=None,
@@ -270,63 +271,44 @@ class TFCTRLMainLayer(tf.keras.layers.Layer):
         training=False,
         **kwargs,
     ):
-        inputs = input_processing(
-            func=self.call,
-            config=self.config,
-            input_ids=input_ids,
-            past=past,
-            attention_mask=attention_mask,
-            token_type_ids=token_type_ids,
-            position_ids=position_ids,
-            head_mask=head_mask,
-            inputs_embeds=inputs_embeds,
-            use_cache=use_cache,
-            output_attentions=output_attentions,
-            output_hidden_states=output_hidden_states,
-            return_dict=return_dict,
-            training=training,
-            kwargs_call=kwargs,
-        )
 
         # If using past key value states, only the last tokens
         # should be given as an input
-        if inputs["past"] is not None:
-            if inputs["input_ids"] is not None:
-                inputs["input_ids"] = inputs["input_ids"][:, -1:]
-            if inputs["inputs_embeds"] is not None:
-                inputs["inputs_embeds"] = inputs["inputs_embeds"][:, -1:]
-            if inputs["token_type_ids"] is not None:
-                inputs["token_type_ids"] = inputs["token_type_ids"][:, -1:]
-
-        if inputs["input_ids"] is not None and inputs["inputs_embeds"] is not None:
+        if past_key_values is not None:
+            if input_ids is not None:
+                input_ids = input_ids[:, -1:]
+            if inputs_embeds is not None:
+                inputs_embeds = inputs_embeds[:, -1:]
+            if token_type_ids is not None:
+                token_type_ids = token_type_ids[:, -1:]
+
+        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")
-        elif inputs["input_ids"] is not None:
-            input_shape = shape_list(inputs["input_ids"])
-            inputs["input_ids"] = tf.reshape(inputs["input_ids"], [-1, input_shape[-1]])
-        elif inputs["inputs_embeds"] is not None:
-            input_shape = shape_list(inputs["inputs_embeds"])[:-1]
+        elif input_ids is not None:
+            input_shape = shape_list(input_ids)
+            input_ids = tf.reshape(input_ids, [-1, input_shape[-1]])
+        elif inputs_embeds is not None:
+            input_shape = shape_list(inputs_embeds)[:-1]
         else:
             raise ValueError("You have to specify either input_ids or inputs_embeds")
 
-        if inputs["past"] is None:
+        if past_key_values is None:
             past_length = 0
-            inputs["past"] = [None] * len(self.h)
+            past_key_values = [None] * len(self.h)
         else:
-            past_length = shape_list(inputs["past"][0][0])[-2]
-        if inputs["position_ids"] is None:
-            inputs["position_ids"] = tf.expand_dims(
-                tf.range(past_length, input_shape[-1] + past_length, dtype=tf.int32), axis=0
-            )
-            inputs["position_ids"] = tf.tile(inputs["position_ids"], [input_shape[0], 1])
+            past_length = shape_list(past_key_values[0][0])[-2]
+        if position_ids is None:
+            position_ids = tf.expand_dims(tf.range(past_length, input_shape[-1] + past_length, dtype=tf.int32), axis=0)
+            position_ids = tf.tile(position_ids, [input_shape[0], 1])
 
         # Attention mask.
-        if inputs["attention_mask"] is not None:
+        if attention_mask is not None:
             # We create a 3D attention mask from a 2D tensor mask.
             # Sizes are [batch_size, 1, 1, to_seq_length]
             # So we can broadcast to [batch_size, num_heads, from_seq_length, to_seq_length]
             # this attention mask is more simple than the triangular masking of causal attention
             # used in OpenAI GPT, we just need to prepare the broadcast dimension here.
-            inputs["attention_mask"] = tf.reshape(inputs["attention_mask"], (input_shape[0], 1, 1, input_shape[1]))
+            attention_mask = tf.reshape(attention_mask, (input_shape[0], 1, 1, input_shape[1]))
 
             # Since attention_mask is 1.0 for positions we want to attend and 0.0 for
             # masked positions, this operation will create a tensor which is 0.0 for
@@ -336,77 +318,75 @@ class TFCTRLMainLayer(tf.keras.layers.Layer):
 
             one_cst = tf.constant(1.0)
             ten_thousand_cst = tf.constant(-10000.0)
-            inputs["attention_mask"] = tf.cast(inputs["attention_mask"], dtype=one_cst.dtype)
-            inputs["attention_mask"] = tf.multiply(tf.subtract(one_cst, inputs["attention_mask"]), ten_thousand_cst)
+            attention_mask = tf.cast(attention_mask, dtype=one_cst.dtype)
+            attention_mask = tf.multiply(tf.subtract(one_cst, attention_mask), ten_thousand_cst)
 
         # Prepare head mask if needed
         # 1.0 in head_mask indicate we keep the head
         # attention_probs has shape bsz x n_heads x N x N
         # head_mask has shape n_layer x batch x n_heads x N x N
-        if inputs["head_mask"] is not None:
+        if head_mask is not None:
             raise NotImplementedError
         else:
-            inputs["head_mask"] = [None] * self.num_layers
+            head_mask = [None] * self.num_layers
 
-        if inputs["token_type_ids"] is not None:
-            inputs["token_type_ids"] = tf.reshape(
-                inputs["token_type_ids"], [-1, shape_list(inputs["token_type_ids"])[-1]]
-            )
-            token_type_embeds = self.w(inputs["token_type_ids"], mode="embedding")
+        if token_type_ids is not None:
+            token_type_ids = tf.reshape(token_type_ids, [-1, shape_list(token_type_ids)[-1]])
+            token_type_embeds = self.w(token_type_ids, mode="embedding")
             token_type_embeds *= tf.math.sqrt(tf.cast(self.d_model_size, dtype=token_type_embeds.dtype))
         else:
             token_type_embeds = tf.constant(0.0)
-        inputs["position_ids"] = tf.reshape(inputs["position_ids"], [-1, shape_list(inputs["position_ids"])[-1]])
+        position_ids = tf.reshape(position_ids, [-1, shape_list(position_ids)[-1]])
 
-        if inputs["inputs_embeds"] is None:
-            inputs["inputs_embeds"] = self.w(inputs["input_ids"], mode="embedding")
+        if inputs_embeds is None:
+            inputs_embeds = self.w(input_ids, mode="embedding")
         seq_len = input_shape[-1]
         mask = 1 - tf.linalg.band_part(tf.ones((seq_len, seq_len)), -1, 0)
 
-        inputs["inputs_embeds"] *= tf.math.sqrt(tf.cast(self.d_model_size, inputs["inputs_embeds"].dtype))
+        inputs_embeds *= tf.math.sqrt(tf.cast(self.d_model_size, inputs_embeds.dtype))
 
-        pos_embeds = tf.gather(self.pos_encoding, inputs["position_ids"])
+        pos_embeds = tf.gather(self.pos_encoding, position_ids)
         pos_embeds = tf.cast(pos_embeds, dtype=token_type_embeds.dtype)
-        hidden_states = inputs["inputs_embeds"] + pos_embeds + token_type_embeds
+        hidden_states = inputs_embeds + pos_embeds + token_type_embeds
 
-        hidden_states = self.dropout(hidden_states, training=inputs["training"])
+        hidden_states = self.dropout(hidden_states, training=training)
 
         output_shape = input_shape + [shape_list(hidden_states)[-1]]
-        presents = () if inputs["use_cache"] else None
-        all_hidden_states = () if inputs["output_hidden_states"] else None
-        all_attentions = () if inputs["output_attentions"] else None
-        for i, (h, layer_past) in enumerate(zip(self.h, inputs["past"])):
-            if inputs["output_hidden_states"]:
+        presents = () if use_cache else None
+        all_hidden_states = () if output_hidden_states else None
+        all_attentions = () if output_attentions else None
+        for i, (h, layer_past) in enumerate(zip(self.h, past_key_values)):
+            if output_hidden_states:
                 all_hidden_states = all_hidden_states + (tf.reshape(hidden_states, output_shape),)
             outputs = h(
                 hidden_states,
                 mask,
                 layer_past,
-                inputs["attention_mask"],
-                inputs["head_mask"][i],
-                inputs["use_cache"],
-                inputs["output_attentions"],
-                training=inputs["training"],
+                attention_mask,
+                head_mask[i],
+                use_cache,
+                output_attentions,
+                training=training,
             )
             hidden_states, present = outputs[:2]
 
-            if inputs["use_cache"]:
+            if use_cache:
                 presents = presents + (present,)
 
-            if inputs["output_attentions"]:
+            if output_attentions:
                 all_attentions = all_attentions + (outputs[2],)
 
         hidden_states = self.layernorm(hidden_states)
         hidden_states = tf.reshape(hidden_states, output_shape)
-        if inputs["output_hidden_states"]:
+        if output_hidden_states:
             all_hidden_states = all_hidden_states + (hidden_states,)
 
-        if inputs["output_attentions"]:
+        if output_attentions:
             # let the number of heads free (-1) so we can extract attention even after head pruning
             attention_output_shape = input_shape[:-1] + [-1] + shape_list(all_attentions[0])[-2:]
             all_attentions = tuple(tf.reshape(t, attention_output_shape) for t in all_attentions)
 
-        if not inputs["return_dict"]:
+        if not return_dict:
             return tuple(v for v in [hidden_states, presents, all_hidden_states, all_attentions] if v is not None)
 
         return TFBaseModelOutputWithPast(
@@ -540,6 +520,7 @@ class TFCTRLModel(TFCTRLPreTrainedModel):
         super().__init__(config, *inputs, **kwargs)
         self.transformer = TFCTRLMainLayer(config, name="transformer")
 
+    @unpack_inputs
     @add_start_docstrings_to_model_forward(CTRL_INPUTS_DOCSTRING)
     @add_code_sample_docstrings(
         processor_class=_TOKENIZER_FOR_DOC,
@@ -550,7 +531,7 @@ class TFCTRLModel(TFCTRLPreTrainedModel):
     def call(
         self,
         input_ids=None,
-        past=None,
+        past_key_values=None,
         attention_mask=None,
         token_type_ids=None,
         position_ids=None,
@@ -563,11 +544,9 @@ class TFCTRLModel(TFCTRLPreTrainedModel):
         training=False,
         **kwargs,
     ):
-        inputs = input_processing(
-            func=self.call,
-            config=self.config,
+        outputs = self.transformer(
             input_ids=input_ids,
-            past=past,
+            past_key_values=past_key_values,
             attention_mask=attention_mask,
             token_type_ids=token_type_ids,
             position_ids=position_ids,
@@ -578,21 +557,6 @@ class TFCTRLModel(TFCTRLPreTrainedModel):
             output_hidden_states=output_hidden_states,
             return_dict=return_dict,
             training=training,
-            kwargs_call=kwargs,
-        )
-        outputs = self.transformer(
-            input_ids=inputs["input_ids"],
-            past=inputs["past"],
-            attention_mask=inputs["attention_mask"],
-            token_type_ids=inputs["token_type_ids"],
-            position_ids=inputs["position_ids"],
-            head_mask=inputs["head_mask"],
-            inputs_embeds=inputs["inputs_embeds"],
-            use_cache=inputs["use_cache"],
-            output_attentions=inputs["output_attentions"],
-            output_hidden_states=inputs["output_hidden_states"],
-            return_dict=inputs["return_dict"],
-            training=inputs["training"],
         )
         return outputs
 
@@ -667,6 +631,7 @@ class TFCTRLLMHeadModel(TFCTRLPreTrainedModel, TFCausalLanguageModelingLoss):
 
         return {"input_ids": input_ids, "past_key_values": past, "use_cache": use_cache}
 
+    @unpack_inputs
     @add_start_docstrings_to_model_forward(CTRL_INPUTS_DOCSTRING)
     @add_code_sample_docstrings(
         processor_class=_TOKENIZER_FOR_DOC,
@@ -677,7 +642,7 @@ class TFCTRLLMHeadModel(TFCTRLPreTrainedModel, TFCausalLanguageModelingLoss):
     def call(
         self,
         input_ids=None,
-        past=None,
+        past_key_values=None,
         attention_mask=None,
         token_type_ids=None,
         position_ids=None,
@@ -696,11 +661,9 @@ class TFCTRLLMHeadModel(TFCTRLPreTrainedModel, TFCausalLanguageModelingLoss):
             Labels for computing the cross entropy classification loss. Indices should be in `[0, ...,
             config.vocab_size - 1]`.
         """
-        inputs = input_processing(
-            func=self.call,
-            config=self.config,
+        transformer_outputs = self.transformer(
             input_ids=input_ids,
-            past=past,
+            past_key_values=past_key_values,
             attention_mask=attention_mask,
             token_type_ids=token_type_ids,
             position_ids=position_ids,
@@ -710,23 +673,7 @@ class TFCTRLLMHeadModel(TFCTRLPreTrainedModel, TFCausalLanguageModelingLoss):
             output_attentions=output_attentions,
             output_hidden_states=output_hidden_states,
             return_dict=return_dict,
-            labels=labels,
             training=training,
-            kwargs_call=kwargs,
-        )
-        transformer_outputs = self.transformer(
-            input_ids=inputs["input_ids"],
-            past=inputs["past"],
-            attention_mask=inputs["attention_mask"],
-            token_type_ids=inputs["token_type_ids"],
-            position_ids=inputs["position_ids"],
-            head_mask=inputs["head_mask"],
-            inputs_embeds=inputs["inputs_embeds"],
-            use_cache=inputs["use_cache"],
-            output_attentions=inputs["output_attentions"],
-            output_hidden_states=inputs["output_hidden_states"],
-            return_dict=inputs["return_dict"],
-            training=inputs["training"],
         )
 
         hidden_states = transformer_outputs[0]
@@ -734,13 +681,13 @@ class TFCTRLLMHeadModel(TFCTRLPreTrainedModel, TFCausalLanguageModelingLoss):
         logits = self.lm_head(hidden_states)
 
         loss = None
-        if inputs["labels"] is not None:
+        if labels is not None:
             # shift labels to the left and cut last logit token
             shifted_logits = logits[:, :-1]
-            labels = inputs["labels"][:, 1:]
+            labels = labels[:, 1:]
             loss = self.hf_compute_loss(labels, shifted_logits)
 
-        if not inputs["return_dict"]:
+        if not return_dict:
             output = (logits,) + transformer_outputs[1:]
             return ((loss,) + output) if loss is not None else output
 
@@ -796,6 +743,7 @@ class TFCTRLForSequenceClassification(TFCTRLPreTrainedModel, TFSequenceClassific
     def get_output_embeddings(self):
         return self.transformer.w
 
+    @unpack_inputs
     @add_start_docstrings_to_model_forward(CTRL_INPUTS_DOCSTRING)
     @add_code_sample_docstrings(
         processor_class=_TOKENIZER_FOR_DOC,
@@ -806,7 +754,7 @@ class TFCTRLForSequenceClassification(TFCTRLPreTrainedModel, TFSequenceClassific
     def call(
         self,
         input_ids=None,
-        past=None,
+        past_key_values=None,
         attention_mask=None,
         token_type_ids=None,
         position_ids=None,
@@ -825,11 +773,10 @@ class TFCTRLForSequenceClassification(TFCTRLPreTrainedModel, TFSequenceClassific
             Labels for computing the cross entropy classification loss. Indices should be in `[0, ...,
             config.vocab_size - 1]`.
         """
-        inputs = input_processing(
-            func=self.call,
-            config=self.config,
+
+        transformer_outputs = self.transformer(
             input_ids=input_ids,
-            past=past,
+            past_key_values=past_key_values,
             attention_mask=attention_mask,
             token_type_ids=token_type_ids,
             position_ids=position_ids,
@@ -839,24 +786,7 @@ class TFCTRLForSequenceClassification(TFCTRLPreTrainedModel, TFSequenceClassific
             output_attentions=output_attentions,
             output_hidden_states=output_hidden_states,
             return_dict=return_dict,
-            labels=labels,
             training=training,
-            kwargs_call=kwargs,
-        )
-
-        transformer_outputs = self.transformer(
-            input_ids=inputs["input_ids"],
-            past=inputs["past"],
-            attention_mask=inputs["attention_mask"],
-            token_type_ids=inputs["token_type_ids"],
-            position_ids=inputs["position_ids"],
-            head_mask=inputs["head_mask"],
-            inputs_embeds=inputs["inputs_embeds"],
-            use_cache=inputs["use_cache"],
-            output_attentions=inputs["output_attentions"],
-            output_hidden_states=inputs["output_hidden_states"],
-            return_dict=inputs["return_dict"],
-            training=inputs["training"],
         )
 
         hidden_states = transformer_outputs[0]
@@ -865,12 +795,12 @@ class TFCTRLForSequenceClassification(TFCTRLPreTrainedModel, TFSequenceClassific
         if self.config.pad_token_id is None:
             sequence_lengths = -1
         else:
-            if inputs["input_ids"] is not None:
+            if input_ids is not None:
                 sequence_lengths = (
                     tf.reduce_sum(
                         tf.cast(
-                            tf.math.not_equal(inputs["input_ids"], self.config.pad_token_id),
-                            dtype=inputs["input_ids"].dtype,
+                            tf.math.not_equal(input_ids, self.config.pad_token_id),
+                            dtype=input_ids.dtype,
                         ),
                         -1,
                         keepdims=False,
@@ -886,11 +816,11 @@ class TFCTRLForSequenceClassification(TFCTRLPreTrainedModel, TFSequenceClassific
                 )
         loss = None
 
-        if inputs["labels"] is not None:
+        if labels is not None:
             if input_ids is not None:
-                batch_size, sequence_length = shape_list(inputs["input_ids"])[:2]
+                batch_size, sequence_length = shape_list(input_ids)[:2]
             else:
-                batch_size, sequence_length = shape_list(inputs["inputs_embeds"])[:2]
+                batch_size, sequence_length = shape_list(inputs_embeds)[:2]
             assert (
                 self.config.pad_token_id is not None or batch_size == 1
             ), "Cannot handle batch sizes > 1 if no padding token is defined."
@@ -898,13 +828,11 @@ class TFCTRLForSequenceClassification(TFCTRLPreTrainedModel, TFSequenceClassific
             if not tf.is_tensor(sequence_lengths):
                 in_logits = logits[0:batch_size, sequence_lengths]
 
-            loss = self.hf_compute_loss(
-                tf.reshape(inputs["labels"], [-1, 1]), tf.reshape(in_logits, [-1, self.num_labels])
-            )
+            loss = self.hf_compute_loss(tf.reshape(labels, [-1, 1]), tf.reshape(in_logits, [-1, self.num_labels]))
 
         pooled_logits = in_logits if in_logits is not None else logits
 
-        if not inputs["return_dict"]:
+        if not return_dict:
             output = (pooled_logits,) + transformer_outputs[1:]
             return ((loss,) + output) if loss is not None else output