transformers/templates/adding_a_new_model/cookiecutter-template-{{cookiecutter.modelname}}/modeling_tf_{{cookiecutter.lowercase_modelname}}.py

# coding=utf-8
# Copyright 2022 {{cookiecutter.authors}} and The HuggingFace Inc. team. All rights reserved.
#
# Licensed under the Apache License, Version 2.0 (the "License");
# you may not use this file except in compliance with the License.
# You may obtain a copy of the License at
#
#     http://www.apache.org/licenses/LICENSE-2.0
#
# Unless required by applicable law or agreed to in writing, software
# distributed under the License is distributed on an "AS IS" BASIS,
# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
# See the License for the specific language governing permissions and
# limitations under the License.
""" TF 2.0 {{cookiecutter.modelname}} model. """

{% if cookiecutter.is_encoder_decoder_model == "False" %}

import math
from typing import Dict, Optional, Tuple, Union

import numpy as np
import tensorflow as tf

from ...activations_tf import get_tf_activation
from ...utils import (
    DUMMY_INPUTS,
    MULTIPLE_CHOICE_DUMMY_INPUTS,
    add_code_sample_docstrings,
    add_start_docstrings,
    add_start_docstrings_to_model_forward,
)
from ...modeling_tf_outputs import (
    TFBaseModelOutputWithPastAndCrossAttentions,
    TFCausalLMOutputWithCrossAttentions,
    TFMaskedLMOutput,
    TFMultipleChoiceModelOutput,
    TFQuestionAnsweringModelOutput,
    TFSequenceClassifierOutput,
    TFTokenClassifierOutput,
)
from ...modeling_tf_utils import (
    TFCausalLanguageModelingLoss,
    TFMaskedLanguageModelingLoss,
    TFModelInputType,
    TFMultipleChoiceLoss,
    TFPreTrainedModel,
    TFQuestionAnsweringLoss,
    TFSequenceClassificationLoss,
    TFSequenceSummary,
    TFTokenClassificationLoss,
    get_initializer,
    keras_serializable,
    unpack_inputs,
)
from ...tf_utils import check_embeddings_within_bounds, shape_list, stable_softmax
from ...utils import logging
from .configuration_{{cookiecutter.lowercase_modelname}} import {{cookiecutter.camelcase_modelname}}Config


logger = logging.get_logger(__name__)

_CHECKPOINT_FOR_DOC = "{{cookiecutter.checkpoint_identifier}}"
_CONFIG_FOR_DOC = "{{cookiecutter.camelcase_modelname}}Config"

TF_{{cookiecutter.uppercase_modelname}}_PRETRAINED_MODEL_ARCHIVE_LIST = [
    "{{cookiecutter.checkpoint_identifier}}",
    # See all {{cookiecutter.modelname}} models at https://huggingface.co/models?filter={{cookiecutter.lowercase_modelname}}
]


# Copied from transformers.models.bert.modeling_tf_bert.TFBertEmbeddings with Bert->{{cookiecutter.camelcase_modelname}}
class TF{{cookiecutter.camelcase_modelname}}Embeddings(tf.keras.layers.Layer):
    """Construct the embeddings from word, position and token_type embeddings."""

    def __init__(self, config: {{cookiecutter.camelcase_modelname}}Config, **kwargs):
        super().__init__(**kwargs)

        self.vocab_size = config.vocab_size
        self.type_vocab_size = config.type_vocab_size
        self.hidden_size = config.hidden_size
        self.max_position_embeddings = config.max_position_embeddings
        self.initializer_range = config.initializer_range
        self.LayerNorm = tf.keras.layers.LayerNormalization(epsilon=config.layer_norm_eps, name="LayerNorm")
        self.dropout = tf.keras.layers.Dropout(rate=config.hidden_dropout_prob)

    def build(self, input_shape: tf.TensorShape):
        with tf.name_scope("word_embeddings"):
            self.weight = self.add_weight(
                name="weight",
                shape=[self.vocab_size, self.hidden_size],
                initializer=get_initializer(self.initializer_range),
            )

        with tf.name_scope("token_type_embeddings"):
            self.token_type_embeddings = self.add_weight(
                name="embeddings",
                shape=[self.type_vocab_size, self.hidden_size],
                initializer=get_initializer(self.initializer_range),
            )

        with tf.name_scope("position_embeddings"):
            self.position_embeddings = self.add_weight(
                name="embeddings",
                shape=[self.max_position_embeddings, self.hidden_size],
                initializer=get_initializer(self.initializer_range),
            )

        super().build(input_shape)

    def call(
        self,
        input_ids: tf.Tensor = None,
        position_ids: tf.Tensor = None,
        token_type_ids: tf.Tensor = None,
        inputs_embeds: tf.Tensor = None,
        past_key_values_length=0,
        training: bool = False,
    ) -> tf.Tensor:
        """
        Applies embedding based on inputs tensor.

        Returns:
            final_embeddings (`tf.Tensor`): output embedding tensor.
        """
        assert not (input_ids is None and inputs_embeds is None)

        if input_ids is not None:
            check_embeddings_within_bounds(input_ids, self.vocab_size)
            inputs_embeds = tf.gather(params=self.weight, indices=input_ids)

        input_shape = shape_list(inputs_embeds)[:-1]

        if token_type_ids is None:
            token_type_ids = tf.fill(dims=input_shape, value=0)

        if position_ids is None:
            position_ids = tf.expand_dims(
                tf.range(start=past_key_values_length, limit=input_shape[1] + past_key_values_length), axis=0
            )

        position_embeds = tf.gather(params=self.position_embeddings, indices=position_ids)
        token_type_embeds = tf.gather(params=self.token_type_embeddings, indices=token_type_ids)
        final_embeddings = inputs_embeds + position_embeds + token_type_embeds
        final_embeddings = self.LayerNorm(inputs=final_embeddings)
        final_embeddings = self.dropout(inputs=final_embeddings, training=training)

        return final_embeddings


# Copied from transformers.models.bert.modeling_tf_bert.TFBertSelfAttention with Bert->{{cookiecutter.camelcase_modelname}}
class TF{{cookiecutter.camelcase_modelname}}SelfAttention(tf.keras.layers.Layer):
    def __init__(self, config: {{cookiecutter.camelcase_modelname}}Config, **kwargs):
        super().__init__(**kwargs)

        if config.hidden_size % config.num_attention_heads != 0:
            raise ValueError(
                f"The hidden size ({config.hidden_size}) is not a multiple of the number "
                f"of attention heads ({config.num_attention_heads})"
            )

        self.num_attention_heads = config.num_attention_heads
        self.attention_head_size = int(config.hidden_size / config.num_attention_heads)
        self.all_head_size = self.num_attention_heads * self.attention_head_size
        self.sqrt_att_head_size = math.sqrt(self.attention_head_size)

        self.query = tf.keras.layers.Dense(
            units=self.all_head_size, kernel_initializer=get_initializer(config.initializer_range), name="query"
        )
        self.key = tf.keras.layers.Dense(
            units=self.all_head_size, kernel_initializer=get_initializer(config.initializer_range), name="key"
        )
        self.value = tf.keras.layers.Dense(
            units=self.all_head_size, kernel_initializer=get_initializer(config.initializer_range), name="value"
        )
        self.dropout = tf.keras.layers.Dropout(rate=config.attention_probs_dropout_prob)

        self.is_decoder = config.is_decoder

    def transpose_for_scores(self, tensor: tf.Tensor, batch_size: int) -> tf.Tensor:
        # Reshape from [batch_size, seq_length, all_head_size] to [batch_size, seq_length, num_attention_heads, attention_head_size]
        tensor = tf.reshape(tensor=tensor, shape=(batch_size, -1, self.num_attention_heads, self.attention_head_size))

        # Transpose the tensor from [batch_size, seq_length, num_attention_heads, attention_head_size] to [batch_size, num_attention_heads, seq_length, attention_head_size]
        return tf.transpose(tensor, perm=[0, 2, 1, 3])

    def call(
        self,
        hidden_states: tf.Tensor,
        attention_mask: tf.Tensor,
        head_mask: tf.Tensor,
        encoder_hidden_states: tf.Tensor,
        encoder_attention_mask: tf.Tensor,
        past_key_value: Tuple[tf.Tensor],
        output_attentions: bool,
        training: bool = False,
    ) -> Tuple[tf.Tensor]:
        batch_size = shape_list(hidden_states)[0]
        mixed_query_layer = self.query(inputs=hidden_states)

        # If this is instantiated as a cross-attention module, the keys
        # and values come from an encoder; the attention mask needs to be
        # such that the encoder's padding tokens are not attended to.
        is_cross_attention = encoder_hidden_states is not None

        if is_cross_attention and past_key_value is not None:
            # reuse k,v, cross_attentions
            key_layer = past_key_value[0]
            value_layer = past_key_value[1]
            attention_mask = encoder_attention_mask
        elif is_cross_attention:
            key_layer = self.transpose_for_scores(self.key(inputs=encoder_hidden_states), batch_size)
            value_layer = self.transpose_for_scores(self.value(inputs=encoder_hidden_states), batch_size)
            attention_mask = encoder_attention_mask
        elif past_key_value is not None:
            key_layer = self.transpose_for_scores(self.key(inputs=hidden_states), batch_size)
            value_layer = self.transpose_for_scores(self.value(inputs=hidden_states), batch_size)
            key_layer = tf.concat([past_key_value[0], key_layer], axis=2)
            value_layer = tf.concat([past_key_value[1], value_layer], axis=2)
        else:
            key_layer = self.transpose_for_scores(self.key(inputs=hidden_states), batch_size)
            value_layer = self.transpose_for_scores(self.value(inputs=hidden_states), batch_size)

        query_layer = self.transpose_for_scores(mixed_query_layer, batch_size)

        if self.is_decoder:
            # if cross_attention save Tuple(tf.Tensor, tf.Tensor) of all cross attention key/value_states.
            # Further calls to cross_attention layer can then reuse all cross-attention
            # key/value_states (first "if" case)
            # if uni-directional self-attention (decoder) save Tuple(tf.Tensor, tf.Tensor) of
            # all previous decoder key/value_states. Further calls to uni-directional self-attention
            # can concat previous decoder key/value_states to current projected key/value_states (third "elif" case)
            # if encoder bi-directional self-attention `past_key_value` is always `None`
            past_key_value = (key_layer, value_layer)

        # Take the dot product between "query" and "key" to get the raw attention scores.
        # (batch size, num_heads, seq_len_q, seq_len_k)
        attention_scores = tf.matmul(query_layer, key_layer, transpose_b=True)
        dk = tf.cast(self.sqrt_att_head_size, dtype=attention_scores.dtype)
        attention_scores = tf.divide(attention_scores, dk)

        if attention_mask is not None:
            # Apply the attention mask is (precomputed for all layers in TF{{cookiecutter.camelcase_modelname}}Model call() function)
            attention_scores = tf.add(attention_scores, attention_mask)

        # Normalize the attention scores to probabilities.
        attention_probs = stable_softmax(logits=attention_scores, axis=-1)

        # This is actually dropping out entire tokens to attend to, which might
        # seem a bit unusual, but is taken from the original Transformer paper.
        attention_probs = self.dropout(inputs=attention_probs, training=training)

        # Mask heads if we want to
        if head_mask is not None:
            attention_probs = tf.multiply(attention_probs, head_mask)

        attention_output = tf.matmul(attention_probs, value_layer)
        attention_output = tf.transpose(attention_output, perm=[0, 2, 1, 3])

        # (batch_size, seq_len_q, all_head_size)
        attention_output = tf.reshape(tensor=attention_output, shape=(batch_size, -1, self.all_head_size))
        outputs = (attention_output, attention_probs) if output_attentions else (attention_output,)

        if self.is_decoder:
            outputs = outputs + (past_key_value,)
        return outputs


# Copied from transformers.models.bert.modeling_tf_bert.TFBertSelfOutput with Bert->{{cookiecutter.camelcase_modelname}}
class TF{{cookiecutter.camelcase_modelname}}SelfOutput(tf.keras.layers.Layer):
    def __init__(self, config: {{cookiecutter.camelcase_modelname}}Config, **kwargs):
        super().__init__(**kwargs)

        self.dense = tf.keras.layers.Dense(
            units=config.hidden_size, kernel_initializer=get_initializer(config.initializer_range), name="dense"
        )
        self.LayerNorm = tf.keras.layers.LayerNormalization(epsilon=config.layer_norm_eps, name="LayerNorm")
        self.dropout = tf.keras.layers.Dropout(rate=config.hidden_dropout_prob)

    def call(self, hidden_states: tf.Tensor, input_tensor: tf.Tensor, training: bool = False) -> tf.Tensor:
        hidden_states = self.dense(inputs=hidden_states)
        hidden_states = self.dropout(inputs=hidden_states, training=training)
        hidden_states = self.LayerNorm(inputs=hidden_states + input_tensor)

        return hidden_states


# Copied from transformers.models.bert.modeling_tf_bert.TFBertAttention with Bert->{{cookiecutter.camelcase_modelname}}
class TF{{cookiecutter.camelcase_modelname}}Attention(tf.keras.layers.Layer):
    def __init__(self, config: {{cookiecutter.camelcase_modelname}}Config, **kwargs):
        super().__init__(**kwargs)

        self.self_attention = TF{{cookiecutter.camelcase_modelname}}SelfAttention(config, name="self")
        self.dense_output = TF{{cookiecutter.camelcase_modelname}}SelfOutput(config, name="output")

    def prune_heads(self, heads):
        raise NotImplementedError

    def call(
        self,
        input_tensor: tf.Tensor,
        attention_mask: tf.Tensor,
        head_mask: tf.Tensor,
        encoder_hidden_states: tf.Tensor,
        encoder_attention_mask: tf.Tensor,
        past_key_value: Tuple[tf.Tensor],
        output_attentions: bool,
        training: bool = False,
    ) -> Tuple[tf.Tensor]:
        self_outputs = self.self_attention(
            hidden_states=input_tensor,
            attention_mask=attention_mask,
            head_mask=head_mask,
            encoder_hidden_states=encoder_hidden_states,
            encoder_attention_mask=encoder_attention_mask,
            past_key_value=past_key_value,
            output_attentions=output_attentions,
            training=training,
        )
        attention_output = self.dense_output(
            hidden_states=self_outputs[0], input_tensor=input_tensor, training=training
        )
        # add attentions (possibly with past_key_value) if we output them
        outputs = (attention_output,) + self_outputs[1:]

        return outputs


# Copied from transformers.models.bert.modeling_tf_bert.TFBertIntermediate with Bert->{{cookiecutter.camelcase_modelname}}
class TF{{cookiecutter.camelcase_modelname}}Intermediate(tf.keras.layers.Layer):
    def __init__(self, config: {{cookiecutter.camelcase_modelname}}Config, **kwargs):
        super().__init__(**kwargs)

        self.dense = tf.keras.layers.Dense(
            units=config.intermediate_size, kernel_initializer=get_initializer(config.initializer_range), name="dense"
        )

        if isinstance(config.hidden_act, str):
            self.intermediate_act_fn = get_tf_activation(config.hidden_act)
        else:
            self.intermediate_act_fn = config.hidden_act

    def call(self, hidden_states: tf.Tensor) -> tf.Tensor:
        hidden_states = self.dense(inputs=hidden_states)
        hidden_states = self.intermediate_act_fn(hidden_states)

        return hidden_states


# Copied from transformers.models.bert.modeling_tf_bert.TFBertOutput with Bert->{{cookiecutter.camelcase_modelname}}
class TF{{cookiecutter.camelcase_modelname}}Output(tf.keras.layers.Layer):
    def __init__(self, config: {{cookiecutter.camelcase_modelname}}Config, **kwargs):
        super().__init__(**kwargs)

        self.dense = tf.keras.layers.Dense(
            units=config.hidden_size, kernel_initializer=get_initializer(config.initializer_range), name="dense"
        )
        self.LayerNorm = tf.keras.layers.LayerNormalization(epsilon=config.layer_norm_eps, name="LayerNorm")
        self.dropout = tf.keras.layers.Dropout(rate=config.hidden_dropout_prob)

    def call(self, hidden_states: tf.Tensor, input_tensor: tf.Tensor, training: bool = False) -> tf.Tensor:
        hidden_states = self.dense(inputs=hidden_states)
        hidden_states = self.dropout(inputs=hidden_states, training=training)
        hidden_states = self.LayerNorm(inputs=hidden_states + input_tensor)

        return hidden_states


# Copied from transformers.models.bert.modeling_tf_bert.TFBertLayer with Bert->{{cookiecutter.camelcase_modelname}}
class TF{{cookiecutter.camelcase_modelname}}Layer(tf.keras.layers.Layer):
    def __init__(self, config: {{cookiecutter.camelcase_modelname}}Config, **kwargs):
        super().__init__(**kwargs)

        self.attention = TF{{cookiecutter.camelcase_modelname}}Attention(config, name="attention")
        self.is_decoder = config.is_decoder
        self.add_cross_attention = config.add_cross_attention
        if self.add_cross_attention:
            if not self.is_decoder:
                raise ValueError(f"{self} should be used as a decoder model if cross attention is added")
            self.crossattention = TF{{cookiecutter.camelcase_modelname}}Attention(config, name="crossattention")
        self.intermediate = TF{{cookiecutter.camelcase_modelname}}Intermediate(config, name="intermediate")
        self.bert_output = TF{{cookiecutter.camelcase_modelname}}Output(config, name="output")

    def call(
        self,
        hidden_states: tf.Tensor,
        attention_mask: tf.Tensor,
        head_mask: tf.Tensor,
        encoder_hidden_states: tf.Tensor | None,
        encoder_attention_mask: tf.Tensor | None,
        past_key_value: Tuple[tf.Tensor] | None,
        output_attentions: bool,
        training: bool = False,
    ) -> Tuple[tf.Tensor]:
        # decoder uni-directional self-attention cached key/values tuple is at positions 1,2
        self_attn_past_key_value = past_key_value[:2] if past_key_value is not None else None
        self_attention_outputs = self.attention(
            input_tensor=hidden_states,
            attention_mask=attention_mask,
            head_mask=head_mask,
            encoder_hidden_states=None,
            encoder_attention_mask=None,
            past_key_value=self_attn_past_key_value,
            output_attentions=output_attentions,
            training=training,
        )
        attention_output = self_attention_outputs[0]

        # if decoder, the last output is tuple of self-attn cache
        if self.is_decoder:
            outputs = self_attention_outputs[1:-1]
            present_key_value = self_attention_outputs[-1]
        else:
            outputs = self_attention_outputs[1:]  # add self attentions if we output attention weights

        cross_attn_present_key_value = None
        if self.is_decoder and encoder_hidden_states is not None:
            if not hasattr(self, "crossattention"):
                raise ValueError(
                    f"If `encoder_hidden_states` are passed, {self} has to be instantiated with cross-attention layers "
                    "by setting `config.add_cross_attention=True`"
                )

            # cross_attn cached key/values tuple is at positions 3,4 of past_key_value tuple
            cross_attn_past_key_value = past_key_value[-2:] if past_key_value is not None else None
            cross_attention_outputs = self.crossattention(
                input_tensor=attention_output,
                attention_mask=attention_mask,
                head_mask=head_mask,
                encoder_hidden_states=encoder_hidden_states,
                encoder_attention_mask=encoder_attention_mask,
                past_key_value=cross_attn_past_key_value,
                output_attentions=output_attentions,
                training=training,
            )
            attention_output = cross_attention_outputs[0]
            outputs = outputs + cross_attention_outputs[1:-1]  # add cross attentions if we output attention weights

            # add cross-attn cache to positions 3,4 of present_key_value tuple
            cross_attn_present_key_value = cross_attention_outputs[-1]
            present_key_value = present_key_value + cross_attn_present_key_value

        intermediate_output = self.intermediate(hidden_states=attention_output)
        layer_output = self.bert_output(
            hidden_states=intermediate_output, input_tensor=attention_output, training=training
        )
        outputs = (layer_output,) + outputs  # add attentions if we output them

        # if decoder, return the attn key/values as the last output
        if self.is_decoder:
            outputs = outputs + (present_key_value,)

        return outputs


# Copied from transformers.models.bert.modeling_tf_bert.TFBertEncoder with Bert->{{cookiecutter.camelcase_modelname}}
class TF{{cookiecutter.camelcase_modelname}}Encoder(tf.keras.layers.Layer):
    def __init__(self, config: {{cookiecutter.camelcase_modelname}}Config, **kwargs):
        super().__init__(**kwargs)
        self.config = config
        self.layer = [TF{{cookiecutter.camelcase_modelname}}Layer(config, name=f"layer_._{i}") for i in range(config.num_hidden_layers)]

    def call(
        self,
        hidden_states: tf.Tensor,
        attention_mask: tf.Tensor,
        head_mask: tf.Tensor,
        encoder_hidden_states: tf.Tensor | None,
        encoder_attention_mask: tf.Tensor | None,
        past_key_values: Tuple[Tuple[tf.Tensor]] | None,
        use_cache: Optional[bool],
        output_attentions: bool,
        output_hidden_states: bool,
        return_dict: bool,
        training: bool = False,
    ) -> Union[TFBaseModelOutputWithPastAndCrossAttentions, Tuple[tf.Tensor]]:
        all_hidden_states = () if output_hidden_states else None
        all_attentions = () if output_attentions else None
        all_cross_attentions = () if output_attentions and self.config.add_cross_attention else None

        next_decoder_cache = () if use_cache else None
        for i, layer_module in enumerate(self.layer):
            if output_hidden_states:
                all_hidden_states = all_hidden_states + (hidden_states,)

            past_key_value = past_key_values[i] if past_key_values is not None else None

            layer_outputs = layer_module(
                hidden_states=hidden_states,
                attention_mask=attention_mask,
                head_mask=head_mask[i],
                encoder_hidden_states=encoder_hidden_states,
                encoder_attention_mask=encoder_attention_mask,
                past_key_value=past_key_value,
                output_attentions=output_attentions,
                training=training,
            )
            hidden_states = layer_outputs[0]

            if use_cache:
                next_decoder_cache += (layer_outputs[-1],)

            if output_attentions:
                all_attentions = all_attentions + (layer_outputs[1],)
                if self.config.add_cross_attention and encoder_hidden_states is not None:
                    all_cross_attentions = all_cross_attentions + (layer_outputs[2],)

        # Add last layer
        if output_hidden_states:
            all_hidden_states = all_hidden_states + (hidden_states,)

        if not return_dict:
            return tuple(
                v for v in [hidden_states, all_hidden_states, all_attentions, all_cross_attentions] if v is not None
            )

        return TFBaseModelOutputWithPastAndCrossAttentions(
            last_hidden_state=hidden_states,
            past_key_values=next_decoder_cache,
            hidden_states=all_hidden_states,
            attentions=all_attentions,
            cross_attentions=all_cross_attentions,
        )


# Copied from transformers.models.bert.modeling_tf_bert.TFBertPredictionHeadTransform with Bert->{{cookiecutter.camelcase_modelname}}
class TF{{cookiecutter.camelcase_modelname}}PredictionHeadTransform(tf.keras.layers.Layer):
    def __init__(self, config: {{cookiecutter.camelcase_modelname}}Config, **kwargs):
        super().__init__(**kwargs)

        self.dense = tf.keras.layers.Dense(
            units=config.hidden_size,
            kernel_initializer=get_initializer(config.initializer_range),
            name="dense",
        )

        if isinstance(config.hidden_act, str):
            self.transform_act_fn = get_tf_activation(config.hidden_act)
        else:
            self.transform_act_fn = config.hidden_act

        self.LayerNorm = tf.keras.layers.LayerNormalization(epsilon=config.layer_norm_eps, name="LayerNorm")

    def call(self, hidden_states: tf.Tensor) -> tf.Tensor:
        hidden_states = self.dense(inputs=hidden_states)
        hidden_states = self.transform_act_fn(hidden_states)
        hidden_states = self.LayerNorm(inputs=hidden_states)

        return hidden_states


# Copied from transformers.models.bert.modeling_tf_bert.TFBertLMPredictionHead with Bert->{{cookiecutter.camelcase_modelname}}
class TF{{cookiecutter.camelcase_modelname}}LMPredictionHead(tf.keras.layers.Layer):
    def __init__(self, config: {{cookiecutter.camelcase_modelname}}Config, input_embeddings: tf.keras.layers.Layer, **kwargs):
        super().__init__(**kwargs)

        self.vocab_size = config.vocab_size
        self.hidden_size = config.hidden_size

        self.transform = TF{{cookiecutter.camelcase_modelname}}PredictionHeadTransform(config, name="transform")

        # The output weights are the same as the input embeddings, but there is
        # an output-only bias for each token.
        self.input_embeddings = input_embeddings

    def build(self, input_shape: tf.TensorShape):
        self.bias = self.add_weight(shape=(self.vocab_size,), initializer="zeros", trainable=True, name="bias")

        super().build(input_shape)

    def get_output_embeddings(self) -> tf.keras.layers.Layer:
        return self.input_embeddings

    def set_output_embeddings(self, value: tf.Variable):
        self.input_embeddings.weight = value
        self.input_embeddings.vocab_size = shape_list(value)[0]

    def get_bias(self) -> Dict[str, tf.Variable]:
        return {"bias": self.bias}

    def set_bias(self, value: tf.Variable):
        self.bias = value["bias"]
        self.vocab_size = shape_list(value["bias"])[0]

    def call(self, hidden_states: tf.Tensor) -> tf.Tensor:
        hidden_states = self.transform(hidden_states=hidden_states)
        seq_length = shape_list(hidden_states)[1]
        hidden_states = tf.reshape(tensor=hidden_states, shape=[-1, self.hidden_size])
        hidden_states = tf.matmul(a=hidden_states, b=self.input_embeddings.weight, transpose_b=True)
        hidden_states = tf.reshape(tensor=hidden_states, shape=[-1, seq_length, self.vocab_size])
        hidden_states = tf.nn.bias_add(value=hidden_states, bias=self.bias)

        return hidden_states


# Copied from transformers.models.bert.modeling_tf_bert.TFBertMLMHead with Bert->{{cookiecutter.camelcase_modelname}}
class TF{{cookiecutter.camelcase_modelname}}MLMHead(tf.keras.layers.Layer):
    def __init__(self, config: {{cookiecutter.camelcase_modelname}}Config, input_embeddings: tf.keras.layers.Layer, **kwargs):
        super().__init__(**kwargs)

        self.predictions = TF{{cookiecutter.camelcase_modelname}}LMPredictionHead(config, input_embeddings, name="predictions")

    def call(self, sequence_output: tf.Tensor) -> tf.Tensor:
        prediction_scores = self.predictions(hidden_states=sequence_output)

        return prediction_scores


@keras_serializable
class TF{{cookiecutter.camelcase_modelname}}MainLayer(tf.keras.layers.Layer):
    config_class = {{cookiecutter.camelcase_modelname}}Config

    def __init__(self, config: {{cookiecutter.camelcase_modelname}}Config, add_pooling_layer: bool = True, **kwargs):
        super().__init__(**kwargs)

        self.config = config
        self.is_decoder = config.is_decoder

        self.embeddings = TF{{cookiecutter.camelcase_modelname}}Embeddings(config, name="embeddings")
        self.encoder = TF{{cookiecutter.camelcase_modelname}}Encoder(config, name="encoder")

    # Copied from transformers.models.bert.modeling_tf_bert.TFBertMainLayer.get_input_embeddings
    def get_input_embeddings(self) -> tf.keras.layers.Layer:
        return self.embeddings

    # Copied from transformers.models.bert.modeling_tf_bert.TFBertMainLayer.set_input_embeddings
    def set_input_embeddings(self, value: tf.Variable):
        self.embeddings.weight = value
        self.embeddings.vocab_size = shape_list(value)[0]

    # Copied from transformers.models.bert.modeling_tf_bert.TFBertMainLayer._prune_heads
    def _prune_heads(self, heads_to_prune):
        """
        Prunes heads of the model. heads_to_prune: dict of {layer_num: list of heads to prune in this layer} See base
        class PreTrainedModel
        """
        raise NotImplementedError

    @unpack_inputs
    def call(
        self,
        input_ids: TFModelInputType | None = None,
        attention_mask: np.ndarray | tf.Tensor | None = None,
        token_type_ids: np.ndarray | tf.Tensor | None = None,
        position_ids: np.ndarray | tf.Tensor | None = None,
        head_mask: np.ndarray | tf.Tensor | None = None,
        inputs_embeds: np.ndarray | tf.Tensor | None = None,
        encoder_hidden_states: np.ndarray | tf.Tensor | None = None,
        encoder_attention_mask: np.ndarray | tf.Tensor | None = None,
        past_key_values: Optional[Tuple[Tuple[Union[np.ndarray, tf.Tensor]]]] = None,
        use_cache: Optional[bool] = None,
        output_attentions: Optional[bool] = None,
        output_hidden_states: Optional[bool] = None,
        return_dict: Optional[bool] = None,
        training: bool = False,
    ) -> Union[TFBaseModelOutputWithPastAndCrossAttentions, Tuple[tf.Tensor]]:

        if not self.config.is_decoder:
            use_cache = False

        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 input_ids is not None:
            input_shape = shape_list(input_ids)
        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")

        batch_size, seq_length = input_shape

        if past_key_values is None:
            past_key_values_length = 0
            past_key_values = [None] * len(self.encoder.layer)
        else:
            past_key_values_length = shape_list(past_key_values[0][0])[-2]

        if attention_mask is None:
            attention_mask = tf.fill(dims=(batch_size, seq_length + past_key_values_length), value=1)

        if token_type_ids is None:
            token_type_ids = tf.fill(dims=input_shape, value=0)

        embedding_output = self.embeddings(
            input_ids=input_ids,
            position_ids=position_ids,
            token_type_ids=token_type_ids,
            inputs_embeds=inputs_embeds,
            past_key_values_length=past_key_values_length,
            training=training,
        )

        # 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.
        attention_mask_shape = shape_list(attention_mask)

        mask_seq_length = seq_length + past_key_values_length
        # Copied from `modeling_tf_t5.py`
        # Provided a padding mask of dimensions [batch_size, mask_seq_length]
        # - if the model is a decoder, apply a causal mask in addition to the padding mask
        # - if the model is an encoder, make the mask broadcastable to [batch_size, num_heads, mask_seq_length, mask_seq_length]
        if self.is_decoder:
            seq_ids = tf.range(mask_seq_length)
            causal_mask = tf.less_equal(
                tf.tile(seq_ids[None, None, :], (batch_size, mask_seq_length, 1)),
                seq_ids[None, :, None],
            )
            causal_mask = tf.cast(causal_mask, dtype=attention_mask.dtype)
            extended_attention_mask = causal_mask * attention_mask[:, None, :]
            attention_mask_shape = shape_list(extended_attention_mask)
            extended_attention_mask = tf.reshape(
                extended_attention_mask, (attention_mask_shape[0], 1, attention_mask_shape[1], attention_mask_shape[2])
            )
            if past_key_values[0] is not None:
                # attention_mask needs to be sliced to the shape `[batch_size, 1, from_seq_length - cached_seq_length, to_seq_length]
                extended_attention_mask = extended_attention_mask[:, :, -seq_length:, :]
        else:
            extended_attention_mask = tf.reshape(
                attention_mask, (attention_mask_shape[0], 1, 1, attention_mask_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
        # positions we want to attend and -10000.0 for masked positions.
        # Since we are adding it to the raw scores before the softmax, this is
        # effectively the same as removing these entirely.
        extended_attention_mask = tf.cast(extended_attention_mask, dtype=embedding_output.dtype)
        one_cst = tf.constant(1.0, dtype=embedding_output.dtype)
        ten_thousand_cst = tf.constant(-10000.0, dtype=embedding_output.dtype)
        extended_attention_mask = tf.multiply(tf.subtract(one_cst, extended_attention_mask), ten_thousand_cst)

        # Copied from `modeling_tf_t5.py` with -1e9 -> -10000
        if self.is_decoder and encoder_attention_mask is not None:
            # If a 2D ou 3D attention mask is provided for the cross-attention
            # we need to make broadcastable to [batch_size, num_heads, mask_seq_length, mask_seq_length]
            # we need to make broadcastable to [batch_size, num_heads, seq_length, seq_length]
            encoder_attention_mask = tf.cast(
                encoder_attention_mask, dtype=extended_attention_mask.dtype
            )
            num_dims_encoder_attention_mask = len(shape_list(encoder_attention_mask))
            if num_dims_encoder_attention_mask == 3:
                encoder_extended_attention_mask = encoder_attention_mask[:, None, :, :]
            if num_dims_encoder_attention_mask == 2:
                encoder_extended_attention_mask = encoder_attention_mask[:, None, None, :]

            # T5 has a mask that can compare sequence ids, we can simulate this here with this transposition
            # Cf. https://github.com/tensorflow/mesh/blob/8d2465e9bc93129b913b5ccc6a59aa97abd96ec6/mesh_tensorflow/transformer/transformer_layers.py#L270
            # encoder_extended_attention_mask = tf.math.equal(encoder_extended_attention_mask,
            #                                         tf.transpose(encoder_extended_attention_mask, perm=(-1, -2)))

            encoder_extended_attention_mask = (1.0 - encoder_extended_attention_mask) * -10000.0
        else:
            encoder_extended_attention_mask = None

        # 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
        # input head_mask has shape [num_heads] or [num_hidden_layers x num_heads]
        # and head_mask is converted to shape [num_hidden_layers x batch x num_heads x seq_length x seq_length]
        if head_mask is not None:
            raise NotImplementedError
        else:
            head_mask = [None] * self.config.num_hidden_layers

        encoder_outputs = self.encoder(
            hidden_states=embedding_output,
            attention_mask=extended_attention_mask,
            head_mask=head_mask,
            encoder_hidden_states=encoder_hidden_states,
            encoder_attention_mask=encoder_extended_attention_mask,
            past_key_values=past_key_values,
            use_cache=use_cache,
            output_attentions=output_attentions,
            output_hidden_states=output_hidden_states,
            return_dict=return_dict,
            training=training,
        )

        sequence_output = encoder_outputs[0]

        if not return_dict:
            return (
                sequence_output,
            ) + encoder_outputs[1:]

        return TFBaseModelOutputWithPastAndCrossAttentions(
            last_hidden_state=sequence_output,
            past_key_values=encoder_outputs.past_key_values,
            hidden_states=encoder_outputs.hidden_states,
            attentions=encoder_outputs.attentions,
            cross_attentions=encoder_outputs.cross_attentions,
        )


class TF{{cookiecutter.camelcase_modelname}}PreTrainedModel(TFPreTrainedModel):
    """An abstract class to handle weights initialization and
    a simple interface for downloading and loading pretrained models.
    """

    config_class = {{cookiecutter.camelcase_modelname}}Config
    base_model_prefix = "{{cookiecutter.lowercase_modelname}}"



{{cookiecutter.uppercase_modelname}}_START_DOCSTRING = r"""

    This model inherits from [`TFPreTrainedModel`]. Check the superclass documentation for the
    generic methods the library implements for all its model (such as downloading or saving, resizing the input
    embeddings, pruning heads etc.)

    This model is also a [tf.keras.Model](https://www.tensorflow.org/api_docs/python/tf/keras/Model) subclass.
    Use it as a regular TF 2.0 Keras Model and refer to the TF 2.0 documentation for all matter related to general
    usage and behavior.

    <Tip>

    TensorFlow models and layers in `transformers` accept two formats as input:

    - having all inputs as keyword arguments (like PyTorch models), or
    - having all inputs as a list, tuple or dict in the first positional argument.

    The reason the second format is supported is that Keras methods prefer this format when passing inputs to models
    and layers. Because of this support, when using methods like `model.fit()` things should "just work" for you - just
    pass your inputs and labels in any format that `model.fit()` supports! If, however, you want to use the second format outside of Keras methods like `fit()` and `predict()`, such as when creating
    your own layers or models with the Keras `Functional` API, there are three possibilities you
    can use to gather all the input Tensors in the first positional argument:

    - a single Tensor with `input_ids` only and nothing else: `model(input_ids)`
    - a list of varying length with one or several input Tensors IN THE ORDER given in the docstring:
    `model([input_ids, attention_mask])` or `model([input_ids, attention_mask, token_type_ids])`
    - a dictionary with one or several input Tensors associated to the input names given in the docstring:
    `model({"input_ids": input_ids, "token_type_ids": token_type_ids})`

    Note that when creating models and layers with (subclassing)[https://keras.io/guides/making_new_layers_and_models_via_subclassing/]
    then you don't need to worry about any of this, as you can just pass inputs like you would to any other Python
    function!

    </Tip>

    Args:
        config ([`~{{cookiecutter.camelcase_modelname}}Config`]): Model configuration class with all the parameters of the model.
            Initializing with a config file does not load the weights associated with the model, only the configuration.
            Check out the [`~PreTrainedModel.from_pretrained`] method to load the model weights.
"""

{{cookiecutter.uppercase_modelname}}_INPUTS_DOCSTRING = r"""
    Args:
        input_ids (`np.ndarray`, `tf.Tensor`, `List[tf.Tensor]`, `Dict[str, tf.Tensor]` or `Dict[str, np.ndarray]` and each example must have the shape `({0})`):
            Indices of input sequence tokens in the vocabulary.

            Indices can be obtained using [`AutoTokenizer`]. See
            [`PreTrainedTokenizer.__call__`] and [`PreTrainedTokenizer.encode`] for
            details.

            [What are input IDs?](../glossary#input-ids)
        attention_mask (`np.ndarray` or `tf.Tensor` of shape `({0})`, *optional*):
            Mask to avoid performing attention on padding token indices. Mask values selected in `[0, 1]`:

            - 1 for tokens that are **not masked**,
            - 0 for tokens that are **masked**.

            [What are attention masks?](../glossary#attention-mask)
        token_type_ids (`np.ndarray` or `tf.Tensor` of shape `({0})`, *optional*):
            Segment token indices to indicate first and second portions of the inputs. Indices are selected in `[0, 1]`:

            - 0 corresponds to a *sentence A* token,
            - 1 corresponds to a *sentence B* token.

            [What are token type IDs?](../glossary#token-type-ids)
        position_ids (`np.ndarray` or `tf.Tensor` of shape `({0})`, *optional*):
            Indices of positions of each input sequence tokens in the position embeddings. Selected in the range `[0, config.max_position_embeddings - 1]`.

            [What are position IDs?](../glossary#position-ids)
        head_mask (`np.ndarray` or `tf.Tensor` of shape `(num_heads,)` or `(num_layers, num_heads)`, *optional*):
            Mask to nullify selected heads of the self-attention modules. Mask values selected in `[0, 1]`:

            - 1 indicates the head is **not masked**,
            - 0 indicates the head is **masked**.

        inputs_embeds (`np.ndarray` or `tf.Tensor` of shape `({0}, hidden_size)`, *optional*):
            Optionally, instead of passing `input_ids` you can choose to directly pass an embedded representation.
            This is useful if you want more control over how to convert `input_ids` indices into associated
            vectors than the model's internal embedding lookup matrix.
        output_attentions (`bool`, *optional*):
            Whether or not to return the attentions tensors of all attention layers. See `attentions` under returned
            tensors for more detail. This argument can be used only in eager mode, in graph mode the value in the
            config will be used instead.
        output_hidden_states (`bool`, *optional*):
            Whether or not to return the hidden states of all layers. See `hidden_states` under returned tensors for
            more detail. This argument can be used only in eager mode, in graph mode the value in the config will be
            used instead.
        return_dict (`bool`, *optional*):
            Whether or not to return a [`~utils.ModelOutput`] instead of a plain tuple. This
            argument can be used in eager mode, in graph mode the value will always be set to True.
        training (`bool`, *optional*, defaults to `False`):
            Whether or not to use the model in training mode (some modules like dropout modules have different
            behaviors between training and evaluation).
"""


@add_start_docstrings(
    "The bare {{cookiecutter.modelname}} Model transformer outputing raw hidden-states without any specific head on top.",
    {{cookiecutter.uppercase_modelname}}_START_DOCSTRING,
)
class TF{{cookiecutter.camelcase_modelname}}Model(TF{{cookiecutter.camelcase_modelname}}PreTrainedModel):
    def __init__(self, config: {{cookiecutter.camelcase_modelname}}Config, *inputs, **kwargs):
        super().__init__(config, *inputs, **kwargs)

        self.{{cookiecutter.lowercase_modelname}} = TF{{cookiecutter.camelcase_modelname}}MainLayer(config, name="{{cookiecutter.lowercase_modelname}}")

    @unpack_inputs
    @add_start_docstrings_to_model_forward({{cookiecutter.uppercase_modelname}}_INPUTS_DOCSTRING.format("batch_size, sequence_length"))
    @add_code_sample_docstrings(
        checkpoint=_CHECKPOINT_FOR_DOC,
        output_type=TFBaseModelOutputWithPastAndCrossAttentions,
        config_class=_CONFIG_FOR_DOC,
    )
    def call(
        self,
        input_ids: TFModelInputType | None = None,
        attention_mask: np.ndarray | tf.Tensor | None = None,
        token_type_ids: np.ndarray | tf.Tensor | None = None,
        position_ids: np.ndarray | tf.Tensor | None = None,
        head_mask: np.ndarray | tf.Tensor | None = None,
        inputs_embeds: np.ndarray | tf.Tensor | None = None,
        encoder_hidden_states: np.ndarray | tf.Tensor | None = None,
        encoder_attention_mask: np.ndarray | tf.Tensor | None = None,
        past_key_values: Optional[Tuple[Tuple[Union[np.ndarray, tf.Tensor]]]] = None,
        use_cache: Optional[bool] = None,
        output_attentions: Optional[bool] = None,
        output_hidden_states: Optional[bool] = None,
        return_dict: Optional[bool] = None,
        training: Optional[bool] = False,
    ) -> Union[TFBaseModelOutputWithPastAndCrossAttentions, Tuple[tf.Tensor]]:
        r"""
        encoder_hidden_states  (`tf.Tensor` of shape `(batch_size, sequence_length, hidden_size)`, *optional*):
            Sequence of hidden-states at the output of the last layer of the encoder. Used in the cross-attention if
            the model is configured as a decoder.
        encoder_attention_mask (`tf.Tensor` of shape `(batch_size, sequence_length)`, *optional*):
            Mask to avoid performing attention on the padding token indices of the encoder input. This mask is used in
            the cross-attention if the model is configured as a decoder. Mask values selected in `[0, 1]`:

            - 1 for tokens that are **not masked**,
            - 0 for tokens that are **masked**.

        past_key_values (`Tuple[Tuple[tf.Tensor]]` of length `config.n_layers`)
            contains precomputed key and value hidden states of the attention blocks. Can be used to speed up decoding.
            If `past_key_values` are used, the user can optionally input only the last `decoder_input_ids`
            (those that don't have their past key value states given to this model) of shape `(batch_size, 1)`
            instead of all `decoder_input_ids` of shape `(batch_size, sequence_length)`.
        use_cache (`bool`, *optional*, defaults to `True`):
            If set to `True`, `past_key_values` key value states are returned and can be used to speed up
            decoding (see `past_key_values`). Set to `False` during training, `True` during generation
        """
        outputs = self.{{cookiecutter.lowercase_modelname}}(
            input_ids=input_ids,
            attention_mask=attention_mask,
            token_type_ids=token_type_ids,
            position_ids=position_ids,
            head_mask=head_mask,
            inputs_embeds=inputs_embeds,
            encoder_hidden_states=encoder_hidden_states,
            encoder_attention_mask=encoder_attention_mask,
            past_key_values=past_key_values,
            use_cache=use_cache,
            output_attentions=output_attentions,
            output_hidden_states=output_hidden_states,
            return_dict=return_dict,
            training=training,
        )

        return outputs



@add_start_docstrings("""{{cookiecutter.modelname}} Model with a `language modeling` head on top. """, {{cookiecutter.uppercase_modelname}}_START_DOCSTRING)
class TF{{cookiecutter.camelcase_modelname}}ForMaskedLM(TF{{cookiecutter.camelcase_modelname}}PreTrainedModel, TFMaskedLanguageModelingLoss):

    def __init__(self, config: {{cookiecutter.camelcase_modelname}}Config, *inputs, **kwargs):
        super().__init__(config, *inputs, **kwargs)

        if config.is_decoder:
            logger.warning(
                "If you want to use `TF{{cookiecutter.camelcase_modelname}}ForMaskedLM` make sure `config.is_decoder=False` for "
                "bi-directional self-attention."
            )

        self.{{cookiecutter.lowercase_modelname}} = TF{{cookiecutter.camelcase_modelname}}MainLayer(config, name="{{cookiecutter.lowercase_modelname}}")
        self.mlm = TF{{cookiecutter.camelcase_modelname}}MLMHead(config, input_embeddings=self.{{cookiecutter.lowercase_modelname}}.embeddings, name="mlm___cls")

    def get_lm_head(self) -> tf.keras.layers.Layer:
        return self.mlm.predictions

    @unpack_inputs
    @add_start_docstrings_to_model_forward({{cookiecutter.uppercase_modelname}}_INPUTS_DOCSTRING.format("batch_size, sequence_length"))
    @add_code_sample_docstrings(
        checkpoint=_CHECKPOINT_FOR_DOC,
        output_type=TFMaskedLMOutput,
        config_class=_CONFIG_FOR_DOC,
    )
    def call(
        self,
        input_ids: TFModelInputType | None = None,
        attention_mask: np.ndarray | tf.Tensor | None = None,
        token_type_ids: np.ndarray | tf.Tensor | None = None,
        position_ids: np.ndarray | tf.Tensor | None = None,
        head_mask: np.ndarray | tf.Tensor | None = None,
        inputs_embeds: np.ndarray | tf.Tensor | None = None,
        output_attentions: Optional[bool] = None,
        output_hidden_states: Optional[bool] = None,
        return_dict: Optional[bool] = None,
        labels: np.ndarray | tf.Tensor | None = None,
        training: Optional[bool] = False,
    ) -> Union[TFMaskedLMOutput, Tuple[tf.Tensor]]:
        r"""
        labels (`tf.Tensor` or `np.ndarray` of shape `(batch_size, sequence_length)`, *optional*):
            Labels for computing the masked language modeling loss. Indices should be in `[-100, 0, ..., config.vocab_size]` (see `input_ids` docstring) Tokens with indices set to `-100` are ignored
            (masked), the loss is only computed for the tokens with labels in `[0, ..., config.vocab_size]`
        """
        outputs = self.{{cookiecutter.lowercase_modelname}}(
            input_ids=input_ids,
            attention_mask=attention_mask,
            token_type_ids=token_type_ids,
            position_ids=position_ids,
            head_mask=head_mask,
            inputs_embeds=inputs_embeds,
            output_attentions=output_attentions,
            output_hidden_states=output_hidden_states,
            return_dict=return_dict,
            training=training,
        )
        sequence_output = outputs[0]
        prediction_scores = self.mlm(sequence_output=sequence_output, training=training)
        loss = (
            None if labels is None else self.hf_compute_loss(labels=labels, logits=prediction_scores)
        )

        if not return_dict:
            output = (prediction_scores,) + outputs[2:]
            return ((loss,) + output) if loss is not None else output

        return TFMaskedLMOutput(
            loss=loss,
            logits=prediction_scores,
            hidden_states=outputs.hidden_states,
            attentions=outputs.attentions,
        )


@add_start_docstrings(
    """{{cookiecutter.modelname}} Model with a `language modeling` head on top for CLM fine-tuning. """, {{cookiecutter.uppercase_modelname}}_START_DOCSTRING
)
class TF{{cookiecutter.camelcase_modelname}}ForCausalLM(TF{{cookiecutter.camelcase_modelname}}PreTrainedModel, TFCausalLanguageModelingLoss):

    def __init__(self, config: {{cookiecutter.camelcase_modelname}}Config, *inputs, **kwargs):
        super().__init__(config, *inputs, **kwargs)

        if not config.is_decoder:
            logger.warning("If you want to use `TF{{cookiecutter.camelcase_modelname}}ForCausalLM` as a standalone, add `is_decoder=True.`")

        self.{{cookiecutter.lowercase_modelname}} = TF{{cookiecutter.camelcase_modelname}}MainLayer(config, name="{{cookiecutter.lowercase_modelname}}")
        self.mlm = TF{{cookiecutter.camelcase_modelname}}MLMHead(config, input_embeddings=self.{{cookiecutter.lowercase_modelname}}.embeddings, name="mlm___cls")

    def get_lm_head(self) -> tf.keras.layers.Layer:
        return self.mlm.predictions

    def prepare_inputs_for_generation(self, inputs, past_key_values=None, attention_mask=None, **model_kwargs):
        # cut decoder_input_ids if past is used
        if past_key_values:
            inputs = tf.expand_dims(inputs[:, -1], -1)

        return {
            "input_ids": inputs,
            "attention_mask": attention_mask,
            "past_key_values": past_key_values,
            "use_cache": model_kwargs["use_cache"],
        }

    @unpack_inputs
    @add_code_sample_docstrings(
        checkpoint=_CHECKPOINT_FOR_DOC,
        output_type=TFCausalLMOutputWithCrossAttentions,
        config_class=_CONFIG_FOR_DOC,
    )
    def call(
        self,
        input_ids: TFModelInputType | None = None,
        attention_mask: np.ndarray | tf.Tensor | None = None,
        token_type_ids: np.ndarray | tf.Tensor | None = None,
        position_ids: np.ndarray | tf.Tensor | None = None,
        head_mask: np.ndarray | tf.Tensor | None = None,
        inputs_embeds: np.ndarray | tf.Tensor | None = None,
        encoder_hidden_states: np.ndarray | tf.Tensor | None = None,
        encoder_attention_mask: np.ndarray | tf.Tensor | None = None,
        past_key_values: Optional[Tuple[Tuple[Union[np.ndarray, tf.Tensor]]]] = None,
        use_cache: Optional[bool] = None,
        output_attentions: Optional[bool] = None,
        output_hidden_states: Optional[bool] = None,
        return_dict: Optional[bool] = None,
        labels: np.ndarray | tf.Tensor | None = None,
        training: Optional[bool] = False,
    ) -> Union[TFCausalLMOutputWithCrossAttentions, Tuple[tf.Tensor]]:
        r"""
        encoder_hidden_states  (`tf.Tensor` of shape `(batch_size, sequence_length, hidden_size)`, *optional*):
            Sequence of hidden-states at the output of the last layer of the encoder. Used in the cross-attention if
            the model is configured as a decoder.
        encoder_attention_mask (`tf.Tensor` of shape `(batch_size, sequence_length)`, *optional*):
            Mask to avoid performing attention on the padding token indices of the encoder input. This mask is used in
            the cross-attention if the model is configured as a decoder. Mask values selected in `[0, 1]`:

            - 1 for tokens that are **not masked**,
            - 0 for tokens that are **masked**.

        past_key_values (`Tuple[Tuple[tf.Tensor]]` of length `config.n_layers`)
            contains precomputed key and value hidden states of the attention blocks. Can be used to speed up decoding.
            If `past_key_values` are used, the user can optionally input only the last `decoder_input_ids`
            (those that don't have their past key value states given to this model) of shape `(batch_size, 1)`
            instead of all `decoder_input_ids` of shape `(batch_size, sequence_length)`.
        use_cache (`bool`, *optional*, defaults to `True`):
            If set to `True`, `past_key_values` key value states are returned and can be used to speed up
            decoding (see `past_key_values`). Set to `False` during training, `True` during generation
        labels (`tf.Tensor` or `np.ndarray` of shape `(batch_size, sequence_length)`, *optional*):
            Labels for computing the cross entropy classification loss. Indices should be in `[0, ..., config.vocab_size - 1]`.
        """
        outputs = self.{{cookiecutter.lowercase_modelname}}(
            input_ids=input_ids,
            attention_mask=attention_mask,
            token_type_ids=token_type_ids,
            position_ids=position_ids,
            head_mask=head_mask,
            inputs_embeds=inputs_embeds,
            encoder_hidden_states=encoder_hidden_states,
            encoder_attention_mask=encoder_attention_mask,
            past_key_values=past_key_values,
            use_cache=use_cache,
            output_attentions=output_attentions,
            output_hidden_states=output_hidden_states,
            return_dict=return_dict,
            training=training,
        )
        sequence_output = outputs[0]
        logits = self.mlm(sequence_output=sequence_output, training=training)
        loss = None

        if labels is not None:
            # shift labels to the left and cut last logit token
            shifted_logits = logits[:, :-1]
            labels = labels[:, 1:]
            loss = self.hf_compute_loss(labels=labels, logits=shifted_logits)

        if not return_dict:
            output = (logits,) + outputs[2:]
            return ((loss,) + output) if loss is not None else output

        return TFCausalLMOutputWithCrossAttentions(
            loss=loss,
            logits=logits,
            past_key_values=outputs.past_key_values,
            hidden_states=outputs.hidden_states,
            attentions=outputs.attentions,
            cross_attentions=outputs.cross_attentions,
        )



class TF{{cookiecutter.camelcase_modelname}}ClassificationHead(tf.keras.layers.Layer):
    """Head for sentence-level classification tasks."""

    def __init__(self, config: {{cookiecutter.camelcase_modelname}}Config, *inputs, **kwargs):
        super().__init__(config, *inputs, **kwargs)

        self.dense = tf.keras.layers.Dense(
            units=config.hidden_size, kernel_initializer=get_initializer(config.initializer_range), name="dense"
        )
        self.dropout = tf.keras.layers.Dropout(rate=config.hidden_dropout_prob)
        self.out_proj = tf.keras.layers.Dense(
            units=config.num_labels, kernel_initializer=get_initializer(config.initializer_range), name="out_proj"
        )

        if isinstance(config.hidden_act, str):
            self.classifier_act_fn = get_tf_activation(config.hidden_act)
        else:
            self.classifier_act_fn = config.hidden_act

    def call(self, hidden_states: tf.Tensor, training: bool = False) -> tf.Tensor:
        hidden_states = hidden_states[:, 0, :]  # take <s> token (equiv. to [CLS])
        hidden_states = self.dropout(inputs=hidden_states, training=training)
        hidden_states = self.dense(inputs=hidden_states)
        hidden_states = self.classifier_act_fn(hidden_states)
        hidden_states = self.dropout(inputs=hidden_states, training=training)
        hidden_states = self.out_proj(hidden_states)

        return hidden_states


@add_start_docstrings(
    """{{cookiecutter.modelname}} Model transformer with a sequence classification/regression head on top
    e.g., for GLUE tasks. """,
    {{cookiecutter.uppercase_modelname}}_START_DOCSTRING,
)
class TF{{cookiecutter.camelcase_modelname}}ForSequenceClassification(TF{{cookiecutter.camelcase_modelname}}PreTrainedModel, TFSequenceClassificationLoss):
    def __init__(self, config: {{cookiecutter.camelcase_modelname}}Config, *inputs, **kwargs):
        super().__init__(config, *inputs, **kwargs)

        self.num_labels = config.num_labels

        self.{{cookiecutter.lowercase_modelname}} = TF{{cookiecutter.camelcase_modelname}}MainLayer(config, name="{{cookiecutter.lowercase_modelname}}")
        self.classifier = TF{{cookiecutter.camelcase_modelname}}ClassificationHead(config, name="classifier")

    @unpack_inputs
    @add_start_docstrings_to_model_forward({{cookiecutter.uppercase_modelname}}_INPUTS_DOCSTRING.format("batch_size, sequence_length"))
    @add_code_sample_docstrings(
        checkpoint=_CHECKPOINT_FOR_DOC,
        output_type=TFSequenceClassifierOutput,
        config_class=_CONFIG_FOR_DOC,
    )
    def call(
        self,
        input_ids: TFModelInputType | None = None,
        attention_mask: np.ndarray | tf.Tensor | None = None,
        token_type_ids: np.ndarray | tf.Tensor | None = None,
        position_ids: np.ndarray | tf.Tensor | None = None,
        head_mask: np.ndarray | tf.Tensor | None = None,
        inputs_embeds: np.ndarray | tf.Tensor | None = None,
        output_attentions: Optional[bool] = None,
        output_hidden_states: Optional[bool] = None,
        return_dict: Optional[bool] = None,
        labels: np.ndarray | tf.Tensor | None = None,
        training: Optional[bool] = False,
    ) -> Union[TFSequenceClassifierOutput, Tuple[tf.Tensor]]:
        r"""
        labels (`tf.Tensor` or `np.ndarray` of shape `(batch_size,)`, *optional*):
            Labels for computing the sequence classification/regression loss. Indices should be in `[0, ..., config.num_labels - 1]`. If `config.num_labels == 1` a regression loss is computed (Mean-Square loss),
            If `config.num_labels > 1` a classification loss is computed (Cross-Entropy).
        """
        outputs = self.{{cookiecutter.lowercase_modelname}}(
            input_ids=input_ids,
            attention_mask=attention_mask,
            token_type_ids=token_type_ids,
            position_ids=position_ids,
            head_mask=head_mask,
            inputs_embeds=inputs_embeds,
            output_attentions=output_attentions,
            output_hidden_states=output_hidden_states,
            return_dict=return_dict,
            training=training,
        )
        logits = self.classifier(hidden_states=outputs[0], training=training)
        loss = None if labels is None else self.hf_compute_loss(labels=labels, logits=logits)

        if not return_dict:
            output = (logits,) + outputs[1:]

            return ((loss,) + output) if loss is not None else output

        return TFSequenceClassifierOutput(
            loss=loss,
            logits=logits,
            hidden_states=outputs.hidden_states,
            attentions=outputs.attentions,
        )


@add_start_docstrings(
    """{{cookiecutter.modelname}} Model with a multiple choice classification head on top (a linear layer on top of
    the pooled output and a softmax) e.g. for RocStories/SWAG tasks. """,
    {{cookiecutter.uppercase_modelname}}_START_DOCSTRING,
)
class TF{{cookiecutter.camelcase_modelname}}ForMultipleChoice(TF{{cookiecutter.camelcase_modelname}}PreTrainedModel, TFMultipleChoiceLoss):
    def __init__(self, config: {{cookiecutter.camelcase_modelname}}Config, *inputs, **kwargs):
        super().__init__(config, *inputs, **kwargs)

        self.{{cookiecutter.lowercase_modelname}} = TF{{cookiecutter.camelcase_modelname}}MainLayer(config, name="{{cookiecutter.lowercase_modelname}}")
        self.sequence_summary = TFSequenceSummary(
            config, config.initializer_range, name="sequence_summary"
        )
        self.classifier = tf.keras.layers.Dense(
            units=1, kernel_initializer=get_initializer(config.initializer_range), name="classifier"
        )

    @unpack_inputs
    @add_start_docstrings_to_model_forward({{cookiecutter.uppercase_modelname}}_INPUTS_DOCSTRING.format("batch_size, num_choices, sequence_length"))
    @add_code_sample_docstrings(
        checkpoint=_CHECKPOINT_FOR_DOC,
        output_type=TFMultipleChoiceModelOutput,
        config_class=_CONFIG_FOR_DOC,
    )
    def call(
        self,
        input_ids: TFModelInputType | None = None,
        attention_mask: np.ndarray | tf.Tensor | None = None,
        token_type_ids: np.ndarray | tf.Tensor | None = None,
        position_ids: np.ndarray | tf.Tensor | None = None,
        head_mask: np.ndarray | tf.Tensor | None = None,
        inputs_embeds: np.ndarray | tf.Tensor | None = None,
        output_attentions: Optional[bool] = None,
        output_hidden_states: Optional[bool] = None,
        return_dict: Optional[bool] = None,
        labels: np.ndarray | tf.Tensor | None = None,
        training: Optional[bool] = False,
    ) -> Union[TFMultipleChoiceModelOutput, Tuple[tf.Tensor]]:
        r"""
        labels (`tf.Tensor` or `np.ndarray` of shape `(batch_size,)`, *optional*):
            Labels for computing the multiple choice classification loss. Indices should be in `[0, ..., num_choices]` where `num_choices` is the size of the second dimension of the input tensors. (See
            `input_ids` above)
        """

        if input_ids is not None:
            num_choices = shape_list(input_ids)[1]
            seq_length = shape_list(input_ids)[2]
        else:
            num_choices = shape_list(inputs_embeds)[1]
            seq_length = shape_list(inputs_embeds)[2]

        flat_input_ids = (
            tf.reshape(tensor=input_ids, shape=(-1, seq_length)) if input_ids is not None else None
        )
        flat_attention_mask = (
            tf.reshape(tensor=attention_mask, shape=(-1, seq_length))
            if attention_mask is not None
            else None
        )
        flat_token_type_ids = (
            tf.reshape(tensor=token_type_ids, shape=(-1, seq_length))
            if token_type_ids is not None
            else None
        )
        flat_position_ids = (
            tf.reshape(tensor=position_ids, shape=(-1, seq_length))
            if position_ids is not None
            else None
        )
        flat_inputs_embeds = (
            tf.reshape(
                tensor=inputs_embeds, shape=(-1, seq_length, shape_list(inputs_embeds)[3])
            )
            if inputs_embeds is not None
            else None
        )
        outputs = self.{{cookiecutter.lowercase_modelname}}(
            input_ids=flat_input_ids,
            attention_mask=flat_attention_mask,
            token_type_ids=flat_token_type_ids,
            position_ids=flat_position_ids,
            head_mask=head_mask,
            inputs_embeds=flat_inputs_embeds,
            output_attentions=output_attentions,
            output_hidden_states=output_hidden_states,
            return_dict=return_dict,
            training=training,
        )
        logits = self.sequence_summary(inputs=outputs[0], training=training)
        logits = self.classifier(inputs=logits)
        reshaped_logits = tf.reshape(tensor=logits, shape=(-1, num_choices))
        loss = None if labels is None else self.hf_compute_loss(labels=labels, logits=reshaped_logits)

        if not return_dict:
            output = (reshaped_logits,) + outputs[1:]

            return ((loss,) + output) if loss is not None else output

        return TFMultipleChoiceModelOutput(
            loss=loss,
            logits=reshaped_logits,
            hidden_states=outputs.hidden_states,
            attentions=outputs.attentions,
        )


@add_start_docstrings(
    """{{cookiecutter.modelname}} Model with a token classification head on top (a linear layer on top of
    the hidden-states output) e.g. for Named-Entity-Recognition (NER) tasks. """,
    {{cookiecutter.uppercase_modelname}}_START_DOCSTRING,
)
class TF{{cookiecutter.camelcase_modelname}}ForTokenClassification(TF{{cookiecutter.camelcase_modelname}}PreTrainedModel, TFTokenClassificationLoss):

    def __init__(self, config: {{cookiecutter.camelcase_modelname}}Config, *inputs, **kwargs):
        super().__init__(config, *inputs, **kwargs)

        self.num_labels = config.num_labels

        self.{{cookiecutter.lowercase_modelname}} = TF{{cookiecutter.camelcase_modelname}}MainLayer(config, name="{{cookiecutter.lowercase_modelname}}")
        self.dropout = tf.keras.layers.Dropout(rate=config.hidden_dropout_prob)
        self.classifier = tf.keras.layers.Dense(
            units=config.num_labels, kernel_initializer=get_initializer(config.initializer_range), name="classifier"
        )

    @unpack_inputs
    @add_start_docstrings_to_model_forward({{cookiecutter.uppercase_modelname}}_INPUTS_DOCSTRING.format("batch_size, sequence_length"))
    @add_code_sample_docstrings(
        checkpoint=_CHECKPOINT_FOR_DOC,
        output_type=TFTokenClassifierOutput,
        config_class=_CONFIG_FOR_DOC,
    )
    def call(
        self,
        input_ids: TFModelInputType | None = None,
        attention_mask: np.ndarray | tf.Tensor | None = None,
        token_type_ids: np.ndarray | tf.Tensor | None = None,
        position_ids: np.ndarray | tf.Tensor | None = None,
        head_mask: np.ndarray | tf.Tensor | None = None,
        inputs_embeds: np.ndarray | tf.Tensor | None = None,
        output_attentions: Optional[bool] = None,
        output_hidden_states: Optional[bool] = None,
        return_dict: Optional[bool] = None,
        labels: np.ndarray | tf.Tensor | None = None,
        training: Optional[bool] = False,
    ) -> Union[TFTokenClassifierOutput, Tuple[tf.Tensor]]:
        r"""
        labels (`tf.Tensor` or `np.ndarray` of shape `(batch_size, sequence_length)`, *optional*):
            Labels for computing the token classification loss. Indices should be in `[0, ..., config.num_labels - 1]`.
        """

        outputs = self.{{cookiecutter.lowercase_modelname}}(
            input_ids=input_ids,
            attention_mask=attention_mask,
            token_type_ids=token_type_ids,
            position_ids=position_ids,
            head_mask=head_mask,
            inputs_embeds=inputs_embeds,
            output_attentions=output_attentions,
            output_hidden_states=output_hidden_states,
            return_dict=return_dict,
            training=training,
        )
        sequence_output = outputs[0]
        sequence_output = self.dropout(inputs=sequence_output, training=training)
        logits = self.classifier(inputs=sequence_output)
        loss = None if labels is None else self.hf_compute_loss(labels=labels, logits=logits)

        if not return_dict:
            output = (logits,) + outputs[1:]
            return ((loss,) + output) if loss is not None else output

        return TFTokenClassifierOutput(
            loss=loss,
            logits=logits,
            hidden_states=outputs.hidden_states,
            attentions=outputs.attentions,
        )


@add_start_docstrings(
    """{{cookiecutter.modelname}} Model with a span classification head on top for extractive question-answering tasks like SQuAD (a linear
    layer on top of the hidden-states output to compute `span start logits` and `span end logits`). """,
    {{cookiecutter.uppercase_modelname}}_START_DOCSTRING,
)
class TF{{cookiecutter.camelcase_modelname}}ForQuestionAnswering(TF{{cookiecutter.camelcase_modelname}}PreTrainedModel, TFQuestionAnsweringLoss):

    def __init__(self, config: {{cookiecutter.camelcase_modelname}}Config, *inputs, **kwargs):
        super().__init__(config, *inputs, **kwargs)

        self.num_labels = config.num_labels

        self.{{cookiecutter.lowercase_modelname}} = TF{{cookiecutter.camelcase_modelname}}MainLayer(config, name="{{cookiecutter.lowercase_modelname}}")
        self.qa_outputs = tf.keras.layers.Dense(
            units=config.num_labels, kernel_initializer=get_initializer(config.initializer_range), name="qa_outputs"
        )

    @unpack_inputs
    @add_start_docstrings_to_model_forward({{cookiecutter.uppercase_modelname}}_INPUTS_DOCSTRING.format("batch_size, sequence_length"))
    @add_code_sample_docstrings(
        checkpoint=_CHECKPOINT_FOR_DOC,
        output_type=TFQuestionAnsweringModelOutput,
        config_class=_CONFIG_FOR_DOC,
    )
    def call(
        self,
        input_ids: TFModelInputType | None = None,
        attention_mask: np.ndarray | tf.Tensor | None = None,
        token_type_ids: np.ndarray | tf.Tensor | None = None,
        position_ids: np.ndarray | tf.Tensor | None = None,
        head_mask: np.ndarray | tf.Tensor | None = None,
        inputs_embeds: np.ndarray | tf.Tensor | None = None,
        output_attentions: Optional[bool] = None,
        output_hidden_states: Optional[bool] = None,
        return_dict: Optional[bool] = None,
        start_positions: np.ndarray | tf.Tensor | None = None,
        end_positions: np.ndarray | tf.Tensor | None = None,
        training: Optional[bool] = False,
    ) -> Union[TFQuestionAnsweringModelOutput, Tuple[tf.Tensor]]:
        r"""
        start_positions (`tf.Tensor` or `np.ndarray` of shape `(batch_size,)`, *optional*):
            Labels for position (index) of the start of the labelled span for computing the token classification loss.
            Positions are clamped to the length of the sequence (`sequence_length`). Position outside of the
            sequence are not taken into account for computing the loss.
        end_positions (`tf.Tensor` or `np.ndarray` of shape `(batch_size,)`, *optional*):
            Labels for position (index) of the end of the labelled span for computing the token classification loss.
            Positions are clamped to the length of the sequence (`sequence_length`). Position outside of the
            sequence are not taken into account for computing the loss.
        """
        outputs = self.{{cookiecutter.lowercase_modelname}}(
            input_ids=input_ids,
            attention_mask=attention_mask,
            token_type_ids=token_type_ids,
            position_ids=position_ids,
            head_mask=head_mask,
            inputs_embeds=inputs_embeds,
            output_attentions=output_attentions,
            output_hidden_states=output_hidden_states,
            return_dict=return_dict,
            training=training,
        )
        sequence_output = outputs[0]
        logits = self.qa_outputs(inputs=sequence_output)
        start_logits, end_logits = tf.split(value=logits, num_or_size_splits=2, axis=-1)
        start_logits = tf.squeeze(input=start_logits, axis=-1)
        end_logits = tf.squeeze(input=end_logits, axis=-1)
        loss = None

        if start_positions is not None and end_positions is not None:
            labels = {"start_position": start_positions}
            labels["end_position"] = end_positions
            loss = self.hf_compute_loss(labels=labels, logits=(start_logits, end_logits))

        if not return_dict:
            output = (start_logits, end_logits) + outputs[2:]
            return ((loss,) + output) if loss is not None else output

        return TFQuestionAnsweringModelOutput(
            loss=loss,
            start_logits=start_logits,
            end_logits=end_logits,
            hidden_states=outputs.hidden_states,
            attentions=outputs.attentions,
        )


{% else %}
import random
from typing import Optional, Tuple, Union

import tensorflow as tf

from ...activations_tf import get_tf_activation
from ...utils import (
    add_code_sample_docstrings,
    add_start_docstrings,
    add_start_docstrings_to_model_forward,
    replace_return_docstrings,
)
from ...modeling_tf_outputs import (
    TFBaseModelOutput,
    TFBaseModelOutputWithPastAndCrossAttentions,
    TFSeq2SeqLMOutput,
    TFSeq2SeqModelOutput,
)

# Public API
from ...modeling_tf_utils import (
    DUMMY_INPUTS,
    TFPreTrainedModel,
    keras_serializable,
    unpack_inputs,
)
from ...tf_utils import check_embeddings_within_bounds, shape_list, stable_softmax
from ...utils import ContextManagers, logging
from .configuration_{{cookiecutter.lowercase_modelname}} import {{cookiecutter.camelcase_modelname}}Config


logger = logging.get_logger(__name__)

_CHECKPOINT_FOR_DOC = "{{cookiecutter.checkpoint_identifier}}"
_CONFIG_FOR_DOC = "{{cookiecutter.camelcase_modelname}}Config"
_TOKENIZER_FOR_DOC = "{{cookiecutter.camelcase_modelname}}Tokenizer"


LARGE_NEGATIVE = -1e8


# Copied from transformers.models.bart.modeling_tf_bart.shift_tokens_right
def shift_tokens_right(input_ids: tf.Tensor, pad_token_id: int, decoder_start_token_id: int):
    pad_token_id = tf.cast(pad_token_id, input_ids.dtype)
    decoder_start_token_id = tf.cast(decoder_start_token_id, input_ids.dtype)
    start_tokens = tf.fill((shape_list(input_ids)[0], 1), tf.convert_to_tensor(decoder_start_token_id, input_ids.dtype))
    shifted_input_ids = tf.concat([start_tokens, input_ids[:, :-1]], -1)
    # replace possible -100 values in labels by `pad_token_id`
    shifted_input_ids = tf.where(
        shifted_input_ids == -100,
        tf.fill(shape_list(shifted_input_ids), tf.convert_to_tensor(pad_token_id, input_ids.dtype)),
        shifted_input_ids,
    )

    # "Verify that `labels` has only positive values and -100"
    assert_gte0 = tf.debugging.assert_greater_equal(shifted_input_ids, tf.constant(0, dtype=shifted_input_ids.dtype))

    # Make sure the assertion op is called by wrapping the result in an identity no-op
    with tf.control_dependencies([assert_gte0]):
        shifted_input_ids = tf.identity(shifted_input_ids)

    return shifted_input_ids


def _make_causal_mask(input_ids_shape: tf.TensorShape, past_key_values_length: int = 0):
    """
    Make causal mask used for bi-directional self-attention.
    """
    bsz, tgt_len = input_ids_shape
    mask = tf.ones((tgt_len, tgt_len)) * LARGE_NEGATIVE
    mask_cond = tf.range(shape_list(mask)[-1])

    mask = tf.where(mask_cond < tf.reshape(mask_cond + 1, (shape_list(mask)[-1], 1)), 0.0, mask)

    if past_key_values_length > 0:
        mask = tf.concat([tf.zeros((tgt_len, past_key_values_length)), mask], axis=-1)

    return tf.tile(mask[None, None, :, :], (bsz, 1, 1, 1))


def _expand_mask(mask: tf.Tensor, tgt_len: Optional[int] = None):
    """
    Expands attention_mask from `[bsz, seq_len]` to `[bsz, 1, tgt_seq_len, src_seq_len]`.
    """
    src_len = shape_list(mask)[1]
    tgt_len = tgt_len if tgt_len is not None else src_len
    one_cst = tf.constant(1.0)
    mask = tf.cast(mask, dtype=one_cst.dtype)
    expanded_mask = tf.tile(mask[:, None, None, :], (1, 1, tgt_len, 1))

    return (one_cst - expanded_mask) * LARGE_NEGATIVE


class TF{{cookiecutter.camelcase_modelname}}LearnedPositionalEmbedding(tf.keras.layers.Embedding):
    """
    This module learns positional embeddings up to a fixed maximum size.
    """

    def __init__(self, num_embeddings: int, embedding_dim: int, **kwargs):
        super().__init__(num_embeddings, embedding_dim, **kwargs)

    def call(self, input_shape: tf.TensorShape, past_key_values_length: int = 0):
        """Input is expected to be of size [bsz x seqlen]."""
        seq_len = input_shape[1]
        position_ids = tf.range(seq_len, delta=1, name="range")
        position_ids += past_key_values_length
        return super().call(tf.cast(position_ids, dtype=tf.int32))


class TF{{cookiecutter.camelcase_modelname}}Attention(tf.keras.layers.Layer):
    """Multi-headed attention from "Attention Is All You Need"""

    def __init__(
        self,
        embed_dim: int,
        num_heads: int,
        dropout: float = 0.0,
        is_decoder: bool = False,
        bias: bool = True,
        **kwargs,
    ):
        super().__init__(**kwargs)
        self.embed_dim = embed_dim

        self.num_heads = num_heads
        self.dropout = tf.keras.layers.Dropout(dropout)
        self.head_dim = embed_dim // num_heads
        assert self.head_dim * num_heads == self.embed_dim, "embed_dim must be divisible by num_heads"
        self.scaling = self.head_dim ** -0.5
        self.is_decoder = is_decoder

        self.k_proj = tf.keras.layers.Dense(embed_dim, use_bias=bias, name="k_proj")
        self.q_proj = tf.keras.layers.Dense(embed_dim, use_bias=bias, name="q_proj")
        self.v_proj = tf.keras.layers.Dense(embed_dim, use_bias=bias, name="v_proj")
        self.out_proj = tf.keras.layers.Dense(embed_dim, use_bias=bias, name="out_proj")

    def _shape(self, tensor: tf.Tensor, seq_len: int, bsz: int):
        return tf.transpose(tf.reshape(tensor, (bsz, seq_len, self.num_heads, self.head_dim)), (0, 2, 1, 3))

    def call(
        self,
        hidden_states: tf.Tensor,
        key_value_states: tf.Tensor | None = None,
        past_key_value: Tuple[Tuple[tf.Tensor]] | None = None,
        attention_mask: tf.Tensor | None = None,
        layer_head_mask: tf.Tensor | None = None,
        training=False,
    ) -> Tuple[tf.Tensor, tf.Tensor | None]:
        """Input shape: Batch x Time x Channel"""

        # if key_value_states are provided this layer is used as a cross-attention layer
        # for the decoder
        is_cross_attention = key_value_states is not None
        bsz, tgt_len, embed_dim = shape_list(hidden_states)

        # get query proj
        query_states = self.q_proj(hidden_states) * self.scaling
        # get key, value proj
        if is_cross_attention and past_key_value is not None:
            # reuse k,v, cross_attentions
            key_states = past_key_value[0]
            value_states = past_key_value[1]
        elif is_cross_attention:
            # cross_attentions
            key_states = self._shape(self.k_proj(key_value_states), -1, bsz)
            value_states = self._shape(self.v_proj(key_value_states), -1, bsz)
        elif past_key_value is not None:
            # reuse k, v, self_attention
            key_states = self._shape(self.k_proj(hidden_states), -1, bsz)
            value_states = self._shape(self.v_proj(hidden_states), -1, bsz)
            key_states = tf.concat([past_key_value[0], key_states], axis=2)
            value_states = tf.concat([past_key_value[1], value_states], axis=2)
        else:
            # self_attention
            key_states = self._shape(self.k_proj(hidden_states), -1, bsz)
            value_states = self._shape(self.v_proj(hidden_states), -1, bsz)

        if self.is_decoder:
            # if cross_attention save Tuple(tf.Tensor, tf.Tensor) of all cross attention key/value_states.
            # Further calls to cross_attention layer can then reuse all cross-attention
            # key/value_states (first "if" case)
            # if uni-directional self-attention (decoder) save Tuple(tf.Tensor, tf.Tensor) of
            # all previous decoder key/value_states. Further calls to uni-directional self-attention
            # can concat previous decoder key/value_states to current projected key/value_states (third "elif" case)
            # if encoder bi-directional self-attention `past_key_value` is always `None`
            past_key_value = (key_states, value_states)

        proj_shape = (bsz * self.num_heads, -1, self.head_dim)
        query_states = tf.reshape(self._shape(query_states, tgt_len, bsz), proj_shape)
        key_states = tf.reshape(key_states, proj_shape)
        value_states = tf.reshape(value_states, proj_shape)

        src_len = shape_list(key_states)[1]
        attn_weights = tf.matmul(query_states, key_states, transpose_b=True)

        tf.debugging.assert_equal(
            shape_list(attn_weights),
            [bsz * self.num_heads, tgt_len, src_len],
            message=f"Attention weights should be of size {(bsz * self.num_heads, tgt_len, src_len)}, but is {shape_list(attn_weights)}",
        )

        if attention_mask is not None:
            tf.debugging.assert_equal(
                shape_list(attention_mask),
                [bsz, 1, tgt_len, src_len],
                message=f"Attention mask should be of size {(bsz, 1, tgt_len, src_len)}, but is {shape_list(attention_mask)}",
            )

            attn_weights = tf.reshape(attn_weights, (bsz, self.num_heads, tgt_len, src_len)) + attention_mask
            attn_weights = tf.reshape(attn_weights, (bsz * self.num_heads, tgt_len, src_len))

        attn_weights = stable_softmax(attn_weights, axis=-1)

        if layer_head_mask is not None:
            tf.debugging.assert_equal(
                shape_list(layer_head_mask),
                [self.num_heads],
                message=f"Head mask for a single layer should be of size {(self.num_heads)}, but is {shape_list(layer_head_mask)}",
            )

            attn_weights = tf.reshape(layer_head_mask, (1, -1, 1, 1)) * tf.reshape(
                attn_weights, (bsz, self.num_heads, tgt_len, src_len)
            )
            attn_weights = tf.reshape(attn_weights, (bsz * self.num_heads, tgt_len, src_len))

        attn_probs = self.dropout(attn_weights, training=training)

        attn_output = tf.matmul(attn_probs, value_states)

        tf.debugging.assert_equal(
            shape_list(attn_output),
            [bsz * self.num_heads, tgt_len, self.head_dim],
            message=f"`attn_output` should be of size {(bsz, self.num_heads, tgt_len, self.head_dim)}, but is {shape_list(attn_output)}",
        )

        attn_output = tf.transpose(
            tf.reshape(attn_output, (bsz, self.num_heads, tgt_len, self.head_dim)), (0, 2, 1, 3)
        )
        attn_output = tf.reshape(attn_output, (bsz, tgt_len, embed_dim))

        attn_output = self.out_proj(attn_output)
        attn_weights: tf.Tensor = tf.reshape(attn_weights, (bsz, self.num_heads, tgt_len, src_len))

        return attn_output, attn_weights, past_key_value


class TF{{cookiecutter.camelcase_modelname}}EncoderLayer(tf.keras.layers.Layer):
    def __init__(self, config: {{cookiecutter.camelcase_modelname}}Config, **kwargs):
        super().__init__(**kwargs)
        self.embed_dim = config.d_model
        self.self_attn = TF{{cookiecutter.camelcase_modelname}}Attention(
            self.embed_dim, config.encoder_attention_heads, dropout=config.attention_dropout, name="self_attn"
        )
        self.self_attn_layer_norm = tf.keras.layers.LayerNormalization(epsilon=1e-5, name="self_attn_layer_norm")
        self.dropout = tf.keras.layers.Dropout(config.dropout)
        self.activation_fn = get_tf_activation(config.activation_function)
        self.activation_dropout = tf.keras.layers.Dropout(config.activation_dropout)
        self.fc1 = tf.keras.layers.Dense(config.encoder_ffn_dim, name="fc1")
        self.fc2 = tf.keras.layers.Dense(self.embed_dim, name="fc2")
        self.final_layer_norm = tf.keras.layers.LayerNormalization(epsilon=1e-5, name="final_layer_norm")

    def call(self, hidden_states: tf.Tensor, attention_mask: tf.Tensor, layer_head_mask: tf.Tensor, training=False):
        """
        Args:
            hidden_states (`tf.Tensor`): input to the layer of shape *(batch, seq_len, embed_dim)*
            attention_mask (`tf.Tensor`): attention mask of size
                *(batch, 1, tgt_len, src_len)* where padding elements are indicated by very large negative values.
            layer_head_mask (`tf.Tensor`): mask for attention heads in a given layer of size
                *(encoder_attention_heads,)*
        """
        residual = hidden_states
        hidden_states, self_attn_weights, _ = self.self_attn(
            hidden_states=hidden_states, attention_mask=attention_mask, layer_head_mask=layer_head_mask
        )

        tf.debugging.assert_equal(
            shape_list(hidden_states),
            shape_list(residual),
            message=f"Self attn modified the shape of query {shape_list(residual)} to {shape_list(hidden_states)}",
        )

        hidden_states = self.dropout(hidden_states, training=training)
        hidden_states = residual + hidden_states
        hidden_states = self.self_attn_layer_norm(hidden_states)

        residual = hidden_states
        hidden_states = self.activation_fn(self.fc1(hidden_states))
        hidden_states = self.activation_dropout(hidden_states, training=training)
        hidden_states = self.fc2(hidden_states)
        hidden_states = self.dropout(hidden_states, training=training)
        hidden_states = residual + hidden_states
        hidden_states = self.final_layer_norm(hidden_states)

        return hidden_states, self_attn_weights


class TF{{cookiecutter.camelcase_modelname}}DecoderLayer(tf.keras.layers.Layer):
    def __init__(self, config: {{cookiecutter.camelcase_modelname}}Config, **kwargs):
        super().__init__(**kwargs)
        self.embed_dim = config.d_model
        self.self_attn = TF{{cookiecutter.camelcase_modelname}}Attention(
            embed_dim=self.embed_dim,
            num_heads=config.decoder_attention_heads,
            dropout=config.attention_dropout,
            name="self_attn",
            is_decoder=True,
        )
        self.dropout = tf.keras.layers.Dropout(config.dropout)
        self.activation_fn = get_tf_activation(config.activation_function)
        self.activation_dropout = tf.keras.layers.Dropout(config.activation_dropout)

        self.self_attn_layer_norm = tf.keras.layers.LayerNormalization(epsilon=1e-5, name="self_attn_layer_norm")
        self.encoder_attn = TF{{cookiecutter.camelcase_modelname}}Attention(
            self.embed_dim,
            config.decoder_attention_heads,
            dropout=config.attention_dropout,
            name="encoder_attn",
            is_decoder=True,
        )
        self.encoder_attn_layer_norm = tf.keras.layers.LayerNormalization(epsilon=1e-5, name="encoder_attn_layer_norm")
        self.fc1 = tf.keras.layers.Dense(config.decoder_ffn_dim, name="fc1")
        self.fc2 = tf.keras.layers.Dense(self.embed_dim, name="fc2")
        self.final_layer_norm = tf.keras.layers.LayerNormalization(epsilon=1e-5, name="final_layer_norm")

    def call(
        self,
        hidden_states,
        attention_mask: tf.Tensor | None = None,
        encoder_hidden_states: tf.Tensor | None = None,
        encoder_attention_mask: tf.Tensor | None = None,
        layer_head_mask: tf.Tensor | None = None,
        cross_attn_layer_head_mask: tf.Tensor | None = None,
        past_key_value: Tuple[tf.Tensor] | None = None,
        training=False,
    ) -> Tuple[tf.Tensor, tf.Tensor, Tuple[Tuple[tf.Tensor]]]:
        """
        Args:
            hidden_states (`tf.Tensor`): input to the layer of shape *(batch, seq_len, embed_dim)*
            attention_mask (`tf.Tensor`): attention mask of size
                *(batch, 1, tgt_len, src_len)* where padding elements are indicated by very large negative values.
            encoder_hidden_states (`tf.Tensor`): cross attention input to the layer of shape *(batch, seq_len, embed_dim)*
            encoder_attention_mask (`tf.Tensor`): encoder attention mask of size
                *(batch, 1, tgt_len, src_len)* where padding elements are indicated by very large negative values.
            layer_head_mask (`tf.Tensor`): mask for attention heads in a given layer of size
                *(decoder_attention_heads,)*
            cross_attn_layer_head_mask (`tf.Tensor`): mask for heads of the cross-attention module.
                *(decoder_attention_heads,)*
            past_key_value (`Tuple(tf.Tensor)`): cached past key and value projection states
        """
        residual = hidden_states

        # Self Attention
        # decoder uni-directional self-attention cached key/values tuple is at positions 1,2
        self_attn_past_key_value = past_key_value[:2] if past_key_value is not None else None
        # add present self-attn cache to positions 1,2 of present_key_value tuple
        hidden_states, self_attn_weights, present_key_value = self.self_attn(
            hidden_states=hidden_states,
            past_key_value=self_attn_past_key_value,
            attention_mask=attention_mask,
            layer_head_mask=layer_head_mask,
        )
        hidden_states = self.dropout(hidden_states, training=training)
        hidden_states = residual + hidden_states
        hidden_states = self.self_attn_layer_norm(hidden_states)

        # Cross-Attention Block
        cross_attn_present_key_value = None
        cross_attn_weights = None
        if encoder_hidden_states is not None:
            residual = hidden_states

            # cross_attn cached key/values tuple is at positions 3,4 of present_key_value tuple
            cross_attn_past_key_value = past_key_value[-2:] if past_key_value is not None else None
            hidden_states, cross_attn_weights, cross_attn_present_key_value = self.encoder_attn(
                hidden_states=hidden_states,
                key_value_states=encoder_hidden_states,
                attention_mask=encoder_attention_mask,
                layer_head_mask=cross_attn_layer_head_mask,
                past_key_value=cross_attn_past_key_value,
            )
            hidden_states = self.dropout(hidden_states, training=training)
            hidden_states = residual + hidden_states
            hidden_states = self.encoder_attn_layer_norm(hidden_states)

            # add cross-attn to positions 3,4 of present_key_value tuple
            present_key_value = present_key_value + cross_attn_present_key_value

        # Fully Connected
        residual = hidden_states
        hidden_states = self.activation_fn(self.fc1(hidden_states))
        hidden_states = self.activation_dropout(hidden_states, training=training)
        hidden_states = self.fc2(hidden_states)
        hidden_states = self.dropout(hidden_states, training=training)
        hidden_states = residual + hidden_states
        hidden_states = self.final_layer_norm(hidden_states)

        return (
            hidden_states,
            self_attn_weights,
            cross_attn_weights,
            present_key_value,
        )


class TF{{cookiecutter.camelcase_modelname}}PreTrainedModel(TFPreTrainedModel):
    config_class = {{cookiecutter.camelcase_modelname}}Config
    base_model_prefix = "model"


{{cookiecutter.uppercase_modelname}}_START_DOCSTRING = r"""
    This model inherits from [`TFPreTrainedModel`]. Check the superclass documentation for the
    generic methods the library implements for all its model (such as downloading or saving, resizing the input
    embeddings, pruning heads etc.)

    This model is also a [tf.keras.Model](https://www.tensorflow.org/api_docs/python/tf/keras/Model) subclass. Use
    it as a regular TF 2.0 Keras Model and refer to the TF 2.0 documentation for all matter related to general usage
    and behavior.

    <Tip>

    TensorFlow models and layers in `transformers` accept two formats as input:

    - having all inputs as keyword arguments (like PyTorch models), or
    - having all inputs as a list, tuple or dict in the first positional argument.

    The reason the second format is supported is that Keras methods prefer this format when passing inputs to models
    and layers. Because of this support, when using methods like `model.fit()` things should "just work" for you - just
    pass your inputs and labels in any format that `model.fit()` supports! If, however, you want to use the second format outside of Keras methods like `fit()` and `predict()`, such as when creating
    your own layers or models with the Keras `Functional` API, there are three possibilities you
    can use to gather all the input Tensors in the first positional argument:

    - a single Tensor with `input_ids` only and nothing else: `model(input_ids)`
    - a list of varying length with one or several input Tensors IN THE ORDER given in the docstring:
    `model([input_ids, attention_mask])` or `model([input_ids, attention_mask, token_type_ids])`
    - a dictionary with one or several input Tensors associated to the input names given in the docstring:
    `model({"input_ids": input_ids, "token_type_ids": token_type_ids})`

    Note that when creating models and layers with (subclassing)[https://keras.io/guides/making_new_layers_and_models_via_subclassing/]
    then you don't need to worry about any of this, as you can just pass inputs like you would to any other Python
    function!

    </Tip>

    Args:
        config ([`~{{cookiecutter.camelcase_modelname}}Config`]): Model configuration class with all the parameters of the model.
            Initializing with a config file does not load the weights associated with the model, only the
            configuration. Check out the [`~TFPreTrainedModel.from_pretrained`] method to load the
            model weights.
"""

{{cookiecutter.uppercase_modelname}}_INPUTS_DOCSTRING = r"""
    Args:
        input_ids (`tf.Tensor` of shape `({0})`):
            Indices of input sequence tokens in the vocabulary.

            Indices can be obtained using [`~{{cookiecutter.camelcase_modelname}}Tokenizer`]. See
            [`PreTrainedTokenizer.encode`] and [`PreTrainedTokenizer.__call__`] for
            details.

            [What are input IDs?](../glossary#input-ids)
        attention_mask (`tf.Tensor` of shape `({0})`, *optional*):
            Mask to avoid performing attention on padding token indices. Mask values selected in `[0, 1]`:

            - 1 for tokens that are **not masked**,
            - 0 for tokens that are **masked**.

            [What are attention masks?](../glossary#attention-mask)
        decoder_input_ids (`tf.Tensor` of shape `(batch_size, target_sequence_length)`, *optional*):
            Indices of decoder input sequence tokens in the vocabulary.

            Indices can be obtained using [`~{{cookiecutter.camelcase_modelname}}Tokenizer`]. See
            [`PreTrainedTokenizer.encode`] and [`PreTrainedTokenizer.__call__`] for
            details.

            [What are input IDs?](../glossary#input-ids)

            {{cookiecutter.camelcase_modelname}} uses the `eos_token_id` as the starting token for
            `decoder_input_ids` generation. If `past_key_values` is used, optionally only the last
            `decoder_input_ids` have to be input (see `past_key_values`).

            For translation and summarization training, `decoder_input_ids` should be provided. If no
            `decoder_input_ids` is provided, the model will create this tensor by shifting the `input_ids` to
            the right for denoising pre-training following the paper.
        decoder_attention_mask (`tf.Tensor` of shape `(batch_size, target_sequence_length)`, *optional*):
            will be made by default and ignore pad tokens. It is not recommended to set this for most use cases.
        head_mask (`tf.Tensor` of shape `(encoder_layers, encoder_attention_heads)`, *optional*):
            Mask to nullify selected heads of the attention modules in the encoder. Mask values selected in `[0, 1]`:

            - 1 indicates the head is **not masked**,
            - 0 indicates the head is **masked**.

        decoder_head_mask (`tf.Tensor` of shape `(decoder_layers, decoder_attention_heads)`, *optional*):
            Mask to nullify selected heads of the attention modules in the decoder. Mask values selected in `[0, 1]`:

            - 1 indicates the head is **not masked**,
            - 0 indicates the head is **masked**.

        cross_attn_head_mask (`tf.Tensor` of shape `(decoder_layers, decoder_attention_heads)`, *optional*):
            Mask to nullify selected heads of the cross-attention modules. Mask values selected in `[0, 1]`:

            - 1 indicates the head is **not masked**,
            - 0 indicates the head is **masked**.

        encoder_outputs (`tf.FloatTensor`, *optional*):
            hidden states at the output of the last layer of the encoder. Used in the cross-attention of the decoder.
            of shape `(batch_size, sequence_length, hidden_size)` is a sequence of
        past_key_values (`Tuple[Tuple[tf.Tensor]]` of length `config.n_layers`)
            contains precomputed key and value hidden states of the attention blocks. Can be used to speed up decoding.
            If `past_key_values` are used, the user can optionally input only the last `decoder_input_ids`
            (those that don't have their past key value states given to this model) of shape `(batch_size, 1)`
            instead of all `decoder_input_ids` of shape `(batch_size, sequence_length)`.
        use_cache (`bool`, *optional*, defaults to `True`):
            If set to `True`, `past_key_values` key value states are returned and can be used to speed up
            decoding (see `past_key_values`). Set to `False` during training, `True` during generation
        output_attentions (`bool`, *optional*):
            Whether or not to return the attentions tensors of all attention layers. See `attentions` under returned
            tensors for more detail. This argument can be used only in eager mode, in graph mode the value in the
            config will be used instead.
        output_hidden_states (`bool`, *optional*):
            Whether or not to return the hidden states of all layers. See `hidden_states` under returned tensors for
            more detail. This argument can be used only in eager mode, in graph mode the value in the config will be
            used instead.
        return_dict (`bool`, *optional*):
            Whether or not to return a [`~utils.ModelOutput`] instead of a plain tuple. This
            argument can be used in eager mode, in graph mode the value will always be set to True.
        training (`bool`, *optional*, defaults to `False`):
            Whether or not to use the model in training mode (some modules like dropout modules have different
            behaviors between training and evaluation).
"""


@keras_serializable
class TF{{cookiecutter.camelcase_modelname}}Encoder(tf.keras.layers.Layer):
    config_class = {{cookiecutter.camelcase_modelname}}Config
    """
    Transformer encoder consisting of *config.encoder_layers* self attention layers. Each layer is a
    [`TF{{cookiecutter.camelcase_modelname}}EncoderLayer`].

    Args:
        config: {{cookiecutter.camelcase_modelname}}Config
    """

    def __init__(self, config: {{cookiecutter.camelcase_modelname}}Config, embed_tokens: Optional[tf.keras.layers.Embedding] = None, **kwargs):
        super().__init__(**kwargs)
        self.config = config
        self.dropout = tf.keras.layers.Dropout(config.dropout)
        self.layerdrop = config.encoder_layerdrop
        self.padding_idx = config.pad_token_id
        self.max_source_positions = config.max_position_embeddings
        self.embed_scale = tf.math.sqrt(float(config.d_model)) if config.scale_embedding else 1.0

        self.embed_tokens = embed_tokens
        self.embed_positions = TF{{cookiecutter.camelcase_modelname}}LearnedPositionalEmbedding(
            config.max_position_embeddings,
            config.d_model,
            name="embed_positions",
        )
        self.layers = [TF{{cookiecutter.camelcase_modelname}}EncoderLayer(config, name=f"layers.{i}") for i in range(config.encoder_layers)]
        self.layernorm_embedding = tf.keras.layers.LayerNormalization(epsilon=1e-5, name="layernorm_embedding")

    def get_embed_tokens(self):
        return self.embed_tokens

    def set_embed_tokens(self, embed_tokens):
        self.embed_tokens = embed_tokens

    @unpack_inputs
    def call(
        self,
        input_ids=None,
        inputs_embeds=None,
        attention_mask=None,
        head_mask=None,
        output_attentions=None,
        output_hidden_states=None,
        return_dict=None,
        training=False,
    ):
        """
        Args:
            input_ids (`tf.Tensor` of shape `(batch_size, sequence_length)`):
                Indices of input sequence tokens in the vocabulary. Padding will be ignored by default should you
                provide it.

                Indices can be obtained using [`~{{cookiecutter.camelcase_modelname}}Tokenizer`]. See
                [`PreTrainedTokenizer.encode`] and [`PreTrainedTokenizer.__call__`]
                for details.

                [What are input IDs?](../glossary#input-ids)
            attention_mask (`tf.Tensor` of shape `(batch_size, sequence_length)`, *optional*):
                Mask to avoid performing attention on padding token indices. Mask values selected in `[0, 1]`:

                - 1 for tokens that are **not masked**,
                - 0 for tokens that are **masked**.

                [What are attention masks?](../glossary#attention-mask)
            head_mask (`tf.Tensor` of shape `(encoder_layers, encoder_attention_heads)`, `optional): Mask to nullify selected heads of the attention modules. Mask values selected in `[0, 1]`:

                - 1 indicates the head is **not masked**,
                - 0 indicates the head is **masked**.

            inputs_embeds (`tf.Tensor` of shape `(batch_size, sequence_length, hidden_size)`, *optional*):
                Optionally, instead of passing `input_ids` you can choose to directly pass an embedded
                representation. This is useful if you want more control over how to convert `input_ids` indices
                into associated vectors than the model's internal embedding lookup matrix.
            output_attentions (`bool`, *optional*):
                Whether or not to return the attentions tensors of all attention layers. See `attentions` under
                returned tensors for more detail. This argument can be used only in eager mode, in graph mode the value
                in the config will be used instead.
            output_hidden_states (`bool`, *optional*):
                Whether or not to return the hidden states of all layers. See `hidden_states` under returned tensors
                for more detail. This argument can be used only in eager mode, in graph mode the value in the config
                will be used instead.
            return_dict (`bool`, *optional*):
                Whether or not to return a [`~utils.ModelOutput`] instead of a plain tuple. This
                argument can be used in eager mode, in graph mode the value will always be set to True.
            training (`bool`, *optional*, defaults to `False`):
                Whether or not to use the model in training mode (some modules like dropout modules have different
                behaviors between training and evaluation).
        """

        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 input_ids is not None:
            input_shape = shape_list(input_ids)
        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_embeds is None:
            # if `self.embed_tokens.load_weight_prefix` is set, runs the embedding operation with the correct name
            # scope, so that its weights are registered with the desired name for loading/storing. When `tf.name_scope`
            # is used with a name ending in `/`, that name replaces the current name scope.
            # (embeddings with tf.name_scope: self.embed_tokens.load_weight_prefix/self.embed_tokens.name/embeddings:0)
            context = []
            if hasattr(self.embed_tokens, "load_weight_prefix"):
                context.append(tf.name_scope(self.embed_tokens.load_weight_prefix + "/"))
            with ContextManagers(context):
                check_embeddings_within_bounds(input_ids, self.embed_tokens.input_dim)
                inputs_embeds = self.embed_tokens(input_ids) * self.embed_scale

        embed_pos = self.embed_positions(input_shape)
        hidden_states = inputs_embeds + embed_pos
        hidden_states = self.layernorm_embedding(hidden_states)
        hidden_states = self.dropout(hidden_states, training=training)

        # check attention mask and invert
        if attention_mask is not None:
            # [bsz, seq_len] -> [bsz, 1, tgt_seq_len, src_seq_len]
            attention_mask = _expand_mask(attention_mask)

        encoder_states = () if output_hidden_states else None
        all_attentions = () if output_attentions else None

        # check if head_mask has a correct number of layers specified if desired
        if head_mask is not None:
            tf.debugging.assert_equal(
                shape_list(head_mask)[0],
                len(self.layers),
                message=f"The head_mask should be specified for {len(self.layers)} layers, but it is for {shape_list(head_mask)[0]}.",
            )

        # encoder layers
        for idx, encoder_layer in enumerate(self.layers):

            if output_hidden_states:
                encoder_states = encoder_states + (hidden_states,)
            # add LayerDrop (see https://arxiv.org/abs/1909.11556 for description)
            dropout_probability = random.uniform(0, 1)
            if training and (dropout_probability < self.layerdrop):  # skip the layer
                continue

            hidden_states, attn = encoder_layer(
                hidden_states,
                attention_mask,
                head_mask[idx] if head_mask is not None else None,
            )

            if output_attentions:
                all_attentions += (attn,)

        if output_hidden_states:
            encoder_states = encoder_states + (hidden_states,)

        if not return_dict:
            return tuple(v for v in [hidden_states, encoder_states, all_attentions] if v is not None)
        return TFBaseModelOutput(
            last_hidden_state=hidden_states, hidden_states=encoder_states, attentions=all_attentions
        )


@keras_serializable
class TF{{cookiecutter.camelcase_modelname}}Decoder(tf.keras.layers.Layer):
    config_class = {{cookiecutter.camelcase_modelname}}Config
    """
    Transformer decoder consisting of *config.decoder_layers* layers. Each layer is a [`TF{{cookiecutter.camelcase_modelname}}DecoderLayer`]

    Args:
        config: {{cookiecutter.camelcase_modelname}}Config
        embed_tokens: output embedding
    """

    def __init__(self, config: {{cookiecutter.camelcase_modelname}}Config, embed_tokens: Optional[tf.keras.layers.Embedding] = None, **kwargs):
        super().__init__(**kwargs)
        self.config = config
        self.padding_idx = config.pad_token_id
        self.embed_tokens = embed_tokens
        self.layerdrop = config.decoder_layerdrop
        self.embed_positions = TF{{cookiecutter.camelcase_modelname}}LearnedPositionalEmbedding(
            config.max_position_embeddings,
            config.d_model,
            name="embed_positions",
        )
        self.embed_scale = tf.math.sqrt(float(config.d_model)) if config.scale_embedding else 1.0
        self.layers = [TF{{cookiecutter.camelcase_modelname}}DecoderLayer(config, name=f"layers.{i}") for i in range(config.decoder_layers)]
        self.layernorm_embedding = tf.keras.layers.LayerNormalization(epsilon=1e-5, name="layernorm_embedding")

        self.dropout = tf.keras.layers.Dropout(config.dropout)

    def get_embed_tokens(self):
        return self.embed_tokens

    def set_embed_tokens(self, embed_tokens):
        self.embed_tokens = embed_tokens

    @unpack_inputs
    def call(
        self,
        input_ids=None,
        inputs_embeds=None,
        attention_mask=None,
        encoder_hidden_states=None,
        encoder_attention_mask=None,
        head_mask=None,
        cross_attn_head_mask=None,
        past_key_values=None,
        use_cache=None,
        output_attentions=None,
        output_hidden_states=None,
        return_dict=None,
        training=False,
    ):
        r"""
        Args:
            input_ids (`tf.Tensor` of shape `(batch_size, sequence_length)`):
                Indices of input sequence tokens in the vocabulary. Padding will be ignored by default should you
                provide it.

                Indices can be obtained using [`~{{cookiecutter.camelcase_modelname}}Tokenizer`]. See
                [`PreTrainedTokenizer.encode`] and [`PreTrainedTokenizer.__call__`]
                for details.

                [What are input IDs?](../glossary#input-ids)
            attention_mask (`tf.Tensor` of shape `(batch_size, sequence_length)`, *optional*):
                Mask to avoid performing attention on padding token indices. Mask values selected in `[0, 1]`:

                - 1 for tokens that are **not masked**,
                - 0 for tokens that are **masked**.

                [What are attention masks?](../glossary#attention-mask)
            encoder_hidden_states (`tf.Tensor` of shape `(batch_size, encoder_sequence_length, hidden_size)`, *optional*):
                Sequence of hidden-states at the output of the last layer of the encoder. Used in the cross-attention
                of the decoder.
            encoder_attention_mask (`tf.Tensor` of shape `(batch_size, encoder_sequence_length)`, *optional*):
                Mask to avoid performing cross-attention on padding tokens indices of encoder input_ids. Mask values
                selected in `[0, 1]`:

                - 1 for tokens that are **not masked**,
                - 0 for tokens that are **masked**.

                [What are attention masks?](../glossary#attention-mask)
            head_mask (`tf.Tensor` of shape `(decoder_layers, decoder_attention_heads)`, *optional*):
                Mask to nullify selected heads of the attention modules. Mask values selected in `[0, 1]`:

                - 1 indicates the head is **not masked**,
                - 0 indicates the head is **masked**.

            cross_attn_head_mask (`tf.Tensor` of shape `(decoder_layers, decoder_attention_heads)`, *optional*):
                Mask to nullify selected heads of the cross-attention modules. Mask values selected in `[0, 1]`:

                - 1 indicates the head is **not masked**,
                - 0 indicates the head is **masked**.

            past_key_values (`Tuple[Tuple[tf.Tensor]]` of length `config.n_layers` with each tuple having 2 tuples each of which has 2 tensors of shape `(batch_size, num_heads, sequence_length - 1, embed_size_per_head)`):
                Contains precomputed key and value hidden-states of the attention blocks. Can be used to speed up
                decoding.

                If `past_key_values` are used, the user can optionally input only the last
                `decoder_input_ids` (those that don't have their past key value states given to this model) of
                shape `(batch_size, 1)` instead of all `decoder_input_ids` of shape `(batch_size,
                sequence_length)`.
            inputs_embeds (`tf.Tensor` of shape `(batch_size, sequence_length, hidden_size)`, *optional*):
                Optionally, instead of passing `input_ids` you can choose to directly pass an embedded representation.
                This is useful if you want more control over how to convert `input_ids` indices
                into associated vectors than the model's internal embedding lookup matrix.
            output_attentions (`bool`, *optional*):
                Whether or not to return the attentions tensors of all attention layers. See `attentions` under
                returned tensors for more detail. This argument can be used only in eager mode, in graph mode the value
                in the config will be used instead.
            output_hidden_states (`bool`, *optional*):
                Whether or not to return the hidden states of all layers. See `hidden_states` under returned tensors
                for more detail. This argument can be used only in eager mode, in graph mode the value in the config
                will be used instead.
            return_dict (`bool`, *optional*):
                Whether or not to return a [`~utils.ModelOutput`] instead of a plain tuple. This
                argument can be used in eager mode, in graph mode the value will always be set to True.
            training (`bool`, *optional*, defaults to `False`):
                Whether or not to use the model in training mode (some modules like dropout modules have different
                behaviors between training and evaluation).
        """

        if input_ids is not None and inputs_embeds is not None:
            raise ValueError("You cannot specify both decoder_input_ids and decoder_inputs_embeds at the same time")
        elif input_ids is not None:
            input_shape = shape_list(input_ids)
        elif inputs_embeds is not None:
            input_shape = shape_list(inputs_embeds)[:-1]
        else:
            raise ValueError("You have to specify either decoder_input_ids or decoder_inputs_embeds")

        past_key_values_length = (
            shape_list(past_key_values[0][0])[2] if past_key_values is not None else 0
        )

        # embed positions
        positions = self.embed_positions(input_shape, past_key_values_length)

        if inputs_embeds is None:
            # if `self.embed_tokens.load_weight_prefix` is set, runs the embedding operation with the correct name
            # scope, so that its weights are registered with the desired name for loading/storing. When `tf.name_scope`
            # is used with a name ending in `/`, that name replaces the current name scope.
            # (embeddings with tf.name_scope: self.embed_tokens.load_weight_prefix/self.embed_tokens.name/embeddings:0)
            context = []
            if hasattr(self.embed_tokens, "load_weight_prefix"):
                context.append(tf.name_scope(self.embed_tokens.load_weight_prefix + "/"))
            with ContextManagers(context):
                check_embeddings_within_bounds(input_ids, self.embed_tokens.input_dim)
                inputs_embeds = self.embed_tokens(input_ids)

        hidden_states = inputs_embeds

        attention_mask, combined_attention_mask = self.compute_combined_attns_mask(
            input_ids, attention_mask, input_shape, past_key_values_length
        )

        if encoder_hidden_states is not None and encoder_attention_mask is not None:
            # [bsz, seq_len] -> [bsz, 1, tgt_seq_len, src_seq_len]
            encoder_attention_mask = _expand_mask(encoder_attention_mask, tgt_len=input_shape[-1])

        hidden_states = self.layernorm_embedding(hidden_states + positions)
        hidden_states = self.dropout(hidden_states, training=training)

        # decoder layers
        all_hidden_states = () if output_hidden_states else None
        all_self_attns = () if output_attentions else None
        all_cross_attns = () if (output_attentions and encoder_hidden_states is not None) else None
        present_key_values = () if use_cache else None

        # check if head_mask and cross_attn_head_mask have a correct number of layers specified if desired
        for attn_mask_name, attn_mask in [("head_mask", head_mask), ("cross_attn_head_mask", cross_attn_head_mask)]:
            if attn_mask is not None:
                tf.debugging.assert_equal(
                    shape_list(attn_mask)[0],
                    len(self.layers),
                    message=f"The {attn_mask_name} should be specified for {len(self.layers)} layers, but it is for {shape_list(attn_mask)[0]}.",
                )

        for idx, decoder_layer in enumerate(self.layers):
            # add LayerDrop (see https://arxiv.org/abs/1909.11556 for description)
            if output_hidden_states:
                all_hidden_states += (hidden_states,)

            dropout_probability = random.uniform(0, 1)

            if training and (dropout_probability < self.layerdrop):
                continue

            past_key_value = past_key_values[idx] if past_key_values is not None else None

            hidden_states, layer_self_attn, layer_cross_attn, present_key_value = decoder_layer(
                hidden_states,
                attention_mask=combined_attention_mask,
                encoder_hidden_states=encoder_hidden_states,
                encoder_attention_mask=encoder_attention_mask,
                layer_head_mask=head_mask[idx] if head_mask is not None else None,
                cross_attn_layer_head_mask=cross_attn_head_mask[idx]
                if cross_attn_head_mask is not None
                else None,
                past_key_value=past_key_value,
            )

            if use_cache:
                present_key_values += (present_key_value,)

            if output_attentions:
                all_self_attns += (layer_self_attn,)

                if encoder_hidden_states is not None:
                    all_cross_attns += (layer_cross_attn,)

        if output_hidden_states:
            all_hidden_states += (hidden_states,)

        if not return_dict:
            return hidden_states, present_key_values, all_hidden_states, all_self_attns, all_cross_attns
        else:
            return TFBaseModelOutputWithPastAndCrossAttentions(
                last_hidden_state=hidden_states,
                past_key_values=present_key_values,
                hidden_states=all_hidden_states,
                attentions=all_self_attns,
                cross_attentions=all_cross_attns,
            )

    @tf.function
    def compute_combined_attns_mask(self, input_ids, attention_mask, input_shape, past_key_values_length):
        # [bsz, seq_len] -> [bsz, 1, tgt_seq_len, src_seq_len]
        combined_attention_mask = None
        if input_shape[-1] > 1:
            combined_attention_mask = _make_causal_mask(input_shape, past_key_values_length=past_key_values_length)
        else:
            combined_attention_mask = _expand_mask(
                tf.ones((input_shape[0], input_shape[1] + past_key_values_length)), tgt_len=input_shape[-1]
            )

        if attention_mask is None and input_ids is not None and input_shape[-1] > 1:
            attention_mask = tf.cast(
                tf.math.not_equal(input_ids, self.config.pad_token_id), input_ids.dtype
            )
            attention_mask = tf.concat(
                [
                    tf.ones((input_shape[0], past_key_values_length), dtype=attention_mask.dtype),
                    attention_mask,
                ],
                axis=-1,
            )
        else:
            attention_mask = tf.ones((input_shape[0], input_shape[1] + past_key_values_length))

        return attention_mask, combined_attention_mask


@keras_serializable
class TF{{cookiecutter.camelcase_modelname}}MainLayer(tf.keras.layers.Layer):
    config_class = {{cookiecutter.camelcase_modelname}}Config

    def __init__(self, config: {{cookiecutter.camelcase_modelname}}Config, **kwargs):
        super().__init__(**kwargs)

        self.config = config
        self.shared = tf.keras.layers.Embedding(
            input_dim=config.vocab_size,
            output_dim=config.d_model,
            embeddings_initializer=tf.keras.initializers.TruncatedNormal(stddev=self.config.init_std),
            name="model.shared"
        )
        # Additional attribute to specify the expected name scope of the layer (for loading/storing weights)
        self.shared.load_weight_prefix = "model.shared"

        self.encoder = TF{{cookiecutter.camelcase_modelname}}Encoder(config, self.shared, name="encoder")
        self.decoder = TF{{cookiecutter.camelcase_modelname}}Decoder(config, self.shared, name="decoder")

    def get_input_embeddings(self):
        return self.shared

    def set_input_embeddings(self, new_embeddings):
        self.shared = new_embeddings
        self.encoder.embed_tokens = self.shared
        self.decoder.embed_tokens = self.shared

    @unpack_inputs
    def call(
        self,
        input_ids=None,
        attention_mask=None,
        decoder_input_ids=None,
        decoder_attention_mask=None,
        head_mask=None,
        decoder_head_mask=None,
        cross_attn_head_mask=None,
        encoder_outputs: Optional[Union[Tuple, TFBaseModelOutput]] = None,
        past_key_values=None,
        inputs_embeds=None,
        decoder_inputs_embeds=None,
        use_cache=None,
        output_attentions=None,
        output_hidden_states=None,
        return_dict=None,
        training=False,
        **kwargs
    ):

        if decoder_input_ids is None and decoder_inputs_embeds is None:
            use_cache = False

        if encoder_outputs is None:
            encoder_outputs = self.encoder(
                input_ids=input_ids,
                attention_mask=attention_mask,
                head_mask=head_mask,
                inputs_embeds=inputs_embeds,
                output_attentions=output_attentions,
                output_hidden_states=output_hidden_states,
                return_dict=return_dict,
                training=training,
            )
        # If the user passed a tuple for encoder_outputs, we wrap it in a TFBaseModelOutput when return_dict=True
        elif return_dict and not isinstance(encoder_outputs, TFBaseModelOutput):
            encoder_outputs = TFBaseModelOutput(
                last_hidden_state=encoder_outputs[0],
                hidden_states=encoder_outputs[1] if len(encoder_outputs) > 1 else None,
                attentions=encoder_outputs[2] if len(encoder_outputs) > 2 else None,
            )
        # If the user passed a TFBaseModelOutput for encoder_outputs, we wrap it in a tuple when return_dict=False
        elif not return_dict and not isinstance(encoder_outputs, tuple):
            encoder_outputs = encoder_outputs.to_tuple()

        decoder_outputs = self.decoder(
            decoder_input_ids,
            attention_mask=decoder_attention_mask,
            encoder_hidden_states=encoder_outputs[0],
            encoder_attention_mask=attention_mask,
            head_mask=decoder_head_mask,
            cross_attn_head_mask=cross_attn_head_mask,
            past_key_values=past_key_values,
            inputs_embeds=decoder_inputs_embeds,
            use_cache=use_cache,
            output_attentions=output_attentions,
            output_hidden_states=output_hidden_states,
            return_dict=return_dict,
            training=training,
        )

        if not return_dict:
            return decoder_outputs + encoder_outputs

        return TFSeq2SeqModelOutput(
            last_hidden_state=decoder_outputs.last_hidden_state,
            past_key_values=decoder_outputs.past_key_values,
            decoder_hidden_states=decoder_outputs.hidden_states,
            decoder_attentions=decoder_outputs.attentions,
            cross_attentions=decoder_outputs.cross_attentions,
            encoder_last_hidden_state=encoder_outputs.last_hidden_state,
            encoder_hidden_states=encoder_outputs.hidden_states,
            encoder_attentions=encoder_outputs.attentions,
        )


@add_start_docstrings(
    "The bare {{cookiecutter.uppercase_modelname}} Model outputting raw hidden-states without any specific head on top.",
    {{cookiecutter.uppercase_modelname}}_START_DOCSTRING,
)
class TF{{cookiecutter.camelcase_modelname}}Model(TF{{cookiecutter.camelcase_modelname}}PreTrainedModel):
    def __init__(self, config: {{cookiecutter.camelcase_modelname}}Config, *inputs, **kwargs):
        super().__init__(config, *inputs, **kwargs)

        self.model = TF{{cookiecutter.camelcase_modelname}}MainLayer(config, name="model")

    def get_encoder(self):
        return self.model.encoder

    def get_decoder(self):
        return self.model.decoder

    @unpack_inputs
    @add_start_docstrings_to_model_forward({{cookiecutter.uppercase_modelname}}_INPUTS_DOCSTRING.format("batch_size, sequence_length"))
    @add_code_sample_docstrings(
        checkpoint=_CHECKPOINT_FOR_DOC,
        output_type=TFSeq2SeqModelOutput,
        config_class=_CONFIG_FOR_DOC,
    )
    def call(
        self,
        input_ids=None,
        attention_mask=None,
        decoder_input_ids=None,
        decoder_attention_mask=None,
        head_mask=None,
        decoder_head_mask=None,
        cross_attn_head_mask=None,
        encoder_outputs: Optional[Union[Tuple, TFBaseModelOutput]] = None,
        past_key_values=None,
        inputs_embeds=None,
        decoder_inputs_embeds=None,
        use_cache=None,
        output_attentions=None,
        output_hidden_states=None,
        return_dict=None,
        training=False,
        **kwargs
    ):

        outputs = self.model(
            input_ids=input_ids,
            attention_mask=attention_mask,
            decoder_input_ids=decoder_input_ids,
            decoder_attention_mask=decoder_attention_mask,
            head_mask=head_mask,
            decoder_head_mask=decoder_head_mask,
            cross_attn_head_mask=cross_attn_head_mask,
            encoder_outputs=encoder_outputs,
            past_key_values=past_key_values,
            inputs_embeds=inputs_embeds,
            decoder_inputs_embeds=decoder_inputs_embeds,
            use_cache=use_cache,
            output_attentions=output_attentions,
            output_hidden_states=output_hidden_states,
            return_dict=return_dict,
            training=training,
        )

        return outputs


# Copied from transformers.models.bart.modeling_tf_bart.BiasLayer
class BiasLayer(tf.keras.layers.Layer):
    """
    Bias as a layer. It is used for serialization purposes: `tf.keras.Model.save_weights` stores on a per-layer basis,
    so all weights have to be registered in a layer.
    """

    def __init__(self, shape, initializer, trainable, name, **kwargs):
        super().__init__(name=name, **kwargs)
        # Note: the name of this variable will NOT be scoped when serialized, i.e. it will not be in the format of
        # "outer_layer/inner_layer/.../name:0". Instead, it will be "name:0". For further details, see:
        # https://github.com/huggingface/transformers/pull/18833#issuecomment-1233090214
        self.bias = self.add_weight(name=name, shape=shape, initializer=initializer, trainable=trainable)

    def call(self, x):
        return x + self.bias


@add_start_docstrings(
    "The {{cookiecutter.uppercase_modelname}} Model with a language modeling head. Can be used for summarization.",
    {{cookiecutter.uppercase_modelname}}_START_DOCSTRING,
)
class TF{{cookiecutter.camelcase_modelname}}ForConditionalGeneration(TF{{cookiecutter.camelcase_modelname}}PreTrainedModel):
    _keys_to_ignore_on_load_unexpected = [
        r"model.encoder.embed_tokens.weight",
        r"model.decoder.embed_tokens.weight",
    ]

    def __init__(self, config, *inputs, **kwargs):
        super().__init__(config, *inputs, **kwargs)
        self.model = TF{{cookiecutter.camelcase_modelname}}MainLayer(config, name="model")
        self.use_cache = config.use_cache
        # final_bias_logits is registered as a buffer in pytorch, so not trainable for the sake of consistency.
        self.bias_layer = BiasLayer(
            name="final_logits_bias", shape=[1, config.vocab_size], initializer="zeros", trainable=False
        )

    def get_decoder(self):
        return self.model.decoder

    def get_encoder(self):
        return self.model.encoder

    def get_bias(self):
        return {"final_logits_bias": self.bias_layer.bias}

    def set_bias(self, value):
        # Replaces the existing layers containing bias for correct (de)serialization.
        vocab_size = value["final_logits_bias"].shape[-1]
        self.bias_layer = BiasLayer(
            name="final_logits_bias", shape=[1, vocab_size], initializer="zeros", trainable=False
        )
        self.bias_layer.bias.assign(value["final_logits_bias"])

    def get_output_embeddings(self):
        return self.get_input_embeddings()

    def set_output_embeddings(self, value):
        self.set_input_embeddings(value)

    @unpack_inputs
    @add_start_docstrings_to_model_forward({{cookiecutter.uppercase_modelname}}_INPUTS_DOCSTRING)
    @replace_return_docstrings(output_type=TFSeq2SeqLMOutput, config_class=_CONFIG_FOR_DOC)
    def call(
        self,
        input_ids=None,
        attention_mask=None,
        decoder_input_ids=None,
        decoder_attention_mask=None,
        head_mask=None,
        decoder_head_mask=None,
        cross_attn_head_mask=None,
        encoder_outputs: Optional[TFBaseModelOutput] = None,
        past_key_values=None,
        inputs_embeds=None,
        decoder_inputs_embeds=None,
        use_cache=None,
        output_attentions=None,
        output_hidden_states=None,
        return_dict=None,
        labels=None,
        training=False,
    ):
        """
        Returns:

        Examples:

        ```python
        >>> from transformers import {{cookiecutter.camelcase_modelname}}Tokenizer, TF{{cookiecutter.camelcase_modelname}}ForConditionalGeneration
        >>> import tensorflow as tf
        >>> mname = '{{cookiecutter.checkpoint_identifier}}'
        >>> tokenizer = {{cookiecutter.camelcase_modelname}}Tokenizer.from_pretrained(mname)
        >>> TXT = "My friends are <mask> but they eat too many carbs."
        >>> model = TF{{cookiecutter.camelcase_modelname}}ForConditionalGeneration.from_pretrained(mname)
        >>> batch = tokenizer([TXT], return_tensors='tf')
        >>> logits = model(inputs=batch.input_ids).logits
        >>> probs = tf.nn.softmax(logits[0])
        >>> # probs[5] is associated with the mask token
        ```"""

        if labels is not None:
            use_cache = False
            if decoder_input_ids is None and decoder_inputs_embeds is None:
                decoder_input_ids = shift_tokens_right(
                    labels, self.config.pad_token_id, self.config.decoder_start_token_id
                )

        outputs = self.model(
            input_ids,
            attention_mask=attention_mask,
            decoder_input_ids=decoder_input_ids,
            encoder_outputs=encoder_outputs,
            decoder_attention_mask=decoder_attention_mask,
            head_mask=head_mask,
            decoder_head_mask=decoder_head_mask,
            cross_attn_head_mask=cross_attn_head_mask,
            past_key_values=past_key_values,
            inputs_embeds=inputs_embeds,
            decoder_inputs_embeds=decoder_inputs_embeds,
            use_cache=use_cache,
            output_attentions=output_attentions,
            output_hidden_states=output_hidden_states,
            return_dict=return_dict,
            training=training
        )
        lm_logits = tf.matmul(outputs[0], self.model.shared.weights, transpose_b=True)
        lm_logits = self.bias_layer(lm_logits)
        masked_lm_loss = None if labels is None else self.hf_compute_loss(labels, lm_logits)

        if not return_dict:
            output = (lm_logits,) + outputs[1:]
            return ((masked_lm_loss,) + output) if masked_lm_loss is not None else output
        return TFSeq2SeqLMOutput(
            loss=masked_lm_loss,
            logits=lm_logits,
            past_key_values=outputs.past_key_values,  # index 1 of d outputs
            decoder_hidden_states=outputs.decoder_hidden_states,  # index 2 of d outputs
            decoder_attentions=outputs.decoder_attentions,  # index 3 of d outputs
            cross_attentions=outputs.cross_attentions,  # index 4 of d outputs
            encoder_last_hidden_state=outputs.encoder_last_hidden_state,  # index 0 of encoder outputs
            encoder_hidden_states=outputs.encoder_hidden_states,  # 1 of e out
            encoder_attentions=outputs.encoder_attentions,  # 2 of e out
        )

    def prepare_inputs_for_generation(
        self,
        decoder_input_ids,
        past_key_values=None,
        attention_mask=None,
        head_mask=None,
        decoder_head_mask=None,
        cross_attn_head_mask=None,
        use_cache=None,
        encoder_outputs=None,
        **kwargs
    ):
        # cut decoder_input_ids if past is used
        if past_key_values is not None:
            decoder_input_ids = decoder_input_ids[:, -1:]

        return {
            "input_ids": None,  # needs to be passed to make Keras.layer.__call__ happy
            "encoder_outputs": encoder_outputs,
            "past_key_values": past_key_values,
            "decoder_input_ids": decoder_input_ids,
            "attention_mask": attention_mask,
            "head_mask": head_mask,
            "decoder_head_mask": decoder_head_mask,
            "cross_attn_head_mask": cross_attn_head_mask,
            "use_cache": use_cache,  # change this to avoid caching (presumably for debugging)
        }

    def hf_compute_loss(self, labels, logits):
        """CrossEntropyLoss that ignores pad tokens"""
        loss_fn = tf.keras.losses.SparseCategoricalCrossentropy(
            from_logits=True,
            reduction=tf.keras.losses.Reduction.NONE,
        )
        melted_labels = tf.reshape(labels, (-1,))
        active_loss = tf.not_equal(melted_labels, self.config.pad_token_id)
        reduced_logits = tf.boolean_mask(tf.reshape(logits, (-1, shape_list(logits)[2])), active_loss)
        labels = tf.boolean_mask(melted_labels, active_loss)
        return loss_fn(labels, reduced_logits)
{% endif -%}