# coding=utf-8
# Copyright 2021 The Fairseq 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 MBart model."""

from __future__ import annotations

import random
from typing import Optional, Tuple, Union

import tensorflow as tf

from ...activations_tf import get_tf_activation
from ...modeling_tf_outputs import (
    TFBaseModelOutput,
    TFBaseModelOutputWithPastAndCrossAttentions,
    TFSeq2SeqLMOutput,
    TFSeq2SeqModelOutput,
)

# Public API
from ...modeling_tf_utils import (
    TFCausalLanguageModelingLoss,
    TFModelInputType,
    TFPreTrainedModel,
    keras,
    keras_serializable,
    unpack_inputs,
)
from ...tf_utils import check_embeddings_within_bounds, shape_list, stable_softmax
from ...utils import (
    add_code_sample_docstrings,
    add_end_docstrings,
    add_start_docstrings,
    add_start_docstrings_to_model_forward,
    logging,
    replace_return_docstrings,
)
from .configuration_mbart import MBartConfig


logger = logging.get_logger(__name__)

_CHECKPOINT_FOR_DOC = "facebook/mbart-large-cc25"
_CONFIG_FOR_DOC = "MBartConfig"


LARGE_NEGATIVE = -1e8


def shift_tokens_right(input_ids: tf.Tensor, pad_token_id: int):
    """
    Shift input ids one token to the right, and wrap the last non pad token (the <LID> token) Note that MBart does not
    have a single `decoder_start_token_id` in contrast to other Bart-like models.
    """
    if pad_token_id is None:
        raise ValueError("self.model.config.pad_token_id has to be defined.")
    # replace possible -100 values in labels by `pad_token_id`
    input_ids = tf.where(
        input_ids == -100, tf.fill(shape_list(input_ids), tf.cast(pad_token_id, input_ids.dtype)), input_ids
    )
    language_id_index = (
        tf.reduce_sum(tf.cast(tf.math.not_equal(input_ids, pad_token_id), dtype=input_ids.dtype), axis=-1) - 1
    )
    language_id_index = tf.stack(
        [tf.range(shape_list(input_ids)[0], dtype=input_ids.dtype), language_id_index], axis=-1
    )
    languages_ids = tf.gather_nd(input_ids, language_id_index)

    shifted_input_ids = tf.concat([tf.expand_dims(languages_ids, axis=-1), input_ids[:, :-1]], axis=-1)

    return shifted_input_ids


# Copied from transformers.models.bart.modeling_tf_bart._make_causal_mask
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 = input_ids_shape[0]
    tgt_len = input_ids_shape[1]
    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))


# Copied from transformers.models.bart.modeling_tf_bart._expand_mask
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


# Copied from transformers.models.bart.modeling_tf_bart.TFBartLearnedPositionalEmbedding with Bart->MBart
class TFMBartLearnedPositionalEmbedding(keras.layers.Embedding):
    """
    This module learns positional embeddings up to a fixed maximum size.
    """

    def __init__(self, num_embeddings: int, embedding_dim: int, **kwargs):
        # MBart is set up so that if padding_idx is specified then offset the embedding ids by 2
        # and adjust num_embeddings appropriately. Other models don't have this hack
        self.offset = 2
        super().__init__(num_embeddings + self.offset, embedding_dim, **kwargs)

    def call(
        self,
        input_shape: Optional[tf.TensorShape] = None,
        past_key_values_length: int = 0,
        position_ids: tf.Tensor | None = None,
    ):
        """Input is expected to be of size [bsz x seqlen]."""
        if position_ids is None:
            seq_len = input_shape[1]
            position_ids = tf.range(seq_len, delta=1, name="range")
            position_ids += past_key_values_length

        offset_dtype = position_ids.dtype if isinstance(position_ids, tf.Tensor) else tf.int32
        return super().call(position_ids + tf.constant(self.offset, dtype=offset_dtype))


# Copied from transformers.models.bart.modeling_tf_bart.TFBartAttention with Bart->MBart
class TFMBartAttention(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 = keras.layers.Dropout(dropout)
        self.head_dim = embed_dim // num_heads
        if (self.head_dim * num_heads) != self.embed_dim:
            raise ValueError(
                f"embed_dim must be divisible by num_heads (got `embed_dim`: {self.embed_dim}"
                f" and `num_heads`: {num_heads})."
            )
        self.scaling = self.head_dim**-0.5
        self.is_decoder = is_decoder

        self.k_proj = keras.layers.Dense(embed_dim, use_bias=bias, name="k_proj")
        self.q_proj = keras.layers.Dense(embed_dim, use_bias=bias, name="q_proj")
        self.v_proj = keras.layers.Dense(embed_dim, use_bias=bias, name="v_proj")
        self.out_proj = 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: Optional[bool] = 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"
                f" {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"
                    f" {shape_list(attention_mask)}"
                ),
            )

            attention_mask = tf.cast(attention_mask, dtype=attn_weights.dtype)
            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"
                    f" {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"
                f" {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

    def build(self, input_shape=None):
        if self.built:
            return
        self.built = True
        if getattr(self, "k_proj", None) is not None:
            with tf.name_scope(self.k_proj.name):
                self.k_proj.build([None, None, self.embed_dim])
        if getattr(self, "q_proj", None) is not None:
            with tf.name_scope(self.q_proj.name):
                self.q_proj.build([None, None, self.embed_dim])
        if getattr(self, "v_proj", None) is not None:
            with tf.name_scope(self.v_proj.name):
                self.v_proj.build([None, None, self.embed_dim])
        if getattr(self, "out_proj", None) is not None:
            with tf.name_scope(self.out_proj.name):
                self.out_proj.build([None, None, self.embed_dim])


class TFMBartEncoderLayer(keras.layers.Layer):
    def __init__(self, config: MBartConfig, **kwargs):
        super().__init__(**kwargs)
        self.embed_dim = config.d_model
        self.self_attn = TFMBartAttention(
            self.embed_dim, config.encoder_attention_heads, dropout=config.attention_dropout, name="self_attn"
        )
        self.self_attn_layer_norm = keras.layers.LayerNormalization(epsilon=1e-5, name="self_attn_layer_norm")
        self.dropout = keras.layers.Dropout(config.dropout)
        self.activation_fn = get_tf_activation(config.activation_function)
        self.activation_dropout = keras.layers.Dropout(config.activation_dropout)
        self.fc1 = keras.layers.Dense(config.encoder_ffn_dim, name="fc1")
        self.fc2 = keras.layers.Dense(self.embed_dim, name="fc2")
        self.final_layer_norm = keras.layers.LayerNormalization(epsilon=1e-5, name="final_layer_norm")
        self.config = config

    def call(
        self,
        hidden_states: tf.Tensor,
        attention_mask: tf.Tensor,
        layer_head_mask: tf.Tensor,
        training: Optional[bool] = 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.self_attn_layer_norm(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

        residual = hidden_states
        hidden_states = self.final_layer_norm(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

        return hidden_states, self_attn_weights

    def build(self, input_shape=None):
        if self.built:
            return
        self.built = True
        if getattr(self, "self_attn", None) is not None:
            with tf.name_scope(self.self_attn.name):
                self.self_attn.build(None)
        if getattr(self, "self_attn_layer_norm", None) is not None:
            with tf.name_scope(self.self_attn_layer_norm.name):
                self.self_attn_layer_norm.build([None, None, self.embed_dim])
        if getattr(self, "fc1", None) is not None:
            with tf.name_scope(self.fc1.name):
                self.fc1.build([None, None, self.embed_dim])
        if getattr(self, "fc2", None) is not None:
            with tf.name_scope(self.fc2.name):
                self.fc2.build([None, None, self.config.encoder_ffn_dim])
        if getattr(self, "final_layer_norm", None) is not None:
            with tf.name_scope(self.final_layer_norm.name):
                self.final_layer_norm.build([None, None, self.embed_dim])


class TFMBartDecoderLayer(keras.layers.Layer):
    def __init__(self, config: MBartConfig, **kwargs):
        super().__init__(**kwargs)
        self.embed_dim = config.d_model
        self.self_attn = TFMBartAttention(
            embed_dim=self.embed_dim,
            num_heads=config.decoder_attention_heads,
            dropout=config.attention_dropout,
            name="self_attn",
            is_decoder=True,
        )
        self.dropout = keras.layers.Dropout(config.dropout)
        self.activation_fn = get_tf_activation(config.activation_function)
        self.activation_dropout = keras.layers.Dropout(config.activation_dropout)

        self.self_attn_layer_norm = keras.layers.LayerNormalization(epsilon=1e-5, name="self_attn_layer_norm")
        self.encoder_attn = TFMBartAttention(
            self.embed_dim,
            config.decoder_attention_heads,
            dropout=config.attention_dropout,
            name="encoder_attn",
            is_decoder=True,
        )
        self.encoder_attn_layer_norm = keras.layers.LayerNormalization(epsilon=1e-5, name="encoder_attn_layer_norm")
        self.fc1 = keras.layers.Dense(config.decoder_ffn_dim, name="fc1")
        self.fc2 = keras.layers.Dense(self.embed_dim, name="fc2")
        self.final_layer_norm = keras.layers.LayerNormalization(epsilon=1e-5, name="final_layer_norm")
        self.config = config

    def call(
        self,
        hidden_states: tf.Tensor,
        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: Optional[bool] = 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
        hidden_states = self.self_attn_layer_norm(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

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

            # 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.final_layer_norm(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

        return (
            hidden_states,
            self_attn_weights,
            cross_attn_weights,
            present_key_value,
        )

    def build(self, input_shape=None):
        if self.built:
            return
        self.built = True
        if getattr(self, "self_attn", None) is not None:
            with tf.name_scope(self.self_attn.name):
                self.self_attn.build(None)
        if getattr(self, "self_attn_layer_norm", None) is not None:
            with tf.name_scope(self.self_attn_layer_norm.name):
                self.self_attn_layer_norm.build([None, None, self.embed_dim])
        if getattr(self, "encoder_attn", None) is not None:
            with tf.name_scope(self.encoder_attn.name):
                self.encoder_attn.build(None)
        if getattr(self, "encoder_attn_layer_norm", None) is not None:
            with tf.name_scope(self.encoder_attn_layer_norm.name):
                self.encoder_attn_layer_norm.build([None, None, self.embed_dim])
        if getattr(self, "fc1", None) is not None:
            with tf.name_scope(self.fc1.name):
                self.fc1.build([None, None, self.embed_dim])
        if getattr(self, "fc2", None) is not None:
            with tf.name_scope(self.fc2.name):
                self.fc2.build([None, None, self.config.decoder_ffn_dim])
        if getattr(self, "final_layer_norm", None) is not None:
            with tf.name_scope(self.final_layer_norm.name):
                self.final_layer_norm.build([None, None, self.embed_dim])


class TFMBartPreTrainedModel(TFPreTrainedModel):
    config_class = MBartConfig
    base_model_prefix = "model"


MBART_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 [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 ([`MBartConfig`]): 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.
"""

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

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

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

            MBart uses a specific language id token as the starting token for `decoder_input_ids` generation that
            varies according to source and target language, *e.g.* 25004 for *en_XX*, and 25003 for *de_DE*. 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.
        decoder_position_ids (`tf.Tensor` of shape `(batch_size, sequence_length)`, *optional*):
            Indices of positions of each decoder input sequence tokens in the position embeddings. Selected in the
            range `[0, config.max_position_embeddings - 1]`.
        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)`.
        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.
        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).
"""

MBART_GENERATION_EXAMPLE = r"""
    Translation example:

    ```python
    >>> from transformers import AutoTokenizer, TFMBartForConditionalGeneration

    >>> model = TFMBartForConditionalGeneration.from_pretrained("facebook/mbart-large-en-ro")
    >>> tokenizer = AutoTokenizer.from_pretrained("facebook/mbart-large-en-ro")

    >>> example_english_phrase = "42 is the answer"
    >>> inputs = tokenizer(example_english_phrase, return_tensors="tf")

    >>> # Translate
    >>> generated_ids = model.generate(**inputs, num_beams=4, max_length=5)
    >>> tokenizer.batch_decode(generated_ids, skip_special_tokens=True, clean_up_tokenization_spaces=False)[0]
    '42 este răspuns'
    ```

    Mask filling example:

    ```python
    >>> from transformers import AutoTokenizer, TFMBartForConditionalGeneration
    >>> import tensorflow as tf

    >>> model = TFMBartForConditionalGeneration.from_pretrained("facebook/mbart-large-cc25")
    >>> tokenizer = AutoTokenizer.from_pretrained("facebook/mbart-large-cc25")

    >>> # de_DE is the language symbol id <LID> for German
    >>> TXT = "</s> Meine Freunde sind <mask> nett aber sie essen zu viel Kuchen. </s> de_DE"

    >>> input_ids = tokenizer([TXT], add_special_tokens=False, return_tensors="tf")["input_ids"]
    >>> logits = model(input_ids).logits

    >>> masked_index = tf.where(input_ids[0] == tokenizer.mask_token_id)[0, 0]
    >>> probs = tf.nn.softmax(logits[0, masked_index], axis=0)
    >>> values, predictions = tf.math.top_k(probs, 5)

    >>> tokenizer.decode(predictions).split()
    ['nett', 'sehr', 'ganz', 'nicht', 'so']
    ```
"""


@keras_serializable
class TFMBartEncoder(keras.layers.Layer):
    config_class = MBartConfig
    """
    Transformer encoder consisting of *config.encoder_layers* self attention layers. Each layer is a
    [`TFMBartEncoderLayer`].

    Args:
        config: MBartConfig
    """

    def __init__(self, config: MBartConfig, embed_tokens: Optional[keras.layers.Embedding] = None, **kwargs):
        super().__init__(**kwargs)
        self.config = config
        self.dropout = 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 = TFMBartLearnedPositionalEmbedding(
            config.max_position_embeddings,
            config.d_model,
            name="embed_positions",
        )
        self.layers = [TFMBartEncoderLayer(config, name=f"layers.{i}") for i in range(config.encoder_layers)]
        self.layernorm_embedding = keras.layers.LayerNormalization(epsilon=1e-5, name="layernorm_embedding")
        self.layer_norm = keras.layers.LayerNormalization(epsilon=1e-5, name="layer_norm")
        self.embed_dim = config.d_model

    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: TFModelInputType | None = None,
        inputs_embeds: tf.Tensor | None = None,
        attention_mask: tf.Tensor | None = None,
        head_mask: tf.Tensor | None = None,
        output_attentions: Optional[bool] = None,
        output_hidden_states: Optional[bool] = None,
        return_dict: Optional[bool] = None,
        training: Optional[bool] = False,
    ) -> Union[TFBaseModelOutput, Tuple[tf.Tensor]]:
        """
        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 [`AutoTokenizer`]. 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:
            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)
        else:
            attention_mask = None

        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"
                    f" {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,)

        hidden_states = self.layer_norm(hidden_states)

        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
        )

    def build(self, input_shape=None):
        if self.built:
            return
        self.built = True
        if getattr(self, "embed_positions", None) is not None:
            with tf.name_scope(self.embed_positions.name):
                self.embed_positions.build(None)
        if getattr(self, "layernorm_embedding", None) is not None:
            with tf.name_scope(self.layernorm_embedding.name):
                self.layernorm_embedding.build([None, None, self.embed_dim])
        if getattr(self, "layer_norm", None) is not None:
            with tf.name_scope(self.layer_norm.name):
                self.layer_norm.build([None, None, self.config.d_model])
        if getattr(self, "layers", None) is not None:
            for layer in self.layers:
                with tf.name_scope(layer.name):
                    layer.build(None)


@keras_serializable
class TFMBartDecoder(keras.layers.Layer):
    config_class = MBartConfig
    """
    Transformer decoder consisting of *config.decoder_layers* layers. Each layer is a [`TFMBartDecoderLayer`]

    Args:
        config: MBartConfig
        embed_tokens: output embedding
    """

    def __init__(self, config: MBartConfig, embed_tokens: Optional[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 = TFMBartLearnedPositionalEmbedding(
            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 = [TFMBartDecoderLayer(config, name=f"layers.{i}") for i in range(config.decoder_layers)]
        self.layernorm_embedding = keras.layers.LayerNormalization(epsilon=1e-5, name="layernorm_embedding")
        self.layer_norm = keras.layers.LayerNormalization(epsilon=1e-5, name="layer_norm")

        self.dropout = 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: TFModelInputType = None,
        inputs_embeds: tf.Tensor | None = None,
        attention_mask: tf.Tensor | None = None,
        position_ids: tf.Tensor | None = None,
        encoder_hidden_states: tf.Tensor | None = None,
        encoder_attention_mask: tf.Tensor | None = None,
        head_mask: tf.Tensor | None = None,
        cross_attn_head_mask: tf.Tensor | None = None,
        past_key_values: Tuple[Tuple[tf.Tensor]] | None = 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, tf.Tensor, tf.Tensor, tf.Tensor, tf.Tensor]
    ]:
        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 [`AutoTokenizer`]. 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)
            position_ids (`tf.Tensor` of shape `(batch_size, sequence_length)`, *optional*):
                Indices of positions of each decoder input sequence tokens in the position embeddings. Selected in the
                range `[0, config.max_position_embeddings - 1]`.
            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
        if position_ids is None:
            positions = self.embed_positions(input_shape, past_key_values_length)
        else:
            positions = self.embed_positions(input_shape, position_ids=position_ids)

        if inputs_embeds is None:
            check_embeddings_within_bounds(input_ids, self.embed_tokens.input_dim)
            inputs_embeds = self.embed_tokens(input_ids) * self.embed_scale

        hidden_states = inputs_embeds

        # [bsz, seq_len] -> [bsz, 1, tgt_seq_len, src_seq_len]
        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 not None:
            combined_attention_mask = combined_attention_mask + _expand_mask(attention_mask, tgt_len=input_shape[-1])

        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"
                        f" {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,)

        hidden_states = self.layer_norm(hidden_states)

        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,
            )

    def build(self, input_shape=None):
        if self.built:
            return
        self.built = True
        if getattr(self, "embed_positions", None) is not None:
            with tf.name_scope(self.embed_positions.name):
                self.embed_positions.build(None)
        if getattr(self, "layernorm_embedding", None) is not None:
            with tf.name_scope(self.layernorm_embedding.name):
                self.layernorm_embedding.build([None, None, self.config.d_model])
        if getattr(self, "layer_norm", None) is not None:
            with tf.name_scope(self.layer_norm.name):
                self.layer_norm.build([None, None, self.config.d_model])
        if getattr(self, "layers", None) is not None:
            for layer in self.layers:
                with tf.name_scope(layer.name):
                    layer.build(None)


@keras_serializable
class TFMBartMainLayer(keras.layers.Layer):
    config_class = MBartConfig

    def __init__(self, config: MBartConfig, **kwargs):
        super().__init__(**kwargs)

        self.config = config
        self.shared = keras.layers.Embedding(
            input_dim=config.vocab_size,
            output_dim=config.d_model,
            embeddings_initializer=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 = TFMBartEncoder(config, self.shared, name="encoder")
        self.decoder = TFMBartDecoder(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: TFModelInputType = None,
        attention_mask: tf.Tensor | None = None,
        decoder_input_ids: tf.Tensor | None = None,
        decoder_attention_mask: tf.Tensor | None = None,
        decoder_position_ids: tf.Tensor | None = None,
        head_mask: tf.Tensor | None = None,
        decoder_head_mask: tf.Tensor | None = None,
        cross_attn_head_mask: tf.Tensor | None = None,
        encoder_outputs: Optional[Union[Tuple, TFBaseModelOutput]] = None,
        past_key_values: Tuple[Tuple[tf.Tensor]] | None = None,
        inputs_embeds: tf.Tensor | None = None,
        decoder_inputs_embeds: tf.Tensor | None = 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,
        **kwargs,
    ) -> Union[TFSeq2SeqModelOutput, tf.Tensor]:
        if decoder_input_ids is None and decoder_inputs_embeds is None:
            use_cache = False

        output_hidden_states = (
            output_hidden_states if output_hidden_states is not None else self.config.output_hidden_states
        )

        if decoder_input_ids is None and input_ids is not None:
            decoder_input_ids = shift_tokens_right(input_ids, self.config.pad_token_id)

        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,
            position_ids=decoder_position_ids,
            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,
        )

    def build(self, input_shape=None):
        if self.built:
            return
        self.built = True
        # The shared/tied weights expect to be in the model base namespace
        # Adding "/" to the end (not the start!) of a tf.name_scope puts it in the root namespace rather than
        # the current one.
        with tf.name_scope(self.shared.load_weight_prefix + "/" + self.shared.name + "/"):
            self.shared.build(None)
        if getattr(self, "encoder", None) is not None:
            with tf.name_scope(self.encoder.name):
                self.encoder.build(None)
        if getattr(self, "decoder", None) is not None:
            with tf.name_scope(self.decoder.name):
                self.decoder.build(None)


@add_start_docstrings(
    "The bare MBART Model outputting raw hidden-states without any specific head on top.",
    MBART_START_DOCSTRING,
)
class TFMBartModel(TFMBartPreTrainedModel):
    def __init__(self, config: MBartConfig, *inputs, **kwargs):
        super().__init__(config, *inputs, **kwargs)

        self.model = TFMBartMainLayer(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(MBART_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: TFModelInputType = None,
        attention_mask: tf.Tensor | None = None,
        decoder_input_ids: tf.Tensor | None = None,
        decoder_attention_mask: tf.Tensor | None = None,
        decoder_position_ids: tf.Tensor | None = None,
        head_mask: tf.Tensor | None = None,
        decoder_head_mask: tf.Tensor | None = None,
        cross_attn_head_mask: tf.Tensor | None = None,
        encoder_outputs: Optional[Union[Tuple, TFBaseModelOutput]] = None,
        past_key_values: Tuple[Tuple[tf.Tensor]] | None = None,
        inputs_embeds: tf.Tensor | None = None,
        decoder_inputs_embeds: tf.Tensor | None = 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,
        **kwargs,
    ) -> Union[TFSeq2SeqModelOutput, Tuple[tf.Tensor]]:
        outputs = self.model(
            input_ids=input_ids,
            attention_mask=attention_mask,
            decoder_input_ids=decoder_input_ids,
            decoder_attention_mask=decoder_attention_mask,
            decoder_position_ids=decoder_position_ids,
            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.TFBartModel.serving_output
    def serving_output(self, output):
        pkv = tf.tuple(output.past_key_values)[1] if self.config.use_cache else None
        dec_hs = tf.convert_to_tensor(output.decoder_hidden_states) if self.config.output_hidden_states else None
        dec_attns = tf.convert_to_tensor(output.decoder_attentions) if self.config.output_attentions else None
        cross_attns = tf.convert_to_tensor(output.cross_attentions) if self.config.output_attentions else None
        enc_hs = tf.convert_to_tensor(output.encoder_hidden_states) if self.config.output_hidden_states else None
        enc_attns = tf.convert_to_tensor(output.encoder_attentions) if self.config.output_attentions else None

        return TFSeq2SeqModelOutput(
            last_hidden_state=output.last_hidden_state,
            past_key_values=pkv,
            decoder_hidden_states=dec_hs,
            decoder_attentions=dec_attns,
            cross_attentions=cross_attns,
            encoder_last_hidden_state=output.encoder_last_hidden_state,
            encoder_hidden_states=enc_hs,
            encoder_attentions=enc_attns,
        )

    def build(self, input_shape=None):
        if self.built:
            return
        self.built = True
        if getattr(self, "model", None) is not None:
            with tf.name_scope(self.model.name):
                self.model.build(None)


# Copied from transformers.models.bart.modeling_tf_bart.BiasLayer
class BiasLayer(keras.layers.Layer):
    """
    Bias as a layer. It is used for serialization purposes: `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 MBART Model with a language modeling head. Can be used for summarization, after fine-tuning the pretrained models.",
    MBART_START_DOCSTRING,
)
class TFMBartForConditionalGeneration(TFMBartPreTrainedModel, TFCausalLanguageModelingLoss):
    _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 = TFMBartMainLayer(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_output_embeddings(self):
        return self.get_input_embeddings()

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

    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"])

    @unpack_inputs
    @add_start_docstrings_to_model_forward(MBART_INPUTS_DOCSTRING)
    @replace_return_docstrings(output_type=TFSeq2SeqLMOutput, config_class=_CONFIG_FOR_DOC)
    @add_end_docstrings(MBART_GENERATION_EXAMPLE)
    def call(
        self,
        input_ids: TFModelInputType = None,
        attention_mask: tf.Tensor | None = None,
        decoder_input_ids: tf.Tensor | None = None,
        decoder_attention_mask: tf.Tensor | None = None,
        decoder_position_ids: tf.Tensor | None = None,
        head_mask: tf.Tensor | None = None,
        decoder_head_mask: tf.Tensor | None = None,
        cross_attn_head_mask: tf.Tensor | None = None,
        encoder_outputs: Optional[TFBaseModelOutput] = None,
        past_key_values: Tuple[Tuple[tf.Tensor]] = None,
        inputs_embeds: tf.Tensor | None = None,
        decoder_inputs_embeds: tf.Tensor | None = None,
        use_cache: Optional[bool] = None,
        output_attentions: Optional[bool] = None,
        output_hidden_states: Optional[bool] = None,
        return_dict: Optional[bool] = None,
        labels: tf.Tensor | None = None,
        training: Optional[bool] = False,
    ) -> Union[TFSeq2SeqLMOutput, Tuple[tf.Tensor]]:
        """
        labels (`tf.Tensor` of shape `(batch_size, sequence_length)`, *optional*):
            Labels for computing the masked language modeling loss. Indices should either be in `[0, ...,
            config.vocab_size]` or -100 (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]`.

        Returns:

        """

        if labels is not None:
            labels = tf.where(
                labels == self.config.pad_token_id,
                tf.cast(tf.fill(shape_list(labels), -100), labels.dtype),
                labels,
            )
            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)

        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,
            decoder_position_ids=decoder_position_ids,
            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
        )

    # Copied from transformers.models.bart.modeling_tf_bart.TFBartForConditionalGeneration.serving_output
    def serving_output(self, output):
        pkv = tf.tuple(output.past_key_values)[1] if self.config.use_cache else None
        dec_hs = tf.convert_to_tensor(output.decoder_hidden_states) if self.config.output_hidden_states else None
        dec_attns = tf.convert_to_tensor(output.decoder_attentions) if self.config.output_attentions else None
        cross_attns = tf.convert_to_tensor(output.cross_attentions) if self.config.output_attentions else None
        enc_hs = tf.convert_to_tensor(output.encoder_hidden_states) if self.config.output_hidden_states else None
        enc_attns = tf.convert_to_tensor(output.encoder_attentions) if self.config.output_attentions else None

        return TFSeq2SeqLMOutput(
            logits=output.logits,
            past_key_values=pkv,
            decoder_hidden_states=dec_hs,
            decoder_attentions=dec_attns,
            cross_attentions=cross_attns,
            encoder_last_hidden_state=output.encoder_last_hidden_state,
            encoder_hidden_states=enc_hs,
            encoder_attentions=enc_attns,
        )

    # Copied from transformers.models.bart.modeling_tf_bart.TFBartForConditionalGeneration.prepare_inputs_for_generation
    def prepare_inputs_for_generation(
        self,
        decoder_input_ids,
        past_key_values=None,
        attention_mask=None,
        decoder_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_key_values is used
        if past_key_values is not None:
            decoder_input_ids = decoder_input_ids[:, -1:]

        if decoder_attention_mask is not None:  # xla
            decoder_position_ids = tf.math.cumsum(decoder_attention_mask, axis=-1, exclusive=True)[:, -1:]
        elif past_key_values is not None:  # no xla + past_key_values
            decoder_position_ids = past_key_values[0][0].shape[2]
        else:  # no xla + no past_key_values
            decoder_position_ids = tf.range(decoder_input_ids.shape[1])

        return {
            "input_ids": None,  # encoder_outputs is defined. input_ids not needed
            "encoder_outputs": encoder_outputs,
            "past_key_values": past_key_values,
            "decoder_input_ids": decoder_input_ids,
            "attention_mask": attention_mask,
            "decoder_attention_mask": decoder_attention_mask,
            "decoder_position_ids": decoder_position_ids,
            "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 prepare_decoder_input_ids_from_labels(self, labels: tf.Tensor):
        return shift_tokens_right(labels, self.config.pad_token_id)

    def build(self, input_shape=None):
        if self.built:
            return
        self.built = True
        if getattr(self, "model", None) is not None:
            with tf.name_scope(self.model.name):
                self.model.build(None)
        if getattr(self, "bias_layer", None) is not None:
            with tf.name_scope(self.bias_layer.name):
                self.bias_layer.build(None)


__all__ = ["TFMBartForConditionalGeneration", "TFMBartModel", "TFMBartPreTrainedModel"]