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

from __future__ import annotations

import math
import random
from typing import Any, Optional, Tuple, Union

import numpy as np
import tensorflow as tf

from ...activations_tf import get_tf_activation

# Public API
from ...file_utils import (
    add_code_sample_docstrings,
    add_start_docstrings,
    add_start_docstrings_to_model_forward,
    replace_return_docstrings,
)
from ...modeling_tf_outputs import TFBaseModelOutputWithPastAndCrossAttentions, TFCausalLMOutputWithCrossAttentions
from ...modeling_tf_utils import (
    TFCausalLanguageModelingLoss,
    TFModelInputType,
    TFPreTrainedModel,
    TFSharedEmbeddings,
    get_initializer,
    keras,
    keras_serializable,
    unpack_inputs,
)
from ...tf_utils import check_embeddings_within_bounds, shape_list, stable_softmax
from ...utils import logging
from .configuration_xglm import XGLMConfig


logger = logging.get_logger(__name__)

_CHECKPOINT_FOR_DOC = "facebook/xglm-564M"
_CONFIG_FOR_DOC = "XGLMConfig"


LARGE_NEGATIVE = -1e8


def create_sinusoidal_positions(num_positions: int, embedding_dim: int, padding_idx: Optional[int]) -> tf.Tensor:
    half_dim = embedding_dim // 2
    emb = math.log(10000) / (half_dim - 1)
    emb = tf.exp(tf.range(half_dim, dtype=tf.float32) * -emb)
    emb = tf.expand_dims(tf.range(num_positions, dtype=tf.float32), axis=1) * tf.expand_dims(emb, axis=0)
    emb = tf.reshape(tf.concat([tf.sin(emb), tf.cos(emb)], axis=1), (num_positions, -1))
    if embedding_dim % 2 == 1:
        # zero pad
        emb = tf.concat([emb, tf.zeros((num_positions, 1))], axis=1)
    if padding_idx is not None:
        _padding_mask = tf.concat(
            [
                tf.ones((padding_idx, shape_list(emb)[1])),
                tf.zeros((1, shape_list(emb)[1])),
                tf.ones((shape_list(emb)[0] - padding_idx - 1, shape_list(emb)[1])),
            ],
            axis=0,
        )
        emb *= _padding_mask

    return tf.constant(emb, name="embed_positions")


def _create_position_ids_from_input_ids(
    input_ids: tf.Tensor, past_key_values_length: int, padding_idx: Optional[int]
) -> tf.Tensor:
    """
    Replace non-padding symbols with their position numbers. Position numbers begin at padding_idx+1. Padding symbols
    are ignored. This is modified from fairseq's `utils.make_positions`.
    """
    # The series of casts and type-conversions here are carefully balanced to both work with ONNX export and XLA.
    mask = tf.where(input_ids != padding_idx, 1, 0)
    incremental_indices = (tf.cast(tf.cumsum(mask, axis=1), dtype=mask.dtype) + past_key_values_length) * mask
    return tf.cast(incremental_indices, dtype=tf.int64) + padding_idx


def _create_position_ids_from_inputs_embeds(
    inputs_embeds: tf.Tensor, past_key_values_length: int, padding_idx: Optional[int]
) -> tf.Tensor:
    """
    Args:
    We are provided embeddings directly. We cannot infer which are padded so just generate sequential position ids.
        inputs_embeds: tf.Tensor
    Returns: tf.Tensor
    """
    input_shape = shape_list(inputs_embeds)[:-1]
    sequence_length = input_shape[1]

    position_ids = tf.range(padding_idx + 1, sequence_length + padding_idx + 1, dtype=tf.int64)

    return tf.broadcast_to(tf.expand_dims(position_ids, axis=0), input_shape) + past_key_values_length


# 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.TFBartAttention with Bart->XGLM
class TFXGLMAttention(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 TFXGLMDecoderLayer(keras.layers.Layer):
    def __init__(self, config: XGLMConfig, **kwargs: Any) -> None:
        super().__init__(**kwargs)
        self.embed_dim = config.d_model
        self.self_attn = TFXGLMAttention(
            embed_dim=self.embed_dim,
            num_heads=config.attention_heads,
            dropout=config.attention_dropout,
            is_decoder=True,
            name="self_attn",
        )
        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)

        if config.add_cross_attention:
            self.encoder_attn = TFXGLMAttention(
                embed_dim=self.embed_dim,
                num_heads=config.attention_heads,
                dropout=config.attention_dropout,
                is_decoder=True,
                name="encoder_attn",
            )
            self.encoder_attn_layer_norm = keras.layers.LayerNormalization(
                epsilon=1e-5, name="encoder_attn_layer_norm"
            )

        self.self_attn_layer_norm = keras.layers.LayerNormalization(epsilon=1e-5, name="self_attn_layer_norm")
        self.fc1 = keras.layers.Dense(config.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

    # Copied from transformers.models.mbart.modeling_tf_mbart.TFMBartDecoderLayer.call
    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, "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.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])
        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])


@keras_serializable
class TFXGLMMainLayer(keras.layers.Layer):
    config_class = XGLMConfig

    def __init__(
        self, config: XGLMConfig, embed_tokens: Optional[TFSharedEmbeddings] = None, *inputs, **kwargs: Any
    ) -> None:
        super().__init__(*inputs, **kwargs)

        self.config = config
        self.padding_idx = config.pad_token_id
        self.max_target_positions = config.max_position_embeddings
        self.embed_scale = math.sqrt(config.d_model) if config.scale_embedding else 1.0

        if embed_tokens is not None:
            self.embed_tokens = embed_tokens
        else:
            self.embed_tokens = TFSharedEmbeddings(
                config.vocab_size, config.d_model, self.padding_idx, name="embed_tokens"
            )

        self.offset = 2
        self._embed_positions_weights = create_sinusoidal_positions(
            num_positions=config.max_position_embeddings + self.offset,
            embedding_dim=config.d_model,
            padding_idx=config.pad_token_id,
        )

        self.dropout = keras.layers.Dropout(config.dropout)
        self.layers = [TFXGLMDecoderLayer(config, name=f"layers.{i}") for i in range(config.num_layers)]
        self.layerdrop = config.layerdrop
        self.layer_norm = keras.layers.LayerNormalization(epsilon=1e-5, name="layer_norm")

    def get_input_embeddings(self) -> TFSharedEmbeddings:
        return self.embed_tokens

    def set_input_embeddings(self, value: TFSharedEmbeddings) -> None:
        self.embed_tokens = value

    def _prepare_decoder_attention_mask(
        self,
        attention_mask: tf.Tensor | None,
        input_shape: tf.TensorShape,
        past_key_values_length: int,
    ) -> tf.Tensor:
        # create causal mask
        # [bsz, seq_len] -> [bsz, 1, tgt_seq_len, src_seq_len]
        combined_attention_mask = _make_causal_mask(input_shape, past_key_values_length)
        combined_attention_mask = tf.cond(
            input_shape[-1] > 1, lambda: combined_attention_mask, lambda: tf.ones_like(combined_attention_mask)
        )
        if attention_mask is None:
            return combined_attention_mask
        expand_attention_mask = _expand_mask(attention_mask, tgt_len=input_shape[-1])
        return expand_attention_mask + combined_attention_mask

    def embed_positions(self, position_ids: np.ndarray | tf.Tensor | None = None) -> tf.Tensor:
        position_ids += self.offset
        positions = tf.gather(self._embed_positions_weights, position_ids, axis=0)
        return positions

    @unpack_inputs
    def call(
        self,
        input_ids: TFModelInputType | None = None,
        attention_mask: np.ndarray | tf.Tensor | None = None,
        position_ids: np.ndarray | tf.Tensor | None = None,
        encoder_hidden_states: np.ndarray | tf.Tensor | None = None,
        encoder_attention_mask: np.ndarray | tf.Tensor | None = None,
        head_mask: np.ndarray | tf.Tensor | None = None,
        cross_attn_head_mask: np.ndarray | tf.Tensor | None = None,
        past_key_values: Optional[Tuple[Tuple[Union[np.ndarray, tf.Tensor]]]] = None,
        inputs_embeds: np.ndarray | 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: Any,
    ) -> Union[TFBaseModelOutputWithPastAndCrossAttentions, Tuple[tf.Tensor]]:
        output_attentions = output_attentions if output_attentions is not None else self.config.output_attentions
        output_hidden_states = (
            output_hidden_states if output_hidden_states is not None else self.config.output_hidden_states
        )
        use_cache = use_cache if use_cache is not None else self.config.use_cache
        return_dict = return_dict if return_dict is not None else self.config.use_return_dict

        # retrieve input_ids and inputs_embeds
        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 = tf.shape(input_ids)
            input_ids = tf.reshape(input_ids, (-1, input_shape[-1]))
        elif inputs_embeds is not None:
            input_shape = tf.shape(inputs_embeds)[:-1]
        else:
            raise ValueError("You have to specify either input_ids or inputs_embeds")

        past_key_values_length = past_key_values[0][0].shape[2] if past_key_values is not None else 0

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

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

        attention_mask = self._prepare_decoder_attention_mask(attention_mask, input_shape, past_key_values_length)

        # expand encoder attention mask
        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])

        # embed positions
        positions = self.embed_positions(position_ids)

        hidden_states = tf.cast(inputs_embeds, dtype=tf.float32) + 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_attentions = () if (output_attentions and encoder_hidden_states is not None) else None
        next_decoder_cache = () 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=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:
                next_decoder_cache += (present_key_value,)

            if output_attentions:
                all_self_attns += (layer_self_attn,)

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

        hidden_states = self.layer_norm(hidden_states)

        # add hidden states from the last decoder layer
        if output_hidden_states:
            all_hidden_states += (hidden_states,)

        next_cache = next_decoder_cache if use_cache else None
        if not return_dict:
            return tuple(
                v
                for v in [hidden_states, next_cache, all_hidden_states, all_self_attns, all_cross_attentions]
                if v is not None
            )
        return TFBaseModelOutputWithPastAndCrossAttentions(
            last_hidden_state=hidden_states,
            past_key_values=next_cache,
            hidden_states=all_hidden_states,
            attentions=all_self_attns,
            cross_attentions=all_cross_attentions,
        )

    def build(self, input_shape=None):
        if self.built:
            return
        self.built = True
        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, "embed_tokens", None) is not None:
            with tf.name_scope(self.embed_tokens.name):
                self.embed_tokens.build(None)
        if getattr(self, "layers", None) is not None:
            for layer in self.layers:
                with tf.name_scope(layer.name):
                    layer.build(None)


class TFXGLMPreTrainedModel(TFPreTrainedModel):
    config_class = XGLMConfig
    base_model_prefix = "model"


XGLM_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 ([`XGLMConfig`]): 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.
"""

XGLM_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)
        position_ids (`tf.Tensor` or `Numpy array` of shape `(batch_size, sequence_length)`, *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)
        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 `(num_layers, 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**.

        cross_attn_head_mask (`tf.Tensor` of shape `(num_layers, 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.num_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).
"""


@add_start_docstrings(
    "The bare XGLM Model transformer outputting raw hidden-states without any specific head on top.",
    XGLM_START_DOCSTRING,
)
class TFXGLMModel(TFXGLMPreTrainedModel):
    """
    Transformer decoder consisting of *config.num_layers* layers. Each layer is a [`TFXGLMDecoderLayer`]

    Args:
        config: XGLMConfig
        embed_tokens: [TFSharedEmbeddings]: output embedding
    """

    def __init__(
        self, config: XGLMConfig, embed_tokens: Optional[TFSharedEmbeddings] = None, *inputs: Any, **kwargs: Any
    ) -> None:
        super().__init__(config, *inputs, **kwargs)

        self.model = TFXGLMMainLayer(config, embed_tokens=embed_tokens, name="model")

    @unpack_inputs
    @add_start_docstrings_to_model_forward(XGLM_INPUTS_DOCSTRING)
    @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,
        position_ids: np.ndarray | tf.Tensor | None = None,
        encoder_hidden_states: np.ndarray | tf.Tensor | None = None,
        encoder_attention_mask: np.ndarray | tf.Tensor | None = None,
        head_mask: np.ndarray | tf.Tensor | None = None,
        cross_attn_head_mask: np.ndarray | tf.Tensor | None = None,
        past_key_values: Optional[Tuple[Tuple[Union[np.ndarray, tf.Tensor]]]] = None,
        inputs_embeds: np.ndarray | 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: Any,
    ) -> Union[TFBaseModelOutputWithPastAndCrossAttentions, Tuple[tf.Tensor]]:
        outputs = self.model(
            input_ids=input_ids,
            attention_mask=attention_mask,
            encoder_hidden_states=encoder_hidden_states,
            encoder_attention_mask=encoder_attention_mask,
            head_mask=head_mask,
            cross_attn_head_mask=cross_attn_head_mask,
            past_key_values=past_key_values,
            inputs_embeds=inputs_embeds,
            use_cache=use_cache,
            output_attentions=output_attentions,
            output_hidden_states=output_hidden_states,
            return_dict=return_dict,
            training=training,
        )

        return outputs

    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)


@add_start_docstrings(
    """
    The XGLM Model transformer with a language modeling head on top (linear layer with weights tied to the input
    embeddings).
    """,
    XGLM_START_DOCSTRING,
)
class TFXGLMForCausalLM(TFXGLMPreTrainedModel, TFCausalLanguageModelingLoss):
    base_model_prefix = "model"
    _keys_to_ignore_on_load_missing = [
        r"model.embed_positions.weights",
        r"lm_head.weight",
    ]
    _keys_to_ignore_on_save = [
        r"model.embed_positions.weights",
    ]

    def __init__(
        self, config: XGLMConfig, embed_tokens: Optional[TFSharedEmbeddings] = None, *inputs: Any, **kwargs: Any
    ) -> None:
        super().__init__(config, *inputs, **kwargs)

        self.model = TFXGLMMainLayer(config, embed_tokens=embed_tokens, name="model")
        self.lm_head = keras.layers.Dense(
            config.vocab_size,
            use_bias=False,
            kernel_initializer=get_initializer(config.init_std),
            name="lm_head",
        )
        self.config = config

    def get_output_embeddings(self):
        return self.lm_head

    def set_output_embeddings(self, new_embeddings):
        self.lm_head = new_embeddings

    def prepare_inputs_for_generation(self, inputs, past_key_values=None, use_cache=None, **kwargs):
        # only last token for inputs_ids if past is defined in kwargs
        if past_key_values:
            inputs = tf.expand_dims(inputs[:, -1], -1)

        position_ids = kwargs.get("position_ids", None)
        attention_mask = kwargs.get("attention_mask", None)

        if attention_mask is not None and position_ids is None:
            position_ids = tf.math.cumsum(attention_mask, axis=-1, exclusive=True)
            if past_key_values:
                position_ids = tf.expand_dims(position_ids[:, -1], -1)

        return {
            "input_ids": inputs,
            "attention_mask": attention_mask,
            "position_ids": position_ids,
            "past_key_values": past_key_values,
            "use_cache": use_cache,
        }

    @unpack_inputs
    @add_start_docstrings_to_model_forward(XGLM_INPUTS_DOCSTRING)
    @replace_return_docstrings(output_type=TFCausalLMOutputWithCrossAttentions, config_class=_CONFIG_FOR_DOC)
    @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,
        position_ids: np.ndarray | tf.Tensor | None = None,
        encoder_hidden_states: np.ndarray | tf.Tensor | None = None,
        encoder_attention_mask: np.ndarray | tf.Tensor | None = None,
        head_mask: np.ndarray | tf.Tensor | None = None,
        cross_attn_head_mask: np.ndarray | tf.Tensor | None = None,
        past_key_values: Optional[Tuple[Tuple[Union[np.ndarray, tf.Tensor]]]] = None,
        inputs_embeds: np.ndarray | tf.Tensor | None = None,
        labels: np.ndarray | 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: Any,
    ) -> Union[TFCausalLMOutputWithCrossAttentions, Tuple[tf.Tensor]]:
        r"""
        labels (`np.ndarray` or `tf.Tensor` of shape `(batch_size, sequence_length)`, *optional*):
            Labels for language modeling. Note that the labels **are shifted** inside the model, i.e. you can set
            `labels = input_ids` Indices are selected in `[-100, 0, ..., config.vocab_size]` All labels set to `-100`
            are ignored (masked), the loss is only computed for labels in `[0, ..., config.vocab_size]`
        """

        outputs = self.model(
            input_ids=input_ids,
            attention_mask=attention_mask,
            position_ids=position_ids,
            encoder_hidden_states=encoder_hidden_states,
            encoder_attention_mask=encoder_attention_mask,
            head_mask=head_mask,
            cross_attn_head_mask=cross_attn_head_mask,
            past_key_values=past_key_values,
            inputs_embeds=inputs_embeds,
            use_cache=use_cache,
            output_attentions=output_attentions,
            output_hidden_states=output_hidden_states,
            return_dict=return_dict,
            training=training,
        )
        hidden_states = outputs[0]
        lm_logits = self.lm_head(hidden_states)

        loss = None
        if labels is not None:
            # shift labels to the left and cut last logit token
            labels = tf.concat(
                [labels[:, 1:], tf.fill((labels.shape[0], 1), tf.cast(-100, labels.dtype))],
                axis=-1,
            )
            loss = self.hf_compute_loss(labels, lm_logits)

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

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

    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, "lm_head", None) is not None:
            with tf.name_scope(self.lm_head.name):
                self.lm_head.build([None, None, self.config.hidden_size])

    def tf_to_pt_weight_rename(self, tf_weight):
        if tf_weight == "lm_head.weight":
            return tf_weight, "model.embed_tokens.weight"
        else:
            return (tf_weight,)


__all__ = ["TFXGLMForCausalLM", "TFXGLMModel", "TFXGLMPreTrainedModel"]