# 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.
""" TensorFlow Speech2Text model."""


from __future__ import annotations

import random
from typing import Optional, Tuple, Union

import numpy as np
import tensorflow as tf

from ...activations_tf import get_tf_activation, glu
from ...modeling_tf_outputs import (
    TFBaseModelOutput,
    TFBaseModelOutputWithPastAndCrossAttentions,
    TFSeq2SeqLMOutput,
    TFSeq2SeqModelOutput,
)
from ...modeling_tf_utils import (
    TFCausalLanguageModelingLoss,
    TFModelInputType,
    TFPreTrainedModel,
    TFSharedEmbeddings,
    keras_serializable,
    unpack_inputs,
)
from ...tf_utils import check_embeddings_within_bounds, shape_list, stable_softmax
from ...utils import (
    add_code_sample_docstrings,
    add_start_docstrings,
    add_start_docstrings_to_model_forward,
    logging,
    replace_return_docstrings,
)
from .configuration_speech_to_text import Speech2TextConfig


logger = logging.get_logger(__name__)

_CONFIG_FOR_DOC = "Speech2TextConfig"
_CHECKPOINT_FOR_DOC = "facebook/s2t-small-librispeech-asr"


TF_SPEECH_TO_TEXT_PRETRAINED_MODEL_ARCHIVE_LIST = [
    "facebook/s2t-small-librispeech-asr",
    # See all Speech2Text models at https://huggingface.co/models?filter=speech_to_text
]


LARGE_NEGATIVE = -1e8


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

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

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

    return shifted_input_ids


# 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


class TFConv1dSubsampler(tf.keras.layers.Layer):
    """
    Convolutional subsampler: a stack of 1D convolution (along temporal dimension) followed by non-linear activation
    via gated linear units (https://arxiv.org/abs/1911.08460)
    """

    def __init__(self, config: Speech2TextConfig, **kwargs):
        super().__init__(**kwargs)
        self.config = config
        self.num_layers = config.num_conv_layers
        self.in_channels = config.input_feat_per_channel * config.input_channels
        self.mid_channels = config.conv_channels
        self.out_channels = config.d_model
        self.kernel_sizes = config.conv_kernel_sizes

        self.conv_layers = [
            tf.keras.layers.Conv1D(
                filters=self.mid_channels if i < self.num_layers - 1 else self.out_channels * 2,
                kernel_size=k,
                strides=2,
                name=f"conv_layers.{i}",
            )
            for i, k in enumerate(self.kernel_sizes)
        ]

    def call(self, input_features: tf.Tensor) -> tf.Tensor:
        # TF Conv1D assumes Batch x Time x Channels, same as the input
        hidden_states = tf.cast(input_features, tf.float32)
        for i, conv in enumerate(self.conv_layers):
            # equivalent to `padding=k // 2` on PT's `nn.Conv1d`
            pad_len = self.kernel_sizes[i] // 2
            hidden_shapes = shape_list(hidden_states)
            hidden_states = tf.concat(
                (
                    tf.zeros((hidden_shapes[0], pad_len, hidden_shapes[2])),
                    hidden_states,
                    tf.zeros((hidden_shapes[0], pad_len, hidden_shapes[2])),
                ),
                axis=1,
            )

            hidden_states = conv(hidden_states)
            hidden_states = glu(hidden_states, axis=2)  # GLU over the Channel dimension
        return hidden_states


class TFSpeech2TextSinusoidalPositionalEmbedding(tf.keras.layers.Layer):
    """This module produces sinusoidal positional embeddings of any length."""

    def __init__(self, num_positions: int, embedding_dim: int, padding_idx: Optional[int] = None, **kwargs):
        super().__init__(**kwargs)
        self.offset = 2
        self.embedding_dim = embedding_dim
        self.padding_idx = padding_idx
        self.embedding_weights = self._get_embedding(num_positions + self.offset, embedding_dim, padding_idx)

    @staticmethod
    def _get_embedding(num_embeddings: int, embedding_dim: int, padding_idx: Optional[int] = None) -> tf.Tensor:
        """
        Build sinusoidal embeddings. This matches the implementation in tensor2tensor, but differs slightly from the
        description in Section 3.5 of "Attention Is All You Need".
        """
        half_dim = embedding_dim // 2
        emb = tf.math.log(10000.0) / (half_dim - 1)
        emb = tf.math.exp(tf.range(half_dim, dtype=tf.float32) * -emb)
        emb = tf.expand_dims(tf.range(num_embeddings, dtype=tf.float32), axis=1) * tf.expand_dims(emb, axis=0)
        emb = tf.reshape(tf.concat([tf.math.sin(emb), tf.math.cos(emb)], axis=1), shape=[num_embeddings, -1])
        if embedding_dim % 2 == 1:
            # zero pad
            emb = tf.concat([emb, tf.zeros(num_embeddings, 1)], axis=1)
        if padding_idx is not None:
            emb = tf.concat([emb[:padding_idx, :], tf.zeros((1, tf.shape(emb)[1])), emb[padding_idx + 1 :, :]], axis=0)
        return emb

    def call(self, input_ids: tf.Tensor, past_key_values_length: int = 0) -> tf.Tensor:
        bsz, seq_len = shape_list(input_ids)
        # Create the position ids from the input token ids. Any padded tokens remain padded.
        position_ids = self.create_position_ids_from_input_ids(input_ids, self.padding_idx, past_key_values_length)

        # Matt: The PyTorch code does a lot of work to cache the embeddings, setting the cached values as a
        # model attribute in the forward pass. This is extremely forbidden in TF, which wants forward calls to be
        # idempotent. TF doesn't need that caching anyway, since it can just store constants during compilation,
        # so we just remove all of that code.
        embeddings = self._get_embedding(
            self.padding_idx + 1 + seq_len + self.offset + past_key_values_length, self.embedding_dim, self.padding_idx
        )
        return tf.reshape(tf.gather(embeddings, tf.reshape(position_ids, (-1,)), axis=0), (bsz, seq_len, -1))

    @staticmethod
    def create_position_ids_from_input_ids(
        input_ids: tf.Tensor, padding_idx: int, past_key_values_length: Optional[int] = 0
    ) -> 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`.

        Args:
            x: tf.Tensor x:
        Returns: tf.Tensor
        """
        mask = tf.cast(tf.math.not_equal(input_ids, padding_idx), dtype=tf.int32)
        incremental_indices = (tf.math.cumsum(mask, axis=1) + past_key_values_length) * mask
        return tf.cast(incremental_indices, dtype=tf.int64) + padding_idx


# Copied from transformers.models.bart.modeling_tf_bart.TFBartAttention with Bart->Speech2Text
class TFSpeech2TextAttention(tf.keras.layers.Layer):
    """Multi-headed attention from "Attention Is All You Need"""

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

        self.num_heads = num_heads
        self.dropout = tf.keras.layers.Dropout(dropout)
        self.head_dim = embed_dim // num_heads
        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 = tf.keras.layers.Dense(embed_dim, use_bias=bias, name="k_proj")
        self.q_proj = tf.keras.layers.Dense(embed_dim, use_bias=bias, name="q_proj")
        self.v_proj = tf.keras.layers.Dense(embed_dim, use_bias=bias, name="v_proj")
        self.out_proj = tf.keras.layers.Dense(embed_dim, use_bias=bias, name="out_proj")

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

    def call(
        self,
        hidden_states: tf.Tensor,
        key_value_states: tf.Tensor | None = None,
        past_key_value: Tuple[Tuple[tf.Tensor]] | None = None,
        attention_mask: tf.Tensor | None = None,
        layer_head_mask: tf.Tensor | None = None,
        training: 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


class TFSpeech2TextEncoderLayer(tf.keras.layers.Layer):
    def __init__(self, config: Speech2TextConfig, **kwargs):
        super().__init__(**kwargs)
        self.embed_dim = config.d_model
        self.self_attn = TFSpeech2TextAttention(
            self.embed_dim, config.encoder_attention_heads, dropout=config.attention_dropout, name="self_attn"
        )
        self.self_attn_layer_norm = tf.keras.layers.LayerNormalization(epsilon=1e-5, name="self_attn_layer_norm")
        self.dropout = tf.keras.layers.Dropout(config.dropout)
        self.activation_fn = get_tf_activation(config.activation_function)
        self.activation_dropout = tf.keras.layers.Dropout(config.activation_dropout)
        self.fc1 = tf.keras.layers.Dense(config.encoder_ffn_dim, name="fc1")
        self.fc2 = tf.keras.layers.Dense(self.embed_dim, name="fc2")
        self.final_layer_norm = tf.keras.layers.LayerNormalization(epsilon=1e-5, name="final_layer_norm")

    def call(
        self, hidden_states: tf.Tensor, attention_mask: tf.Tensor, layer_head_mask: tf.Tensor, training: 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,
            training=training,
        )

        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


class TFSpeech2TextDecoderLayer(tf.keras.layers.Layer):
    def __init__(self, config: Speech2TextConfig, **kwargs):
        super().__init__(**kwargs)
        self.embed_dim = config.d_model

        self.self_attn = TFSpeech2TextAttention(
            embed_dim=self.embed_dim,
            num_heads=config.decoder_attention_heads,
            dropout=config.attention_dropout,
            name="self_attn",
            is_decoder=True,
        )
        self.dropout = tf.keras.layers.Dropout(config.dropout)
        self.activation_fn = get_tf_activation(config.activation_function)
        self.activation_dropout = tf.keras.layers.Dropout(config.activation_dropout)

        self.self_attn_layer_norm = tf.keras.layers.LayerNormalization(epsilon=1e-5, name="self_attn_layer_norm")
        self.encoder_attn = TFSpeech2TextAttention(
            self.embed_dim,
            config.decoder_attention_heads,
            dropout=config.attention_dropout,
            name="encoder_attn",
            is_decoder=True,
        )
        self.encoder_attn_layer_norm = tf.keras.layers.LayerNormalization(epsilon=1e-5, name="encoder_attn_layer_norm")
        self.fc1 = tf.keras.layers.Dense(config.decoder_ffn_dim, name="fc1")
        self.fc2 = tf.keras.layers.Dense(self.embed_dim, name="fc2")
        self.final_layer_norm = tf.keras.layers.LayerNormalization(epsilon=1e-5, name="final_layer_norm")

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


class TFSpeech2TextPreTrainedModel(TFPreTrainedModel):
    config_class = Speech2TextConfig
    base_model_prefix = "model"
    main_input_name = "input_features"
    _keys_to_ignore_on_load_unexpected = [r"encoder.embed_positions.weights"]

    def _get_feat_extract_output_lengths(self, input_lengths: tf.Tensor):
        """
        Computes the output length of the convolutional layers
        """
        for _ in range(self.config.num_conv_layers):
            input_lengths = (input_lengths - 1) // 2 + 1

        return input_lengths

    @property
    def input_signature(self):
        return {
            "input_features": tf.TensorSpec(
                (None, None, self.config.input_feat_per_channel * self.config.input_channels),
                tf.float32,
                name="input_features",
            ),
            "attention_mask": tf.TensorSpec((None, None), tf.int32, name="attention_mask"),
            "decoder_input_ids": tf.TensorSpec((None, None), tf.int32, name="decoder_input_ids"),
            "decoder_attention_mask": tf.TensorSpec((None, None), tf.int32, name="decoder_attention_mask"),
        }


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

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

    <Tip>

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

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

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

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

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

    </Tip>

    Parameters:
        config ([`Speech2TextConfig`]):
            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.
"""


SPEECH_TO_TEXT_INPUTS_DOCSTRING = r"""
    Args:
        input_features (`tf.Tensor` of shape `(batch_size, sequence_length, feature_size)`):
            Float values of fbank features extracted from the raw speech waveform. Raw speech waveform can be obtained
            by loading a `.flac` or `.wav` audio file into an array of type `List[float]` or a `numpy.ndarray`, *e.g.*
            via the soundfile library (`pip install soundfile`). To prepare the array into `input_features`, the
            [`AutoFeatureExtractor`] should be used for extracting the fbank features, padding and conversion into a
            tensor of floats. See [`~Speech2TextFeatureExtractor.__call__`]
        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 [`Speech2TextTokenizer`]. See [`PreTrainedTokenizer.encode`] and
            [`PreTrainedTokenizer.__call__`] for details.

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

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

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

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

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

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

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

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

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


@keras_serializable
class TFSpeech2TextEncoder(tf.keras.layers.Layer):
    config_class = Speech2TextConfig
    """
    Transformer encoder consisting of *config.encoder_layers* self attention layers. Each layer is a
    [`TFSpeech2TextEncoderLayer`].

    Args:
        config: Speech2TextConfig
    """

    def __init__(self, config: Speech2TextConfig, **kwargs):
        super().__init__(**kwargs)
        self.config = config

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

        embed_dim = config.d_model
        self.padding_idx = config.pad_token_id
        self.max_source_positions = config.max_source_positions
        self.embed_scale = tf.math.sqrt(float(embed_dim)) if config.scale_embedding else 1.0

        self.conv = TFConv1dSubsampler(config, name="conv")

        self.embed_positions = TFSpeech2TextSinusoidalPositionalEmbedding(
            num_positions=config.max_source_positions,
            embedding_dim=embed_dim,
            padding_idx=self.padding_idx,
            name="embed_positions",
        )
        self.layers = [TFSpeech2TextEncoderLayer(config, name=f"layers.{i}") for i in range(config.encoder_layers)]
        self.layer_norm = tf.keras.layers.LayerNormalization(epsilon=1e-5, name="layer_norm")

    def _get_feat_extract_output_lengths(self, input_lengths: tf.Tensor):
        """
        Computes the output length of the convolutional layers
        """
        for _ in range(self.config.num_conv_layers):
            input_lengths = (input_lengths - 1) // 2 + 1

        return input_lengths

    def _get_feature_vector_attention_mask(self, feature_vector_length, attention_mask):
        # generate creates 3D attention mask, because of the shape of input_features
        # convert it to 2D if thats the case
        if len(attention_mask.shape) > 2:
            attention_mask = attention_mask[:, :, -1]

        subsampled_lengths = self._get_feat_extract_output_lengths(tf.math.reduce_sum(attention_mask, -1))
        bsz = shape_list(attention_mask)[0]
        indices = tf.concat(
            (
                tf.expand_dims(tf.range(bsz, dtype=attention_mask.dtype), -1),
                tf.expand_dims(subsampled_lengths - 1, -1),
            ),
            axis=-1,
        )
        attention_mask = tf.scatter_nd(indices=indices, updates=tf.ones(bsz), shape=[bsz, feature_vector_length])
        attention_mask = tf.cast(tf.reverse(tf.math.cumsum(tf.reverse(attention_mask, [-1]), -1), [-1]), tf.int64)
        return attention_mask

    @unpack_inputs
    def call(
        self,
        input_features=None,
        attention_mask=None,
        head_mask=None,
        output_attentions=None,
        output_hidden_states=None,
        return_dict=None,
        training=False,
    ):
        """
        Args:
            input_features (`tf.Tensor` of shape `(batch_size, sequence_length, feature_size)`):
                Float values of fbank features extracted from the raw speech waveform. Raw speech waveform can be
                obtained by loading a `.flac` or `.wav` audio file into an array of type `List[float]` or a
                `numpy.ndarray`, *e.g.* via the soundfile library (`pip install soundfile`). To prepare the array into
                `input_features`, the [`AutoFeatureExtractor`] should be used for extracting the fbank features,
                padding and conversion into a tensor of floats. See [`~Speech2TextFeatureExtractor.__call__`]
            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**.

            output_attentions (`bool`, *optional*):
                Whether or not to return the attentions tensors of all attention layers. See `attentions` under
                returned tensors for more detail.
            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.
            return_dict (`bool`, *optional*):
                Whether or not to return a [`~utils.ModelOutput`] instead of a plain tuple.
        """
        if input_features is None:
            raise ValueError("You have to specify input_features")

        inputs_embeds = self.conv(input_features)
        inputs_embeds = self.embed_scale * inputs_embeds

        # subsample attention mask if necessary
        if attention_mask is not None:
            attention_mask = self._get_feature_vector_attention_mask(tf.shape(inputs_embeds)[1], attention_mask)
            padding_mask = tf.cast(tf.math.not_equal(attention_mask, 1), tf.int64)
        else:
            padding_mask = tf.zeros(tf.shape(inputs_embeds)[:-1], dtype=tf.int64)

        embed_pos = self.embed_positions(padding_mask)

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

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

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

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

        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,
                training=training,
            )

            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
        )


@keras_serializable
class TFSpeech2TextDecoder(tf.keras.layers.Layer):
    config_class = Speech2TextConfig
    """
    Transformer decoder consisting of *config.decoder_layers* layers. Each layer is a [`TFSpeech2TextDecoderLayer`]

    Args:
        config: Speech2TextConfig
    """

    def __init__(self, config: Speech2TextConfig, **kwargs):
        super().__init__(**kwargs)
        self.config = config
        self.layerdrop = config.decoder_layerdrop
        self.padding_idx = config.pad_token_id
        self.max_target_positions = config.max_target_positions
        self.embed_scale = tf.math.sqrt(float(config.d_model)) if config.scale_embedding else 1.0

        self.embed_tokens = TFSharedEmbeddings(config.vocab_size, config.d_model, name="embed_tokens")

        self.embed_positions = TFSpeech2TextSinusoidalPositionalEmbedding(
            num_positions=config.max_target_positions,
            embedding_dim=config.d_model,
            padding_idx=self.padding_idx,
            name="embed_positions",
        )

        self.layers = [TFSpeech2TextDecoderLayer(config, name=f"layers.{i}") for i in range(config.decoder_layers)]
        self.layer_norm = tf.keras.layers.LayerNormalization(epsilon=1e-5, name="layer_norm")

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

    def get_embed_tokens(self):
        return self.embed_tokens

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

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

                Indices can be obtained using [`Speech2TextTokenizer`]. See [`PreTrainedTokenizer.encode`] and
                [`PreTrainedTokenizer.__call__`] for details.

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

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

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

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

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

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

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

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

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

                If `past_key_values` are used, the user can optionally input only the last `decoder_input_ids` (those
                that don't have their past key value states given to this model) of shape `(batch_size, 1)` instead of
                all `decoder_input_ids` of shape `(batch_size, sequence_length)`. inputs_embeds (`tf.Tensor` of shape
                `(batch_size, sequence_length, hidden_size)`, *optional*): Optionally, instead of passing `input_ids`
                you can choose to directly pass an embedded representation. This is useful if you want more control
                over how to convert `input_ids` indices into associated vectors than the model's internal embedding
                lookup matrix.
            output_attentions (`bool`, *optional*):
                Whether or not to return the attentions tensors of all attention layers. See `attentions` under
                returned tensors for more detail.
            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.
            return_dict (`bool`, *optional*):
                Whether or not to return a [`~utils.ModelOutput`] instead of a plain tuple.
        """

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

        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
        else:
            inputs_embeds = 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])

        # 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(input_ids, past_key_values_length=past_key_values_length)

        hidden_states = inputs_embeds + 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
        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
            cross_attn_layer_head_mask = cross_attn_head_mask[idx] if cross_attn_head_mask 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_layer_head_mask,
                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_attns += (layer_cross_attn,)

        hidden_states = self.layer_norm(hidden_states)
        if output_hidden_states:
            all_hidden_states += (hidden_states,)

        next_cache = next_decoder_cache if use_cache else None

        if not return_dict:
            return hidden_states, next_cache, all_hidden_states, all_self_attns, all_cross_attns
        else:
            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_attns,
            )


@keras_serializable
class TFSpeech2TextMainLayer(tf.keras.layers.Layer):
    config_class = Speech2TextConfig

    def __init__(self, config: Speech2TextConfig, **kwargs):
        super().__init__(**kwargs)
        self.config = config

        self.encoder = TFSpeech2TextEncoder(config, name="encoder")
        self.decoder = TFSpeech2TextDecoder(config, name="decoder")

    def get_input_embeddings(self):
        return self.decoder.embed_tokens

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

    @unpack_inputs
    def call(
        self,
        input_features=None,
        attention_mask=None,
        decoder_input_ids=None,
        decoder_attention_mask=None,
        head_mask=None,
        decoder_head_mask=None,
        cross_attn_head_mask=None,
        encoder_outputs=None,
        past_key_values=None,
        decoder_inputs_embeds=None,
        use_cache=None,
        output_attentions=None,
        output_hidden_states=None,
        return_dict=None,
        training=False,
        **kwargs,
    ):
        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

        if encoder_outputs is None:
            encoder_outputs = self.encoder(
                input_features=input_features,
                attention_mask=attention_mask,
                head_mask=head_mask,
                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()

        # downsample encoder attention mask
        if attention_mask is not None:
            encoder_attention_mask = self.encoder._get_feature_vector_attention_mask(
                tf.shape(encoder_outputs[0])[1], attention_mask
            )
        else:
            encoder_attention_mask = None

        # decoder outputs consists of (dec_features, past_key_value, dec_hidden, dec_attn)
        decoder_outputs = self.decoder(
            input_ids=decoder_input_ids,
            attention_mask=decoder_attention_mask,
            encoder_hidden_states=encoder_outputs[0],
            encoder_attention_mask=encoder_attention_mask,
            head_mask=decoder_head_mask,
            cross_attn_head_mask=cross_attn_head_mask,
            past_key_values=past_key_values,
            inputs_embeds=decoder_inputs_embeds,
            use_cache=use_cache,
            output_attentions=output_attentions,
            output_hidden_states=output_hidden_states,
            return_dict=return_dict,
            training=training,
        )

        if not return_dict:
            return decoder_outputs + encoder_outputs

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


@add_start_docstrings(
    "The bare Speech2Text Model outputting raw hidden-states without any specific head on top.",
    SPEECH_TO_TEXT_START_DOCSTRING,
)
class TFSpeech2TextModel(TFSpeech2TextPreTrainedModel):
    def __init__(self, config: Speech2TextConfig, *inputs, **kwargs):
        super().__init__(config, *inputs, **kwargs)

        self.model = TFSpeech2TextMainLayer(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(SPEECH_TO_TEXT_INPUTS_DOCSTRING)
    @add_code_sample_docstrings(
        checkpoint=_CHECKPOINT_FOR_DOC,
        output_type=TFSeq2SeqModelOutput,
        config_class=_CONFIG_FOR_DOC,
    )
    def call(
        self,
        input_features: TFModelInputType | None = None,
        attention_mask: np.ndarray | tf.Tensor | None = None,
        decoder_input_ids: np.ndarray | tf.Tensor | None = None,
        decoder_attention_mask: np.ndarray | tf.Tensor | None = None,
        head_mask: np.ndarray | tf.Tensor | None = None,
        decoder_head_mask: np.ndarray | tf.Tensor | None = None,
        cross_attn_head_mask: np.ndarray | tf.Tensor | None = None,
        encoder_outputs: np.ndarray | tf.Tensor | None = None,
        past_key_values: Optional[Tuple[Tuple[Union[np.ndarray, tf.Tensor]]]] = None,
        decoder_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: bool = False,
        **kwargs,
    ) -> Union[Tuple, TFSeq2SeqModelOutput]:
        outputs = self.model(
            input_features=input_features,
            attention_mask=attention_mask,
            decoder_input_ids=decoder_input_ids,
            decoder_attention_mask=decoder_attention_mask,
            head_mask=head_mask,
            decoder_head_mask=decoder_head_mask,
            cross_attn_head_mask=cross_attn_head_mask,
            encoder_outputs=encoder_outputs,
            past_key_values=past_key_values,
            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

    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,
        )


@add_start_docstrings(
    "The Speech2Text Model with a language modeling head. Can be used for summarization.",
    SPEECH_TO_TEXT_START_DOCSTRING,
)
class TFSpeech2TextForConditionalGeneration(TFSpeech2TextPreTrainedModel, TFCausalLanguageModelingLoss):
    def __init__(self, config: Speech2TextConfig):
        super().__init__(config)
        self.model = TFSpeech2TextMainLayer(config, name="model")
        self.lm_head = tf.keras.layers.Dense(self.config.vocab_size, use_bias=False, name="lm_head")
        # TODO (Joao): investigate why Speech2Text has numerical issues in XLA generate
        self.supports_xla_generation = False

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

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

    def resize_token_embeddings(self, new_num_tokens: int) -> tf.Variable:
        new_embeddings = super().resize_token_embeddings(new_num_tokens)
        return new_embeddings

    def get_output_embeddings(self):
        return self.lm_head

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

    @unpack_inputs
    @add_start_docstrings_to_model_forward(SPEECH_TO_TEXT_INPUTS_DOCSTRING)
    @replace_return_docstrings(output_type=TFSeq2SeqLMOutput, config_class=_CONFIG_FOR_DOC)
    def call(
        self,
        input_features: TFModelInputType | None = None,
        attention_mask: np.ndarray | tf.Tensor | None = None,
        decoder_input_ids: np.ndarray | tf.Tensor | None = None,
        decoder_attention_mask: np.ndarray | tf.Tensor | None = None,
        head_mask: np.ndarray | tf.Tensor | None = None,
        decoder_head_mask: np.ndarray | tf.Tensor | None = None,
        cross_attn_head_mask: np.ndarray | tf.Tensor | None = None,
        encoder_outputs: np.ndarray | tf.Tensor | None = None,
        past_key_values: Optional[Tuple[Tuple[Union[np.ndarray, tf.Tensor]]]] = None,
        decoder_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,
    ) -> Union[Tuple, TFSeq2SeqLMOutput]:
        r"""
        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:

        Example:

        ```python
        >>> import tensorflow as tf
        >>> from transformers import Speech2TextProcessor, TFSpeech2TextForConditionalGeneration
        >>> from datasets import load_dataset
        >>> import soundfile as sf

        >>> model = TFSpeech2TextForConditionalGeneration.from_pretrained(
        ...     "facebook/s2t-small-librispeech-asr", from_pt=True
        ... )
        >>> processor = Speech2TextProcessor.from_pretrained("facebook/s2t-small-librispeech-asr")


        >>> def map_to_array(batch):
        ...     speech, _ = sf.read(batch["file"])
        ...     batch["speech"] = speech
        ...     return batch


        >>> ds = load_dataset("hf-internal-testing/librispeech_asr_dummy", "clean", split="validation")
        >>> ds = ds.map(map_to_array)
        >>> ds.set_format(type="tf")

        >>> input_features = processor(
        ...     ds["speech"][0], sampling_rate=16000, return_tensors="tf"
        ... ).input_features  # Batch size 1
        >>> generated_ids = model.generate(input_features)

        >>> transcription = processor.batch_decode(generated_ids)
        ```"""
        return_dict = return_dict if return_dict is not None else self.config.use_return_dict

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

        outputs = self.model(
            input_features=input_features,
            attention_mask=attention_mask,
            decoder_input_ids=decoder_input_ids,
            encoder_outputs=encoder_outputs,
            decoder_attention_mask=decoder_attention_mask,
            head_mask=head_mask,
            decoder_head_mask=decoder_head_mask,
            cross_attn_head_mask=cross_attn_head_mask,
            past_key_values=past_key_values,
            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 = self.lm_head(outputs[0])
        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,
            decoder_hidden_states=outputs.decoder_hidden_states,
            decoder_attentions=outputs.decoder_attentions,
            cross_attentions=outputs.cross_attentions,
            encoder_last_hidden_state=outputs.encoder_last_hidden_state,
            encoder_hidden_states=outputs.encoder_hidden_states,
            encoder_attentions=outputs.encoder_attentions,
        )

    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,
        )

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

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