AutoAndroidController/py/venv/Lib/site-packages/transformers/models/whisper/modeling_whisper.py

# coding=utf-8
# Copyright 2022 The OpenAI 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.
"""PyTorch Whisper model."""

import math
from typing import Callable, Optional, Union

import numpy as np
import torch
from torch import nn
from torch.nn import CrossEntropyLoss

from ...activations import ACT2FN
from ...cache_utils import Cache, DynamicCache, EncoderDecoderCache
from ...generation import GenerationMixin
from ...masking_utils import create_causal_mask
from ...modeling_flash_attention_utils import (
    FlashAttentionKwargs,
)
from ...modeling_layers import GradientCheckpointingLayer
from ...modeling_outputs import (
    BaseModelOutput,
    BaseModelOutputWithPastAndCrossAttentions,
    CausalLMOutputWithCrossAttentions,
    Seq2SeqLMOutput,
    Seq2SeqModelOutput,
    SequenceClassifierOutput,
)
from ...modeling_utils import ALL_ATTENTION_FUNCTIONS, PreTrainedModel
from ...processing_utils import Unpack
from ...utils import auto_docstring, logging
from ...utils.deprecation import deprecate_kwarg
from .configuration_whisper import WhisperConfig
from .generation_whisper import WhisperGenerationMixin


logger = logging.get_logger(__name__)

_HIDDEN_STATES_START_POSITION = 1


def sinusoids(length: int, channels: int, max_timescale: float = 10000) -> torch.Tensor:
    """Returns sinusoids for positional embedding"""
    if channels % 2 != 0:
        raise ValueError(
            f"Number of channels has to be divisible by 2 for sinusoidal positional embeddings, got {channels} channels."
        )
    log_timescale_increment = math.log(max_timescale) / (channels // 2 - 1)
    inv_timescales = torch.exp(-log_timescale_increment * torch.arange(channels // 2))
    scaled_time = torch.arange(length).view(-1, 1) * inv_timescales.view(1, -1)
    return torch.cat([scaled_time.sin(), scaled_time.cos()], dim=1)


# Copied from transformers.models.bart.modeling_bart.shift_tokens_right
def shift_tokens_right(input_ids: torch.Tensor, pad_token_id: int, decoder_start_token_id: int):
    """
    Shift input ids one token to the right.
    """
    shifted_input_ids = input_ids.new_zeros(input_ids.shape)
    shifted_input_ids[:, 1:] = input_ids[:, :-1].clone()
    shifted_input_ids[:, 0] = decoder_start_token_id

    if pad_token_id is None:
        raise ValueError("self.model.config.pad_token_id has to be defined.")
    # replace possible -100 values in labels by `pad_token_id`
    shifted_input_ids.masked_fill_(shifted_input_ids == -100, pad_token_id)

    return shifted_input_ids


# Copied from transformers.models.wav2vec2.modeling_wav2vec2._compute_mask_indices
def _compute_mask_indices(
    shape: tuple[int, int],
    mask_prob: float,
    mask_length: int,
    attention_mask: Optional[torch.LongTensor] = None,
    min_masks: int = 0,
) -> np.ndarray:
    """
    Computes random mask spans for a given shape. Used to implement [SpecAugment: A Simple Data Augmentation Method for
    ASR](https://huggingface.co/papers/1904.08779). Note that this method is not optimized to run on TPU and should be run on
    CPU as part of the preprocessing during training.

    Args:
        shape: The shape for which to compute masks. This should be of a tuple of size 2 where
               the first element is the batch size and the second element is the length of the axis to span.
        mask_prob:  The percentage of the whole axis (between 0 and 1) which will be masked. The number of
                    independently generated mask spans of length `mask_length` is computed by
                    `mask_prob*shape[1]/mask_length`. Note that due to overlaps, `mask_prob` is an upper bound and the
                    actual percentage will be smaller.
        mask_length: size of the mask
        min_masks: minimum number of masked spans
        attention_mask: A (right-padded) attention mask which independently shortens the feature axis of
                        each batch dimension.
    """
    batch_size, sequence_length = shape

    if mask_length < 1:
        raise ValueError("`mask_length` has to be bigger than 0.")

    if mask_length > sequence_length:
        raise ValueError(
            f"`mask_length` has to be smaller than `sequence_length`, but got `mask_length`: {mask_length}"
            f" and `sequence_length`: {sequence_length}`"
        )

    # epsilon is used for probabilistic rounding
    epsilon = np.random.rand(1).item()

    def compute_num_masked_span(input_length):
        """Given input length, compute how many spans should be masked"""
        num_masked_span = int(mask_prob * input_length / mask_length + epsilon)
        num_masked_span = max(num_masked_span, min_masks)

        # make sure num masked span <= sequence_length
        if num_masked_span * mask_length > sequence_length:
            num_masked_span = sequence_length // mask_length

        # make sure num_masked span is also <= input_length - (mask_length - 1)
        if input_length - (mask_length - 1) < num_masked_span:
            num_masked_span = max(input_length - (mask_length - 1), 0)

        return num_masked_span

    # compute number of masked spans in batch
    input_lengths = (
        attention_mask.detach().sum(-1).tolist()
        if attention_mask is not None
        else [sequence_length for _ in range(batch_size)]
    )

    # SpecAugment mask to fill
    spec_aug_mask = np.zeros((batch_size, sequence_length), dtype=bool)
    spec_aug_mask_idxs = []

    max_num_masked_span = compute_num_masked_span(sequence_length)

    if max_num_masked_span == 0:
        return spec_aug_mask

    for input_length in input_lengths:
        # compute num of masked spans for this input
        num_masked_span = compute_num_masked_span(input_length)

        # get random indices to mask
        spec_aug_mask_idx = np.random.choice(
            np.arange(input_length - (mask_length - 1)), num_masked_span, replace=False
        )

        # pick first sampled index that will serve as a dummy index to pad vector
        # to ensure same dimension for all batches due to probabilistic rounding
        # Picking first sample just pads those vectors twice.
        if len(spec_aug_mask_idx) == 0:
            # this case can only happen if `input_length` is strictly smaller then
            # `sequence_length` in which case the last token has to be a padding
            # token which we can use as a dummy mask id
            dummy_mask_idx = sequence_length - 1
        else:
            dummy_mask_idx = spec_aug_mask_idx[0]

        spec_aug_mask_idx = np.concatenate(
            [spec_aug_mask_idx, np.ones(max_num_masked_span - num_masked_span, dtype=np.int32) * dummy_mask_idx]
        )
        spec_aug_mask_idxs.append(spec_aug_mask_idx)

    spec_aug_mask_idxs = np.array(spec_aug_mask_idxs)

    # expand masked indices to masked spans
    spec_aug_mask_idxs = np.broadcast_to(
        spec_aug_mask_idxs[:, :, None], (batch_size, max_num_masked_span, mask_length)
    )
    spec_aug_mask_idxs = spec_aug_mask_idxs.reshape(batch_size, max_num_masked_span * mask_length)

    # add offset to the starting indexes so that indexes now create a span
    offsets = np.arange(mask_length)[None, None, :]
    offsets = np.broadcast_to(offsets, (batch_size, max_num_masked_span, mask_length)).reshape(
        batch_size, max_num_masked_span * mask_length
    )
    spec_aug_mask_idxs = spec_aug_mask_idxs + offsets

    # ensure that we cannot have indices larger than sequence_length
    if spec_aug_mask_idxs.max() > sequence_length - 1:
        spec_aug_mask_idxs[spec_aug_mask_idxs > sequence_length - 1] = sequence_length - 1

    # scatter indices to mask
    np.put_along_axis(spec_aug_mask, spec_aug_mask_idxs, 1, -1)

    return spec_aug_mask


class WhisperPositionalEmbedding(nn.Embedding):
    def __init__(self, num_positions: int, embedding_dim: int, padding_idx: Optional[int] = None):
        super().__init__(num_positions, embedding_dim)

    def forward(self, input_ids, past_key_values_length=0, position_ids=None):
        if position_ids is None:
            return self.weight[past_key_values_length : past_key_values_length + input_ids.shape[1]]
        else:
            return self.weight[position_ids]


def eager_attention_forward(
    module: nn.Module,
    query: torch.Tensor,
    key: torch.Tensor,
    value: torch.Tensor,
    attention_mask: Optional[torch.Tensor],
    scaling: Optional[float] = None,
    dropout: float = 0.0,
    head_mask: Optional[torch.Tensor] = None,
    **kwargs,
):
    if scaling is None:
        scaling = query.size(-1) ** -0.5

    attn_weights = torch.matmul(query, key.transpose(2, 3)) * scaling
    if attention_mask is not None and attention_mask.ndim == 4:
        attn_weights = attn_weights + attention_mask[:, :, :, : key.shape[-2]]

    attn_weights = nn.functional.softmax(attn_weights, dim=-1)

    if head_mask is not None:
        attn_weights = attn_weights * head_mask.view(1, -1, 1, 1)

    attn_weights = nn.functional.dropout(attn_weights, p=dropout, training=module.training)
    attn_output = torch.matmul(attn_weights, value)
    attn_output = attn_output.transpose(1, 2).contiguous()

    return attn_output, attn_weights


class WhisperAttention(nn.Module):
    """Multi-headed attention from 'Attention Is All You Need' paper"""

    def __init__(
        self,
        embed_dim: int,
        num_heads: int,
        dropout: float = 0.0,
        is_decoder: bool = False,
        bias: bool = True,
        is_causal: bool = False,
        layer_idx: Optional[int] = None,
        config: Optional[WhisperConfig] = None,
    ):
        super().__init__()
        self.embed_dim = embed_dim
        self.num_heads = num_heads
        self.dropout = dropout
        self.head_dim = embed_dim // num_heads
        self.config = config

        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.is_causal = is_causal

        if layer_idx is None and is_decoder:
            logger.warning_once(
                f"Instantiating a decoder {self.__class__.__name__} without passing `layer_idx` is not recommended and "
                "will to errors during the forward call, if caching is used. Please make sure to provide a `layer_idx` "
                "when creating this class."
            )
        self.layer_idx = layer_idx

        self.k_proj = nn.Linear(embed_dim, embed_dim, bias=False)
        self.v_proj = nn.Linear(embed_dim, embed_dim, bias=bias)
        self.q_proj = nn.Linear(embed_dim, embed_dim, bias=bias)
        self.out_proj = nn.Linear(embed_dim, embed_dim, bias=bias)

    @deprecate_kwarg("past_key_value", new_name="past_key_values", version="4.58")
    def forward(
        self,
        hidden_states: torch.Tensor,
        key_value_states: Optional[torch.Tensor] = None,
        past_key_values: Optional[Cache] = None,
        attention_mask: Optional[torch.Tensor] = None,
        layer_head_mask: Optional[torch.Tensor] = None,
        output_attentions: bool = False,
        cache_position: Optional[torch.Tensor] = None,
        # TODO: we need a refactor so that the different attention modules can get their specific kwargs
        # ATM, we have mixed things encoder, decoder, and encoder-decoder attn
        **kwargs: Unpack[FlashAttentionKwargs],
    ) -> tuple[torch.Tensor, Optional[torch.Tensor], Optional[tuple[torch.Tensor]]]:
        """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

        # determine input shapes
        bsz, tgt_len = hidden_states.shape[:-1]
        q_input_shape = (bsz, tgt_len, -1, self.head_dim)

        # Scaling is susceptible to floating point arithmetics' inprecisions
        # which can lead to different results (this is dependent from model
        # to model, e.g. whisper is one such case). We therefore keep the
        # original order of scaling to follow the original implementation
        # and enforce no scaling (1.0) in the attention call below.
        query_states = self.q_proj(hidden_states) * self.scaling
        query_states = query_states.view(*q_input_shape)
        query_states = query_states.transpose(1, 2).contiguous()

        # Check is encoder-decoder model is being used. Otherwise we'll get `DynamicCache`
        if past_key_values is not None and isinstance(past_key_values, EncoderDecoderCache):
            is_updated = past_key_values.is_updated.get(self.layer_idx)
            if is_cross_attention:
                # after the first generated id, we can subsequently re-use all key/value_states from cache
                past_key_values.is_updated[self.layer_idx] = True
                past_key_values = past_key_values.cross_attention_cache
            else:
                past_key_values = past_key_values.self_attention_cache

        # use key_value_states if cross attention
        current_states = key_value_states if key_value_states is not None else hidden_states
        if is_cross_attention and past_key_values and is_updated:
            # reuse k,v, cross_attentions
            key_states = past_key_values.layers[self.layer_idx].keys
            value_states = past_key_values.layers[self.layer_idx].values
        else:
            key_states = self.k_proj(current_states).view(bsz, -1, self.num_heads, self.head_dim)
            value_states = self.v_proj(current_states).view(bsz, -1, self.num_heads, self.head_dim)
            key_states = key_states.transpose(1, 2).contiguous()
            value_states = value_states.transpose(1, 2).contiguous()
            if past_key_values is not None:
                # save all key/value_states to cache to be re-used for fast auto-regressive generation
                cache_position = cache_position if not is_cross_attention else None
                key_states, value_states = past_key_values.update(
                    key_states, value_states, self.layer_idx, {"cache_position": cache_position}
                )

        attention_interface: Callable = eager_attention_forward
        if self.config._attn_implementation != "eager":
            attention_interface = ALL_ATTENTION_FUNCTIONS[self.config._attn_implementation]

        attn_output, attn_weights = attention_interface(
            self,
            query_states,
            key_states,
            value_states,
            attention_mask,
            dropout=0.0 if not self.training else self.dropout,
            scaling=1.0,
            output_attentions=output_attentions,
            head_mask=layer_head_mask,
            **kwargs,
        )

        attn_output = attn_output.reshape(bsz, tgt_len, -1).contiguous()
        attn_output = self.out_proj(attn_output)

        return attn_output, attn_weights


# Copied from transformers.models.mbart.modeling_mbart.MBartEncoderLayer with MBart->Whisper, MBART->WHISPER
class WhisperEncoderLayer(GradientCheckpointingLayer):
    def __init__(self, config: WhisperConfig):
        super().__init__()
        self.embed_dim = config.d_model

        self.self_attn = WhisperAttention(
            embed_dim=self.embed_dim,
            num_heads=config.encoder_attention_heads,
            dropout=config.attention_dropout,
            config=config,
        )
        self.self_attn_layer_norm = nn.LayerNorm(self.embed_dim)
        self.dropout = config.dropout
        self.activation_fn = ACT2FN[config.activation_function]
        self.activation_dropout = config.activation_dropout
        self.fc1 = nn.Linear(self.embed_dim, config.encoder_ffn_dim)
        self.fc2 = nn.Linear(config.encoder_ffn_dim, self.embed_dim)
        self.final_layer_norm = nn.LayerNorm(self.embed_dim)

    def forward(
        self,
        hidden_states: torch.Tensor,
        attention_mask: torch.Tensor,
        layer_head_mask: torch.Tensor,
        output_attentions: bool = False,
    ) -> torch.Tensor:
        """
        Args:
            hidden_states (`torch.FloatTensor`): input to the layer of shape `(batch, seq_len, embed_dim)`
            attention_mask (`torch.FloatTensor`): attention mask of size
                `(batch, 1, tgt_len, src_len)` where padding elements are indicated by very large negative values.
            layer_head_mask (`torch.FloatTensor`): mask for attention heads in a given layer of size
                `(encoder_attention_heads,)`.
            output_attentions (`bool`, *optional*):
                Whether or not to return the attentions tensors of all attention layers. See `attentions` under
                returned tensors for more detail.
        """
        residual = hidden_states
        hidden_states = self.self_attn_layer_norm(hidden_states)
        hidden_states, attn_weights = self.self_attn(
            hidden_states=hidden_states,
            attention_mask=attention_mask,
            layer_head_mask=layer_head_mask,
            output_attentions=output_attentions,
        )
        hidden_states = nn.functional.dropout(hidden_states, p=self.dropout, training=self.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 = nn.functional.dropout(hidden_states, p=self.activation_dropout, training=self.training)
        hidden_states = self.fc2(hidden_states)
        hidden_states = nn.functional.dropout(hidden_states, p=self.dropout, training=self.training)
        hidden_states = residual + hidden_states

        if hidden_states.dtype == torch.float16:
            clamp_value = torch.finfo(hidden_states.dtype).max - 1000
            hidden_states = torch.clamp(hidden_states, min=-clamp_value, max=clamp_value)

        return hidden_states, attn_weights


class WhisperDecoderLayer(GradientCheckpointingLayer):
    def __init__(self, config: WhisperConfig, layer_idx: Optional[int] = None):
        super().__init__()
        self.embed_dim = config.d_model

        self.self_attn = WhisperAttention(
            embed_dim=self.embed_dim,
            num_heads=config.decoder_attention_heads,
            dropout=config.attention_dropout,
            is_decoder=True,
            is_causal=True,
            layer_idx=layer_idx,
            config=config,
        )
        self.dropout = config.dropout
        self.activation_fn = ACT2FN[config.activation_function]
        self.activation_dropout = config.activation_dropout

        self.self_attn_layer_norm = nn.LayerNorm(self.embed_dim)
        self.encoder_attn = WhisperAttention(
            self.embed_dim,
            config.decoder_attention_heads,
            dropout=config.attention_dropout,
            is_decoder=True,
            layer_idx=layer_idx,
            config=config,
        )
        self.encoder_attn_layer_norm = nn.LayerNorm(self.embed_dim)
        self.fc1 = nn.Linear(self.embed_dim, config.decoder_ffn_dim)
        self.fc2 = nn.Linear(config.decoder_ffn_dim, self.embed_dim)
        self.final_layer_norm = nn.LayerNorm(self.embed_dim)

    @deprecate_kwarg("past_key_value", new_name="past_key_values", version="4.58")
    def forward(
        self,
        hidden_states: torch.Tensor,
        attention_mask: Optional[torch.Tensor] = None,
        encoder_hidden_states: Optional[torch.Tensor] = None,
        encoder_attention_mask: Optional[torch.Tensor] = None,
        layer_head_mask: Optional[torch.Tensor] = None,
        cross_attn_layer_head_mask: Optional[torch.Tensor] = None,
        past_key_values: Optional[EncoderDecoderCache] = None,
        output_attentions: Optional[bool] = False,
        use_cache: Optional[bool] = True,
        cache_position: Optional[torch.LongTensor] = None,
    ) -> torch.Tensor:
        """
        Args:
            hidden_states (`torch.FloatTensor`): input to the layer of shape `(batch, seq_len, embed_dim)`
            attention_mask (`torch.FloatTensor`): attention mask of size
                `(batch, 1, tgt_len, src_len)` where padding elements are indicated by very large negative values.
            encoder_hidden_states (`torch.FloatTensor`):
                cross attention input to the layer of shape `(batch, seq_len, embed_dim)`
            encoder_attention_mask (`torch.FloatTensor`): encoder attention mask of size
                `(batch, 1, tgt_len, src_len)` where padding elements are indicated by very large negative values.
            layer_head_mask (`torch.FloatTensor`): mask for attention heads in a given layer of size
                `(encoder_attention_heads,)`.
            cross_attn_layer_head_mask (`torch.FloatTensor`): mask for cross-attention heads in a given layer of
                size `(decoder_attention_heads,)`.
            past_key_values (`Cache`): cached past key and value projection states
            output_attentions (`bool`, *optional*):
                Whether or not to return the attentions tensors of all attention layers. See `attentions` under
                returned tensors for more detail.
        """
        residual = hidden_states
        hidden_states = self.self_attn_layer_norm(hidden_states)

        # Self Attention
        hidden_states, self_attn_weights = self.self_attn(
            hidden_states=hidden_states,
            past_key_values=past_key_values,
            attention_mask=attention_mask,
            layer_head_mask=layer_head_mask,
            output_attentions=output_attentions,
            cache_position=cache_position,
        )
        hidden_states = nn.functional.dropout(hidden_states, p=self.dropout, training=self.training)
        hidden_states = residual + hidden_states

        # Cross-Attention Block
        cross_attn_weights = None
        if encoder_hidden_states is not None:
            residual = hidden_states
            hidden_states = self.encoder_attn_layer_norm(hidden_states)
            hidden_states, cross_attn_weights = 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_values=past_key_values,
                output_attentions=output_attentions,
            )
            hidden_states = nn.functional.dropout(hidden_states, p=self.dropout, training=self.training)
            hidden_states = residual + hidden_states

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

        outputs = (hidden_states,)

        if output_attentions:
            outputs += (self_attn_weights, cross_attn_weights)

        return outputs


@auto_docstring
class WhisperPreTrainedModel(PreTrainedModel):
    config: WhisperConfig
    base_model_prefix = "model"
    main_input_name = "input_features"
    supports_gradient_checkpointing = True
    _no_split_modules = ["WhisperEncoderLayer", "WhisperDecoderLayer"]
    _supports_flash_attn = True
    _supports_sdpa = True
    _supports_flex_attn = True

    _can_compile_fullgraph = True

    def _init_weights(self, module):
        std = self.config.init_std
        if isinstance(module, (nn.Linear, nn.Conv1d)):
            module.weight.data.normal_(mean=0.0, std=std)
            if module.bias is not None:
                module.bias.data.zero_()
        elif isinstance(module, nn.Embedding):
            module.weight.data.normal_(mean=0.0, std=std)
            if module.padding_idx is not None:
                module.weight.data[module.padding_idx].zero_()
        elif isinstance(module, nn.LayerNorm):
            module.weight.data.fill_(1.0)
            module.bias.data.zero_()
        elif isinstance(module, WhisperEncoder):
            module.embed_positions.weight.copy_(sinusoids(*module.embed_positions.weight.shape))
        elif isinstance(module, WhisperForAudioClassification):
            if self.config.use_weighted_layer_sum:
                module.layer_weights.data.fill_(1.0 / (self.config.num_hidden_layers + 1))

    def _get_feat_extract_output_lengths(self, input_lengths: torch.LongTensor):
        """
        Computes the output length of the convolutional layers
        """
        input_lengths = (input_lengths - 1) // 2 + 1

        return input_lengths


class WhisperEncoder(WhisperPreTrainedModel):
    """
    Transformer encoder consisting of *config.encoder_layers* self attention layers. Each layer is a
    [`WhisperEncoderLayer`].

    Args:
        config: WhisperConfig
    """

    def __init__(self, config: WhisperConfig):
        super().__init__(config)
        self.dropout = config.dropout
        self.layerdrop = config.encoder_layerdrop

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

        self.conv1 = nn.Conv1d(self.num_mel_bins, embed_dim, kernel_size=3, padding=1)
        self.conv2 = nn.Conv1d(embed_dim, embed_dim, kernel_size=3, stride=2, padding=1)

        self.embed_positions = nn.Embedding(self.max_source_positions, embed_dim)
        self.embed_positions.requires_grad_(False)

        self.layers = nn.ModuleList([WhisperEncoderLayer(config) for _ in range(config.encoder_layers)])
        self.layer_norm = nn.LayerNorm(config.d_model)

        self.gradient_checkpointing = False
        # Initialize weights and apply final processing
        self.post_init()

    def _freeze_parameters(self):
        for param in self.parameters():
            param.requires_grad = False
        self._requires_grad = False

    def get_input_embeddings(self) -> nn.Module:
        return self.conv1

    def set_input_embeddings(self, value: nn.Module):
        self.conv1 = value

    def forward(
        self,
        input_features,
        attention_mask=None,
        head_mask=None,
        output_attentions=None,
        output_hidden_states=None,
        return_dict=None,
    ):
        r"""
        Args:
            input_features (`torch.LongTensor` of shape `(batch_size, feature_size, sequence_length)`):
                Float values of mel 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]`, a
                `numpy.ndarray` or a `torch.Tensor`, *e.g.* via the torchcodec library (`pip install torchcodec`) or
                the soundfile library (`pip install soundfile`). To prepare the array into
                `input_features`, the [`AutoFeatureExtractor`] should be used for extracting the mel features, padding
                and conversion into a tensor of type `torch.FloatTensor`. See [`~WhisperFeatureExtractor.__call__`]
            attention_mask (`torch.Tensor`)`, *optional*):
                Whisper does not support masking of the `input_features`, this argument is preserved for compatibility,
                but it is not used. By default the silence in the input log mel spectrogram are ignored.
            head_mask (`torch.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.
        """

        expected_seq_length = self.config.max_source_positions * self.conv1.stride[0] * self.conv2.stride[0]
        if input_features.shape[-1] != expected_seq_length:
            raise ValueError(
                f"Whisper expects the mel input features to be of length {expected_seq_length}, but found {input_features.shape[-1]}. Make sure to pad the input mel features to {expected_seq_length}."
            )

        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
        )
        return_dict = return_dict if return_dict is not None else self.config.use_return_dict
        inputs_embeds = nn.functional.gelu(self.conv1(input_features))
        inputs_embeds = nn.functional.gelu(self.conv2(inputs_embeds))

        inputs_embeds = inputs_embeds.permute(0, 2, 1)
        all_positions = torch.arange(self.embed_positions.num_embeddings, device=inputs_embeds.device)

        hidden_states = inputs_embeds + self.embed_positions(all_positions)
        hidden_states = nn.functional.dropout(hidden_states, p=self.dropout, training=self.training)

        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:
            assert head_mask.size()[0] == (len(self.layers)), (
                f"The head_mask should be specified for {len(self.layers)} layers, but it is for {head_mask.size()[0]}."
            )

        for idx, encoder_layer in enumerate(self.layers):
            if output_hidden_states:
                encoder_states = encoder_states + (hidden_states,)
            # add LayerDrop (see https://huggingface.co/papers/1909.11556 for description)
            to_drop = False
            if self.training:
                dropout_probability = torch.rand([])
                if dropout_probability < self.layerdrop:  # skip the layer
                    to_drop = True

            if to_drop:
                layer_outputs = (None, None)
            else:
                layer_outputs = encoder_layer(
                    hidden_states,
                    None,
                    layer_head_mask=(head_mask[idx] if head_mask is not None else None),
                    output_attentions=output_attentions,
                )

                hidden_states = layer_outputs[0]

            if output_attentions:
                all_attentions = all_attentions + (layer_outputs[1],)

        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 BaseModelOutput(
            last_hidden_state=hidden_states, hidden_states=encoder_states, attentions=all_attentions
        )


class WhisperDecoder(WhisperPreTrainedModel):
    """
    Transformer decoder consisting of *config.decoder_layers* layers. Each layer is a [`WhisperDecoderLayer`]

    Args:
        config: WhisperConfig
    """

    main_input_name = "input_ids"

    def __init__(self, config: WhisperConfig):
        super().__init__(config)
        self.dropout = config.dropout
        self.layerdrop = config.decoder_layerdrop
        self.padding_idx = config.pad_token_id
        self.max_target_positions = config.max_target_positions
        self.max_source_positions = config.max_source_positions
        self.embed_scale = math.sqrt(config.d_model) if config.scale_embedding else 1.0

        self.embed_tokens = nn.Embedding(config.vocab_size, config.d_model, self.padding_idx)
        self.embed_positions = WhisperPositionalEmbedding(self.max_target_positions, config.d_model)

        self.layers = nn.ModuleList(
            [WhisperDecoderLayer(config, layer_idx) for layer_idx in range(config.decoder_layers)]
        )

        self.layer_norm = nn.LayerNorm(config.d_model)

        self.gradient_checkpointing = False
        # Initialize weights and apply final processing
        self.post_init()

    def forward(
        self,
        input_ids=None,
        attention_mask=None,
        encoder_hidden_states=None,
        head_mask=None,
        cross_attn_head_mask=None,
        past_key_values=None,
        inputs_embeds=None,
        position_ids=None,
        use_cache=None,
        output_attentions=None,
        output_hidden_states=None,
        return_dict=None,
        cache_position=None,
    ):
        r"""
        Args:
            input_ids (`torch.LongTensor` 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 [`WhisperTokenizer`]. See [`PreTrainedTokenizer.encode`] and
                [`PreTrainedTokenizer.__call__`] for details.

                [What are input IDs?](../glossary#input-ids)
            attention_mask (`torch.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 (`torch.FloatTensor` 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.
            head_mask (`torch.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 (`torch.Tensor` of shape `(decoder_layers, decoder_attention_heads)`, *optional*):
                Mask to nullify selected heads of the attention modules in encoder to avoid performing cross-attention
                on hidden heads. Mask values selected in `[0, 1]`:

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

            past_key_values (`EncoderDecoderCache` or `tuple(tuple(torch.FloatTensor))`, *optional*):
                It is a [`~cache_utils.Cache`] instance. For more details, see our [kv cache guide](https://huggingface.co/docs/transformers/en/kv_cache).

                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 (`torch.FloatTensor` 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.
            cache_position (`torch.LongTensor` of shape `(sequence_length)`, *optional*):
                Indices depicting the position of the input sequence tokens in the sequence. It is used to update the
                cache in the correct position and to infer the complete sequence length.
        """
        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 decoder_input_ids and decoder_inputs_embeds at the same time")
        elif input_ids is not None:
            input_shape = input_ids.size()
            input_ids = input_ids.view(-1, input_shape[-1])
        elif inputs_embeds is not None:
            input_shape = inputs_embeds.size()[:-1]
        else:
            raise ValueError("You have to specify either decoder_input_ids or decoder_inputs_embeds")

        if inputs_embeds is None:
            inputs_embeds = self.embed_tokens(input_ids)

        if use_cache and past_key_values is None:
            if self.config.is_encoder_decoder:
                past_key_values = EncoderDecoderCache(
                    DynamicCache(config=self.config), DynamicCache(config=self.config)
                )
            else:
                past_key_values = DynamicCache(config=self.config)

        past_key_values_length = 0
        if cache_position is not None:
            past_key_values_length = cache_position[0]
        elif past_key_values is not None:
            past_key_values_length = past_key_values.get_seq_length()

        if cache_position is None:
            cache_position = torch.arange(
                past_key_values_length, past_key_values_length + input_shape[1], device=inputs_embeds.device
            )

        if position_ids is None:
            position_ids = cache_position.unsqueeze(0).repeat(input_shape[0], 1)

        # embed positions
        if input_ids is not None:
            positions = self.embed_positions(
                input_ids, past_key_values_length=past_key_values_length, position_ids=position_ids
            )
        else:
            positions = self.embed_positions(
                inputs_embeds, past_key_values_length=past_key_values_length, position_ids=position_ids
            )

        hidden_states = inputs_embeds + positions.to(inputs_embeds.device)
        hidden_states = nn.functional.dropout(hidden_states, p=self.dropout, training=self.training)

        causal_mask = create_causal_mask(
            config=self.config,
            input_embeds=inputs_embeds,
            attention_mask=attention_mask,
            cache_position=cache_position,
            past_key_values=past_key_values,
            position_ids=position_ids,
        )

        if self.gradient_checkpointing and self.training:
            if use_cache:
                logger.warning_once(
                    "`use_cache = True` is incompatible with gradient checkpointing. Setting `use_cache = False`..."
                )
                use_cache = False
        # 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

        # check if head_mask/cross_attn_head_mask has a correct number of layers specified if desired
        for attn_mask, mask_name in zip([head_mask, cross_attn_head_mask], ["head_mask", "cross_attn_head_mask"]):
            if attn_mask is not None:
                assert attn_mask.size()[0] == (len(self.layers)), (
                    f"The `{mask_name}` should be specified for {len(self.layers)} layers, but it is for"
                    f" {head_mask.size()[0]}."
                )
        for idx, decoder_layer in enumerate(self.layers):
            # add LayerDrop (see https://huggingface.co/papers/1909.11556 for description)
            if output_hidden_states:
                all_hidden_states += (hidden_states,)
            if self.training:
                dropout_probability = torch.rand([])
                if dropout_probability < self.layerdrop:
                    continue

            layer_outputs = decoder_layer(
                hidden_states,
                attention_mask=causal_mask,
                encoder_hidden_states=encoder_hidden_states,
                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_values=past_key_values if use_cache else None,
                output_attentions=output_attentions,
                use_cache=use_cache,
                cache_position=cache_position,
            )
            hidden_states = layer_outputs[0]

            if output_attentions:
                all_self_attns += (layer_outputs[1],)

                if encoder_hidden_states is not None:
                    all_cross_attentions += (layer_outputs[2],)

        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 = past_key_values 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 BaseModelOutputWithPastAndCrossAttentions(
            last_hidden_state=hidden_states,
            past_key_values=next_cache,
            hidden_states=all_hidden_states,
            attentions=all_self_attns,
            cross_attentions=all_cross_attentions,
        )


@auto_docstring
class WhisperModel(WhisperPreTrainedModel):
    def __init__(self, config: WhisperConfig):
        super().__init__(config)

        self.encoder = WhisperEncoder(config)
        self.decoder = WhisperDecoder(config)
        # Initialize weights and apply final processing
        self.post_init()

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

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

    def get_encoder(self):
        return self.encoder

    def freeze_encoder(self):
        """
        Calling this function will disable the gradient computation for the Whisper encoder so that its parameters will
        not be updated during training.
        """
        self.encoder._freeze_parameters()

    def _mask_input_features(
        self,
        input_features: torch.FloatTensor,
        attention_mask: Optional[torch.LongTensor] = None,
    ):
        """
        Masks extracted features along time axis and/or along feature axis according to
        [SpecAugment](https://huggingface.co/papers/1904.08779).
        """

        # `config.apply_spec_augment` can set masking to False
        if not getattr(self.config, "apply_spec_augment", True):
            return input_features

        # generate indices & apply SpecAugment along time axis
        batch_size, hidden_size, sequence_length = input_features.size()

        if self.config.mask_time_prob > 0 and self.training:
            # generate indices & apply SpecAugment along time axis
            mask_time_indices = _compute_mask_indices(
                (batch_size, sequence_length),
                mask_prob=self.config.mask_time_prob,
                mask_length=self.config.mask_time_length,
                attention_mask=attention_mask,
                min_masks=self.config.mask_time_min_masks,
            )
            mask_time_indices = torch.tensor(mask_time_indices, device=input_features.device, dtype=torch.bool)
            mask_time_indices = mask_time_indices[:, None].expand(-1, hidden_size, -1)
            input_features[mask_time_indices] = 0

        if self.config.mask_feature_prob > 0 and self.training:
            # generate indices & apply SpecAugment along feature axis
            mask_feature_indices = _compute_mask_indices(
                (batch_size, hidden_size),
                mask_prob=self.config.mask_feature_prob,
                mask_length=self.config.mask_feature_length,
                min_masks=self.config.mask_feature_min_masks,
            )
            mask_feature_indices = torch.tensor(mask_feature_indices, device=input_features.device, dtype=torch.bool)
            input_features[mask_feature_indices] = 0

        return input_features

    @auto_docstring
    def forward(
        self,
        input_features: Optional[torch.FloatTensor] = None,
        attention_mask: Optional[torch.LongTensor] = None,
        decoder_input_ids: Optional[torch.LongTensor] = None,
        decoder_attention_mask: Optional[torch.LongTensor] = None,
        head_mask: Optional[torch.Tensor] = None,
        decoder_head_mask: Optional[torch.Tensor] = None,
        cross_attn_head_mask: Optional[torch.Tensor] = None,
        encoder_outputs: Optional[tuple[tuple[torch.FloatTensor]]] = None,
        past_key_values: Optional[Cache] = None,
        decoder_inputs_embeds: Optional[tuple[torch.FloatTensor]] = None,
        decoder_position_ids: Optional[tuple[torch.LongTensor]] = None,
        use_cache: Optional[bool] = None,
        output_attentions: Optional[bool] = None,
        output_hidden_states: Optional[bool] = None,
        return_dict: Optional[bool] = None,
        cache_position: Optional[torch.LongTensor] = None,
    ) -> Union[tuple[torch.Tensor], Seq2SeqModelOutput]:
        r"""
        decoder_input_ids (`torch.LongTensor` of shape `(batch_size, target_sequence_length)`, *optional*):
            Indices of decoder input sequence tokens in the vocabulary.

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

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

            Whisper uses the `decoder_start_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`).
        decoder_attention_mask (`torch.LongTensor` of shape `(batch_size, target_sequence_length)`, *optional*):
            Default behavior: generate a tensor that ignores pad tokens in `decoder_input_ids`. Causal mask will also
            be used by default.

            If you want to change padding behavior, you should read
            [`modeling_whisper._prepare_decoder_attention_mask`] and modify to your needs. See diagram 1 in [the BART
            paper](https://huggingface.co/papers/1910.13461) for more information on the default strategy.
        cross_attn_head_mask (`torch.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**.
        decoder_position_ids (`torch.LongTensor` 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.n_positions - 1]`.

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

        Example:
         ```python
         >>> import torch
         >>> from transformers import AutoFeatureExtractor, WhisperModel
         >>> from datasets import load_dataset

         >>> model = WhisperModel.from_pretrained("openai/whisper-base")
         >>> feature_extractor = AutoFeatureExtractor.from_pretrained("openai/whisper-base")
         >>> ds = load_dataset("hf-internal-testing/librispeech_asr_dummy", "clean", split="validation")
         >>> inputs = feature_extractor(ds[0]["audio"]["array"], return_tensors="pt")
         >>> input_features = inputs.input_features
         >>> decoder_input_ids = torch.tensor([[1, 1]]) * model.config.decoder_start_token_id
         >>> last_hidden_state = model(input_features, decoder_input_ids=decoder_input_ids).last_hidden_state
         >>> list(last_hidden_state.shape)
         [1, 2, 512]
         ```"""
        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:
            input_features = self._mask_input_features(input_features, attention_mask=attention_mask)

            encoder_outputs = self.encoder(
                input_features,
                head_mask=head_mask,
                output_attentions=output_attentions,
                output_hidden_states=output_hidden_states,
                return_dict=return_dict,
            )
        # If the user passed a tuple for encoder_outputs, we wrap it in a BaseModelOutput when return_dict=True
        elif return_dict and not isinstance(encoder_outputs, BaseModelOutput):
            encoder_outputs = BaseModelOutput(
                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,
            )

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

        if not return_dict:
            return decoder_outputs + encoder_outputs

        return Seq2SeqModelOutput(
            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,
        )


@auto_docstring(
    custom_intro="""
    The Whisper Model with a language modeling head. Can be used for automatic speech recognition.
    """
)
class WhisperForConditionalGeneration(WhisperGenerationMixin, WhisperPreTrainedModel):
    base_model_prefix = "model"
    _tied_weights_keys = ["proj_out.weight"]

    def __init__(self, config: WhisperConfig):
        super().__init__(config)
        self.model = WhisperModel(config)
        self.proj_out = nn.Linear(config.d_model, config.vocab_size, bias=False)
        self.max_target_positions = config.max_target_positions

        # Initialize weights and apply final processing
        self.post_init()

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

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

    def get_output_embeddings(self):
        return self.proj_out

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

    def get_input_embeddings(self) -> nn.Module:
        return self.model.get_input_embeddings()

    def freeze_encoder(self):
        """
        Calling this function will disable the gradient computation for the Whisper encoder so that its parameters will
        not be updated during training.
        """
        self.model.encoder._freeze_parameters()

    @auto_docstring
    def forward(
        self,
        input_features: Optional[torch.FloatTensor] = None,
        attention_mask: Optional[torch.LongTensor] = None,
        decoder_input_ids: Optional[torch.LongTensor] = None,
        decoder_attention_mask: Optional[torch.LongTensor] = None,
        head_mask: Optional[torch.Tensor] = None,
        decoder_head_mask: Optional[torch.Tensor] = None,
        cross_attn_head_mask: Optional[torch.Tensor] = None,
        encoder_outputs: Optional[tuple[tuple[torch.FloatTensor]]] = None,
        past_key_values: Optional[Cache] = None,
        decoder_inputs_embeds: Optional[tuple[torch.FloatTensor]] = None,
        decoder_position_ids: Optional[tuple[torch.LongTensor]] = None,
        labels: Optional[torch.LongTensor] = None,
        use_cache: Optional[bool] = None,
        output_attentions: Optional[bool] = None,
        output_hidden_states: Optional[bool] = None,
        return_dict: Optional[bool] = None,
        cache_position: Optional[torch.LongTensor] = None,
    ) -> Union[tuple[torch.Tensor], Seq2SeqLMOutput]:
        r"""
        decoder_input_ids (`torch.LongTensor` of shape `(batch_size, target_sequence_length)`, *optional*):
            Indices of decoder input sequence tokens in the vocabulary.

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

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

            Whisper uses the `decoder_start_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`).
        decoder_attention_mask (`torch.LongTensor` of shape `(batch_size, target_sequence_length)`, *optional*):
            Default behavior: generate a tensor that ignores pad tokens in `decoder_input_ids`. Causal mask will also
            be used by default.

            If you want to change padding behavior, you should read
            [`modeling_whisper._prepare_decoder_attention_mask`] and modify to your needs. See diagram 1 in [the BART
            paper](https://huggingface.co/papers/1910.13461) for more information on the default strategy.
        cross_attn_head_mask (`torch.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**.
        decoder_position_ids (`torch.LongTensor` 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.n_positions - 1]`.

            [What are position IDs?](../glossary#position-ids)
        labels (`torch.LongTensor` of shape `(batch_size, sequence_length)`, *optional*):
            Labels for computing the 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]`. `sequence_length` should be smaller than or equal to `config.max_target_positions`.

        Example:

        ```python
        >>> import torch
        >>> from transformers import AutoProcessor, WhisperForConditionalGeneration
        >>> from datasets import load_dataset

        >>> processor = AutoProcessor.from_pretrained("openai/whisper-tiny.en")
        >>> model = WhisperForConditionalGeneration.from_pretrained("openai/whisper-tiny.en")

        >>> ds = load_dataset("hf-internal-testing/librispeech_asr_dummy", "clean", split="validation")

        >>> inputs = processor(ds[0]["audio"]["array"], return_tensors="pt")
        >>> input_features = inputs.input_features

        >>> generated_ids = model.generate(inputs=input_features)

        >>> transcription = processor.batch_decode(generated_ids, skip_special_tokens=True)[0]
        >>> transcription
        ' Mr. Quilter is the apostle of the middle classes, and we are glad to welcome his gospel.'
        ```"""
        return_dict = return_dict if return_dict is not None else self.config.use_return_dict

        if labels is not None:
            if labels.shape[1] > self.max_target_positions:
                raise ValueError(
                    f"Labels' sequence length {labels.shape[1]} cannot exceed the maximum allowed length of {self.max_target_positions} tokens."
                )
            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,
            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,
            decoder_position_ids=decoder_position_ids,
            use_cache=use_cache,
            output_attentions=output_attentions,
            output_hidden_states=output_hidden_states,
            return_dict=return_dict,
            cache_position=cache_position,
        )
        lm_logits = self.proj_out(outputs[0])

        loss = None
        if labels is not None:
            loss_fct = CrossEntropyLoss()
            # move labels to correct device to enable PP
            labels = labels.to(lm_logits.device)
            loss = loss_fct(lm_logits.view(-1, self.config.vocab_size), labels.reshape(-1))

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

        return Seq2SeqLMOutput(
            loss=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,
        )


class WhisperDecoderWrapper(WhisperPreTrainedModel):
    """
    This wrapper class is a helper class to correctly load pretrained checkpoints when the causal language model is
    used in combination with the [`EncoderDecoderModel`] framework.
    """

    def __init__(self, config):
        super().__init__(config)
        config.is_encoder_decoder = False
        self.decoder = WhisperDecoder(config)

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

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

    def forward(self, *args, **kwargs):
        return self.decoder(*args, **kwargs)


@auto_docstring(
    custom_intro="""
    Whisper decoder with a language modeling head on top (linear layer with weights tied to the input embeddings).
    """
)
class WhisperForCausalLM(WhisperPreTrainedModel, GenerationMixin):
    _tied_weights_keys = ["proj_out.weight"]
    main_input_name = "input_ids"

    def __init__(self, config):
        super().__init__(config)
        config.is_encoder_decoder = False
        self.model = WhisperDecoderWrapper(config)

        self.proj_out = nn.Linear(config.hidden_size, config.vocab_size, bias=False)

        # Initialize weights and apply final processing
        self.post_init()

    def get_output_embeddings(self):
        return self.proj_out

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

    def get_input_embeddings(self) -> nn.Module:
        return self.model.get_input_embeddings()

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

    def set_decoder(self, decoder):
        self.model.decoder = decoder

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

    @auto_docstring
    def forward(
        self,
        input_ids: Optional[torch.LongTensor] = None,
        attention_mask: Optional[torch.Tensor] = None,
        encoder_outputs: Optional[tuple[torch.FloatTensor]] = None,
        head_mask: Optional[torch.Tensor] = None,
        cross_attn_head_mask: Optional[torch.Tensor] = None,
        past_key_values: Optional[Cache] = None,
        inputs_embeds: Optional[torch.FloatTensor] = None,
        labels: Optional[torch.LongTensor] = None,
        use_cache: Optional[bool] = None,
        output_attentions: Optional[bool] = None,
        output_hidden_states: Optional[bool] = None,
        return_dict: Optional[bool] = None,
        cache_position: Optional[torch.LongTensor] = None,
    ) -> Union[tuple, CausalLMOutputWithCrossAttentions]:
        r"""
        encoder_outputs (`torch.FloatTensor` of shape `(batch_size, sequence_length, hidden_size)`, *optional*):
            Sequence of hidden-states at the output of the last layer of the encoder. Used in the cross-attention
            if the model is configured as a decoder.
        cross_attn_head_mask (`torch.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**.
        labels (`torch.LongTensor` 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]`.

        Example:

        ```python
        >>> from transformers import WhisperForCausalLM, WhisperForConditionalGeneration, WhisperProcessor
        >>> import torch
        >>> from datasets import load_dataset

        >>> processor = WhisperProcessor.from_pretrained("openai/whisper-large-v2")
        >>> model = WhisperForConditionalGeneration.from_pretrained("openai/whisper-large-v2")

        >>> assistant_model = WhisperForCausalLM.from_pretrained("distil-whisper/distil-large-v2")

        >>> ds = load_dataset("hf-internal-testing/librispeech_asr_dummy", "clean", split="validation")
        >>> sample = ds[0]["audio"]
        >>> input_features = processor(
        ...     sample["array"], sampling_rate=sample["sampling_rate"], return_tensors="pt"
        ... ).input_features

        >>> predicted_ids = model.generate(input_features, assistant_model=assistant_model)

        >>> # decode token ids to text
        >>> transcription = processor.batch_decode(predicted_ids, skip_special_tokens=True)[0]
        >>> transcription
        ' Mr. Quilter is the apostle of the middle classes and we are glad to welcome his gospel.'
        ```"""
        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
        )
        return_dict = return_dict if return_dict is not None else self.config.use_return_dict

        # If the user passed a tuple or `BaseModelOutput` for encoder_outputs, we extract only the hidden states
        if isinstance(encoder_outputs, (BaseModelOutput, tuple, list)):
            encoder_outputs = encoder_outputs[0]

        # decoder outputs consists of (dec_features, layer_state, dec_hidden, dec_attn)
        outputs = self.model.decoder(
            input_ids=input_ids,
            attention_mask=attention_mask,
            encoder_hidden_states=encoder_outputs,
            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,
            cache_position=cache_position,
        )

        logits = self.proj_out(outputs[0])

        loss = None
        if labels is not None:
            labels = labels.to(logits.device)
            loss_fct = CrossEntropyLoss()
            loss = loss_fct(logits.view(-1, self.config.vocab_size), labels.view(-1))

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

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


@auto_docstring(
    custom_intro="""
    Whisper Encoder Model with a sequence classification head on top (a linear layer over the pooled output) for tasks
    like SUPERB Keyword Spotting.
    """
)
class WhisperForAudioClassification(WhisperPreTrainedModel):
    def __init__(self, config):
        super().__init__(config)

        self.encoder = WhisperEncoder(config)
        num_layers = config.num_hidden_layers + 1  # transformer layers + input embeddings
        if config.use_weighted_layer_sum:
            self.layer_weights = nn.Parameter(torch.ones(num_layers) / num_layers)
        self.projector = nn.Linear(config.hidden_size, config.classifier_proj_size)
        self.classifier = nn.Linear(config.classifier_proj_size, config.num_labels)

        # Initialize weights and apply final processing
        self.post_init()

    def freeze_encoder(self):
        """
        Calling this function will disable the gradient computation for the Whisper encoder so that its parameters will
        not be updated during training. Only the projection layers and classification head will be updated.
        """
        self.encoder._freeze_parameters()

    def get_input_embeddings(self) -> nn.Module:
        return self.encoder.get_input_embeddings()

    def set_input_embeddings(self, value: nn.Module):
        self.encoder.set_input_embeddings(value)

    @auto_docstring
    def forward(
        self,
        input_features: Optional[torch.LongTensor] = None,
        head_mask: Optional[torch.Tensor] = None,
        encoder_outputs: Optional[tuple[tuple[torch.FloatTensor]]] = None,
        labels: Optional[torch.LongTensor] = None,
        output_attentions: Optional[bool] = None,
        output_hidden_states: Optional[bool] = None,
        return_dict: Optional[bool] = None,
    ) -> Union[tuple[torch.Tensor], SequenceClassifierOutput]:
        r"""
        labels (`torch.LongTensor` of shape `(batch_size,)`, *optional*):
            Labels for computing the sequence classification/regression loss. Indices should be in `[0, ...,
            config.num_labels - 1]`. If `config.num_labels == 1` a regression loss is computed (Mean-Square loss), If
            `config.num_labels > 1` a classification loss is computed (Cross-Entropy).

        Example:

        ```python
        >>> import torch
        >>> from transformers import AutoFeatureExtractor, WhisperForAudioClassification
        >>> from datasets import load_dataset

        >>> feature_extractor = AutoFeatureExtractor.from_pretrained("sanchit-gandhi/whisper-medium-fleurs-lang-id")
        >>> model = WhisperForAudioClassification.from_pretrained("sanchit-gandhi/whisper-medium-fleurs-lang-id")

        >>> ds = load_dataset("google/fleurs", "all", split="validation", streaming=True)
        >>> sample = next(iter(ds))

        >>> inputs = feature_extractor(
        ...     sample["audio"]["array"], sampling_rate=sample["audio"]["sampling_rate"], return_tensors="pt"
        ... )
        >>> input_features = inputs.input_features

        >>> with torch.no_grad():
        ...     logits = model(input_features).logits

        >>> predicted_class_ids = torch.argmax(logits).item()
        >>> predicted_label = model.config.id2label[predicted_class_ids]
        >>> predicted_label
        'Afrikaans'
        ```"""

        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
        )
        if self.config.use_weighted_layer_sum:
            output_hidden_states = True
        elif output_hidden_states is None:
            output_hidden_states = self.config.output_hidden_states

        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,
                head_mask=head_mask,
                output_attentions=output_attentions,
                output_hidden_states=output_hidden_states,
                return_dict=return_dict,
            )

        if self.config.use_weighted_layer_sum:
            hidden_states = encoder_outputs[_HIDDEN_STATES_START_POSITION]
            hidden_states = torch.stack(hidden_states, dim=1)
            norm_weights = nn.functional.softmax(self.layer_weights, dim=-1)
            hidden_states = (hidden_states * norm_weights.view(-1, 1, 1)).sum(dim=1)
        else:
            hidden_states = encoder_outputs[0]

        hidden_states = self.projector(hidden_states)
        pooled_output = hidden_states.mean(dim=1)

        logits = self.classifier(pooled_output)

        loss = None

        if labels is not None:
            loss_fct = CrossEntropyLoss()
            # move labels to correct device to enable PP
            labels = labels.to(logits.device)
            loss = loss_fct(logits.view(-1, self.config.num_labels), labels.view(-1))

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

        return SequenceClassifierOutput(
            loss=loss,
            logits=logits,
            hidden_states=encoder_outputs.hidden_states,
            attentions=encoder_outputs.attentions,
        )


__all__ = [
    "WhisperForCausalLM",
    "WhisperForConditionalGeneration",
    "WhisperModel",
    "WhisperPreTrainedModel",
    "WhisperForAudioClassification",
]