image-match/python/py/Lib/site-packages/transformers/models/sew/modeling_sew.py

#                🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨
#           This file was automatically generated from src/transformers/models/sew/modular_sew.py.
#               Do NOT edit this file manually as any edits will be overwritten by the generation of
#             the file from the modular. If any change should be done, please apply the change to the
#                          modular_sew.py file directly. One of our CI enforces this.
#                🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨
# Copyright 2021 ASAPP Inc. 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.

import math
from collections.abc import Callable

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

from ... import initialization as init
from ...activations import ACT2FN
from ...integrations.deepspeed import is_deepspeed_zero3_enabled
from ...integrations.fsdp import is_fsdp_managed_module
from ...modeling_flash_attention_utils import FlashAttentionKwargs
from ...modeling_layers import GradientCheckpointingLayer
from ...modeling_outputs import BaseModelOutput, CausalLMOutput, SequenceClassifierOutput
from ...modeling_utils import ALL_ATTENTION_FUNCTIONS, PreTrainedModel, get_torch_context_manager_or_global_device
from ...processing_utils import Unpack
from ...utils import TransformersKwargs, auto_docstring, logging
from ...utils.generic import is_flash_attention_requested
from .configuration_sew import SEWConfig


logger = logging.get_logger(__name__)


class SEWNoLayerNormConvLayer(GradientCheckpointingLayer):
    def __init__(self, config, layer_id=0):
        super().__init__()
        self.in_conv_dim = config.conv_dim[layer_id - 1] if layer_id > 0 else 1
        self.out_conv_dim = config.conv_dim[layer_id]

        self.conv = nn.Conv1d(
            self.in_conv_dim,
            self.out_conv_dim,
            kernel_size=config.conv_kernel[layer_id],
            stride=config.conv_stride[layer_id],
            bias=config.conv_bias,
        )
        self.activation = ACT2FN[config.feat_extract_activation]

    def forward(self, hidden_states):
        hidden_states = self.conv(hidden_states)
        hidden_states = self.activation(hidden_states)
        return hidden_states


class SEWLayerNormConvLayer(GradientCheckpointingLayer):
    def __init__(self, config, layer_id=0):
        super().__init__()
        self.in_conv_dim = config.conv_dim[layer_id - 1] if layer_id > 0 else 1
        self.out_conv_dim = config.conv_dim[layer_id]

        self.conv = nn.Conv1d(
            self.in_conv_dim,
            self.out_conv_dim,
            kernel_size=config.conv_kernel[layer_id],
            stride=config.conv_stride[layer_id],
            bias=config.conv_bias,
        )
        self.layer_norm = nn.LayerNorm(self.out_conv_dim, elementwise_affine=True)
        self.activation = ACT2FN[config.feat_extract_activation]

    def forward(self, hidden_states):
        hidden_states = self.conv(hidden_states)

        hidden_states = hidden_states.transpose(-2, -1)
        hidden_states = self.layer_norm(hidden_states)
        hidden_states = hidden_states.transpose(-2, -1)

        hidden_states = self.activation(hidden_states)
        return hidden_states


class SEWGroupNormConvLayer(GradientCheckpointingLayer):
    def __init__(self, config, layer_id=0):
        super().__init__()
        self.in_conv_dim = config.conv_dim[layer_id - 1] if layer_id > 0 else 1
        self.out_conv_dim = config.conv_dim[layer_id]

        self.conv = nn.Conv1d(
            self.in_conv_dim,
            self.out_conv_dim,
            kernel_size=config.conv_kernel[layer_id],
            stride=config.conv_stride[layer_id],
            bias=config.conv_bias,
        )
        self.activation = ACT2FN[config.feat_extract_activation]

        self.layer_norm = nn.GroupNorm(num_groups=self.out_conv_dim, num_channels=self.out_conv_dim, affine=True)

    def forward(self, hidden_states):
        hidden_states = self.conv(hidden_states)
        hidden_states = self.layer_norm(hidden_states)
        hidden_states = self.activation(hidden_states)
        return hidden_states


class SEWPositionalConvEmbedding(nn.Module):
    def __init__(self, config):
        super().__init__()
        self.conv = nn.Conv1d(
            config.hidden_size,
            config.hidden_size,
            kernel_size=config.num_conv_pos_embeddings,
            padding=config.num_conv_pos_embeddings // 2,
            groups=config.num_conv_pos_embedding_groups,
            stride=config.squeeze_factor,
        )

        weight_norm = nn.utils.weight_norm
        if hasattr(nn.utils.parametrizations, "weight_norm"):
            weight_norm = nn.utils.parametrizations.weight_norm

        if is_deepspeed_zero3_enabled():
            import deepspeed

            with deepspeed.zero.GatheredParameters(self.conv.weight, modifier_rank=0):
                self.conv = weight_norm(self.conv, name="weight", dim=2)
            if hasattr(self.conv, "parametrizations"):
                weight_g = self.conv.parametrizations.weight.original0
                weight_v = self.conv.parametrizations.weight.original1
            else:
                weight_g = self.conv.weight_g
                weight_v = self.conv.weight_v
            deepspeed.zero.register_external_parameter(self, weight_v)
            deepspeed.zero.register_external_parameter(self, weight_g)
        else:
            self.conv = weight_norm(self.conv, name="weight", dim=2)

        self.padding = SEWSamePadLayer(config.num_conv_pos_embeddings)
        self.activation = ACT2FN[config.feat_extract_activation]

    def forward(self, hidden_states):
        hidden_states = self.conv(hidden_states)
        hidden_states = self.padding(hidden_states)
        hidden_states = self.activation(hidden_states)

        return hidden_states


class SEWSamePadLayer(nn.Module):
    def __init__(self, num_conv_pos_embeddings):
        super().__init__()
        self.num_pad_remove = 1 if num_conv_pos_embeddings % 2 == 0 else 0

    def forward(self, hidden_states):
        if self.num_pad_remove > 0:
            hidden_states = hidden_states[:, :, : -self.num_pad_remove]
        return hidden_states


class SEWUpsampling(nn.Module):
    def __init__(self, config):
        super().__init__()
        self.projection = nn.Linear(config.hidden_size, config.hidden_size * config.squeeze_factor)
        self.activation = ACT2FN[config.feat_extract_activation]
        self.squeeze_factor = config.squeeze_factor

    def forward(self, hidden_states):
        hidden_states = self.projection(hidden_states)
        hidden_states = self.activation(hidden_states)

        if self.squeeze_factor > 1:
            # transform embedding channels to sequence length
            bsz, src_len, src_embed_dim = hidden_states.size()
            tgt_len = src_len * self.squeeze_factor
            tgt_embed_dim = src_embed_dim // self.squeeze_factor
            hidden_states = hidden_states.reshape(bsz, src_len, self.squeeze_factor, tgt_embed_dim)
            hidden_states = hidden_states.reshape(bsz, tgt_len, tgt_embed_dim)

        return hidden_states


class SEWFeatureEncoder(nn.Module):
    """Construct the features from raw audio waveform"""

    def __init__(self, config):
        super().__init__()

        if config.feat_extract_norm == "group":
            conv_layers = [SEWGroupNormConvLayer(config, layer_id=0)] + [
                SEWNoLayerNormConvLayer(config, layer_id=i + 1) for i in range(config.num_feat_extract_layers - 1)
            ]
        elif config.feat_extract_norm == "layer":
            conv_layers = [SEWLayerNormConvLayer(config, layer_id=i) for i in range(config.num_feat_extract_layers)]
        else:
            raise ValueError(
                f"`config.feat_extract_norm` is {config.feat_extract_norm}, but has to be one of ['group', 'layer']"
            )
        self.conv_layers = nn.ModuleList(conv_layers)
        self.gradient_checkpointing = False
        self._requires_grad = True

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

    def forward(self, input_values):
        hidden_states = input_values[:, None]

        # make sure hidden_states require grad for gradient_checkpointing
        if self._requires_grad and self.training:
            hidden_states.requires_grad = True

        for conv_layer in self.conv_layers:
            hidden_states = conv_layer(hidden_states)

        return hidden_states


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

    # Take the dot product between "query" and "key" to get the raw attention scores.
    attn_weights = torch.matmul(query, key.transpose(2, 3)) * scaling

    if attention_mask is not None:
        attn_weights = attn_weights + attention_mask

    attn_weights = nn.functional.softmax(attn_weights, dim=-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 SEWAttention(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,
        config: SEWConfig | None = 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

        self.k_proj = nn.Linear(embed_dim, embed_dim, bias=bias)
        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)

    def forward(
        self,
        hidden_states: torch.Tensor,
        key_value_states: torch.Tensor | None = None,
        attention_mask: torch.Tensor | None = None,
        output_attentions: bool | None = False,
        # 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, torch.Tensor | None, tuple[torch.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

        # determine input shapes
        input_shape = hidden_states.shape[:-1]

        hidden_shape = (*input_shape, -1, self.head_dim)

        # get query proj
        query_states = self.q_proj(hidden_states).view(hidden_shape).transpose(1, 2)

        current_states = key_value_states if is_cross_attention else hidden_states
        kv_shape = (*current_states.shape[:-1], -1, self.head_dim)
        key_states = self.k_proj(current_states).view(kv_shape).transpose(1, 2)
        value_states = self.v_proj(current_states).view(kv_shape).transpose(1, 2)

        attention_interface: Callable = ALL_ATTENTION_FUNCTIONS.get_interface(
            self.config._attn_implementation, eager_attention_forward
        )

        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=self.scaling,
            output_attentions=output_attentions,
            **kwargs,
        )

        attn_output = attn_output.reshape(*input_shape, -1).contiguous()
        attn_output = self.out_proj(attn_output)

        return attn_output, attn_weights, None


class SEWFeedForward(nn.Module):
    def __init__(self, config):
        super().__init__()
        self.intermediate_dropout = nn.Dropout(config.activation_dropout)

        self.intermediate_dense = nn.Linear(config.hidden_size, config.intermediate_size)
        if isinstance(config.hidden_act, str):
            self.intermediate_act_fn = ACT2FN[config.hidden_act]
        else:
            self.intermediate_act_fn = config.hidden_act

        self.output_dense = nn.Linear(config.intermediate_size, config.hidden_size)
        self.output_dropout = nn.Dropout(config.hidden_dropout)

    def forward(self, hidden_states):
        hidden_states = self.intermediate_dense(hidden_states)
        hidden_states = self.intermediate_act_fn(hidden_states)
        hidden_states = self.intermediate_dropout(hidden_states)

        hidden_states = self.output_dense(hidden_states)
        hidden_states = self.output_dropout(hidden_states)
        return hidden_states


class SEWEncoderLayer(GradientCheckpointingLayer):
    def __init__(self, config):
        super().__init__()
        self.attention = SEWAttention(
            embed_dim=config.hidden_size,
            num_heads=config.num_attention_heads,
            dropout=config.attention_dropout,
            is_decoder=False,
            config=config,
        )

        self.dropout = nn.Dropout(config.hidden_dropout)
        self.layer_norm = nn.LayerNorm(config.hidden_size, eps=config.layer_norm_eps)
        self.feed_forward = SEWFeedForward(config)
        self.final_layer_norm = nn.LayerNorm(config.hidden_size, eps=config.layer_norm_eps)

    def forward(self, hidden_states, attention_mask=None, output_attentions=False):
        attn_residual = hidden_states
        hidden_states, attn_weights, _ = self.attention(
            hidden_states, attention_mask=attention_mask, output_attentions=output_attentions
        )
        hidden_states = self.dropout(hidden_states)
        hidden_states = attn_residual + hidden_states

        hidden_states = self.layer_norm(hidden_states)
        hidden_states = hidden_states + self.feed_forward(hidden_states)
        hidden_states = self.final_layer_norm(hidden_states)

        outputs = (hidden_states,)

        if output_attentions:
            outputs += (attn_weights,)

        return outputs


class SEWEncoder(nn.Module):
    def __init__(self, config):
        super().__init__()
        self.config = config
        self.pos_conv_embed = SEWPositionalConvEmbedding(config)
        self.pool = nn.AvgPool1d(config.squeeze_factor, config.squeeze_factor)
        self.layer_norm = nn.LayerNorm(config.hidden_size, eps=config.layer_norm_eps)
        self.dropout = nn.Dropout(config.hidden_dropout)
        self.layers = nn.ModuleList([SEWEncoderLayer(config) for _ in range(config.num_hidden_layers)])
        self.upsample = SEWUpsampling(config)
        self.gradient_checkpointing = False

    def forward(
        self,
        hidden_states,
        attention_mask=None,
        output_attentions=False,
        output_hidden_states=False,
        return_dict=True,
    ):
        all_hidden_states = () if output_hidden_states else None
        all_self_attentions = () if output_attentions else None

        if attention_mask is not None:
            expand_attention_mask = attention_mask.unsqueeze(-1).repeat(1, 1, hidden_states.shape[2])
            if is_flash_attention_requested(self.config):
                # make sure padded tokens output 0
                hidden_states[~expand_attention_mask] = 0.0
                # 2d mask is passed through the layers
                attention_mask = attention_mask if (attention_mask is not None and 0 in attention_mask) else None
            else:
                # make sure padded tokens output 0
                hidden_states[~expand_attention_mask] = 0.0
                input_lengths = (attention_mask.long()).sum(-1)
                # apply pooling formula to get real output_lengths
                output_lengths = input_lengths // self.config.squeeze_factor
                max_encoder_length = hidden_states.shape[1] // self.config.squeeze_factor
                attention_ids = (
                    torch.arange(0, max_encoder_length, device=output_lengths.device)
                    .view(1, -1)
                    .expand(output_lengths.shape[0], -1)
                )
                attention_mask = (attention_ids < output_lengths.view(-1, 1)).long()

                # extend attention_mask
                attention_mask = 1.0 - attention_mask[:, None, None, :].to(dtype=hidden_states.dtype)
                attention_mask = attention_mask * torch.finfo(hidden_states.dtype).min
                attention_mask = attention_mask.expand(
                    attention_mask.shape[0], 1, attention_mask.shape[-1], attention_mask.shape[-1]
                )

        n_input_timesteps = hidden_states.shape[1]

        hidden_states = hidden_states.transpose(1, 2)
        position_embeddings = self.pos_conv_embed(hidden_states)
        pooled_hidden_states = self.pool(hidden_states)
        min_length = min(position_embeddings.size(-1), pooled_hidden_states.size(-1))
        hidden_states = pooled_hidden_states[..., :min_length] + position_embeddings[..., :min_length]
        hidden_states = hidden_states.transpose(1, 2)

        hidden_states = self.layer_norm(hidden_states)
        hidden_states = self.dropout(hidden_states)

        synced_gpus = is_deepspeed_zero3_enabled() or is_fsdp_managed_module(self)

        for layer in self.layers:
            if output_hidden_states:
                all_hidden_states = all_hidden_states + (hidden_states,)

            # add LayerDrop (see https://huggingface.co/papers/1909.11556 for description)
            dropout_probability = torch.rand([])

            skip_the_layer = self.training and dropout_probability < self.config.layerdrop
            if not skip_the_layer or synced_gpus:
                # under fsdp or deepspeed zero3 all gpus must run in sync
                layer_outputs = layer(
                    hidden_states, attention_mask=attention_mask, output_attentions=output_attentions
                )
                hidden_states = layer_outputs[0]

            if skip_the_layer:
                layer_outputs = (None, None)

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

        if output_hidden_states:
            all_hidden_states = all_hidden_states + (hidden_states,)

        hidden_states = self.upsample(hidden_states)
        if hidden_states.shape[1] < n_input_timesteps:
            hidden_states = nn.functional.pad(hidden_states, (0, 0, 0, n_input_timesteps - hidden_states.shape[1]))

        if not return_dict:
            return tuple(v for v in [hidden_states, all_hidden_states, all_self_attentions] if v is not None)
        return BaseModelOutput(
            last_hidden_state=hidden_states,
            hidden_states=all_hidden_states,
            attentions=all_self_attentions,
        )


@auto_docstring
class SEWPreTrainedModel(PreTrainedModel):
    config: SEWConfig
    base_model_prefix = "sew"
    main_input_name = "input_values"
    input_modalities = "audio"
    supports_gradient_checkpointing = True
    _supports_flash_attn = True
    _supports_sdpa = True
    _supports_flex_attn = False  # needs a proper look into the mask creation

    @torch.no_grad()
    def _init_weights(self, module):
        """Initialize the weights"""
        if isinstance(module, SEWPositionalConvEmbedding):
            init.normal_(
                module.conv.weight,
                mean=0,
                std=2 * math.sqrt(1 / (module.conv.kernel_size[0] * module.conv.in_channels)),
            )
            init.constant_(module.conv.bias, 0)
        elif isinstance(module, nn.Linear):
            init.normal_(module.weight, mean=0.0, std=self.config.initializer_range)
        elif isinstance(module, (nn.LayerNorm, nn.GroupNorm)):
            init.zeros_(module.bias)
            init.ones_(module.weight)
        elif isinstance(module, nn.Conv1d):
            if is_deepspeed_zero3_enabled():
                import deepspeed

                if hasattr(module, "weight_v") and hasattr(module, "weight_g"):
                    with deepspeed.zero.GatheredParameters([module.weight_v, module.weight_g], modifier_rank=0):
                        init.kaiming_normal_(module.weight)
                else:
                    with deepspeed.zero.GatheredParameters(module.weight, modifier_rank=0):
                        init.kaiming_normal_(module.weight)
            else:
                init.kaiming_normal_(module.weight)

        if isinstance(module, (nn.Linear, nn.Conv1d)) and module.bias is not None:
            init.zeros_(module.bias)

    def _get_feat_extract_output_lengths(self, input_lengths: torch.LongTensor | int):
        """
        Computes the output length of the convolutional layers
        """

        def _conv_out_length(input_length, kernel_size, stride):
            # 1D convolutional layer output length formula taken
            # from https://pytorch.org/docs/stable/generated/torch.nn.Conv1d.html
            return torch.div(input_length - kernel_size, stride, rounding_mode="floor") + 1

        for kernel_size, stride in zip(self.config.conv_kernel, self.config.conv_stride):
            input_lengths = _conv_out_length(input_lengths, kernel_size, stride)

        return input_lengths

    def _get_feature_vector_attention_mask(self, feature_vector_length: int, attention_mask: torch.LongTensor):
        output_lengths = self._get_feat_extract_output_lengths(attention_mask.sum(-1)).to(torch.long)
        batch_size = attention_mask.shape[0]

        attention_mask = torch.zeros(
            (batch_size, feature_vector_length), dtype=attention_mask.dtype, device=attention_mask.device
        )
        # these two operations makes sure that all values before the output lengths idxs are attended to
        attention_mask[(torch.arange(attention_mask.shape[0], device=attention_mask.device), output_lengths - 1)] = 1
        attention_mask = attention_mask.flip([-1]).cumsum(-1).flip([-1]).bool()
        return attention_mask


def _compute_mask_indices(
    shape: tuple[int, int],
    mask_prob: float,
    mask_length: int,
    attention_mask: torch.LongTensor | None = 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


@auto_docstring
class SEWModel(SEWPreTrainedModel):
    def __init__(self, config: SEWConfig):
        super().__init__(config)
        self.config = config
        self.feature_extractor = SEWFeatureEncoder(config)
        self.layer_norm = nn.LayerNorm(config.conv_dim[-1], eps=config.layer_norm_eps)

        self.project_features = config.conv_dim[-1] != config.hidden_size
        if self.project_features:
            self.feature_projection = nn.Linear(config.conv_dim[-1], config.hidden_size)
        self.feature_dropout = nn.Dropout(config.feat_proj_dropout)

        if config.mask_time_prob > 0.0 or config.mask_feature_prob > 0.0:
            self.masked_spec_embed = nn.Parameter(torch.Tensor(config.hidden_size).uniform_())

        self.encoder = SEWEncoder(config)

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

    # Copied from transformers.models.wav2vec2.modeling_wav2vec2.Wav2Vec2Model._mask_hidden_states
    def _mask_hidden_states(
        self,
        hidden_states: torch.FloatTensor,
        mask_time_indices: torch.FloatTensor | None = None,
        attention_mask: torch.LongTensor | None = 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 hidden_states

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

        if mask_time_indices is not None:
            # apply SpecAugment along time axis with given mask_time_indices
            hidden_states[mask_time_indices] = self.masked_spec_embed.to(hidden_states.dtype)
        elif self.config.mask_time_prob > 0 and self.training:
            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=hidden_states.device, dtype=torch.bool)
            hidden_states[mask_time_indices] = self.masked_spec_embed.to(hidden_states.dtype)

        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=hidden_states.device, dtype=torch.bool)
            mask_feature_indices = mask_feature_indices[:, None].expand(-1, sequence_length, -1)
            hidden_states[mask_feature_indices] = 0

        return hidden_states

    @auto_docstring
    def forward(
        self,
        input_values: torch.Tensor | None,
        attention_mask: torch.Tensor | None = None,
        mask_time_indices: torch.FloatTensor | None = None,
        output_attentions: bool | None = None,
        output_hidden_states: bool | None = None,
        return_dict: bool | None = None,
        **kwargs,
    ) -> tuple | BaseModelOutput:
        r"""
        mask_time_indices (`torch.BoolTensor` of shape `(batch_size, sequence_length)`, *optional*):
            Indices to mask extracted features for contrastive loss. When in training mode, model learns to predict
            masked extracted features in *config.proj_codevector_dim* space.
        """
        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.return_dict

        extract_features = self.feature_extractor(input_values)
        extract_features = extract_features.transpose(1, 2)
        extract_features = self.layer_norm(extract_features)

        if self.project_features:
            extract_features = self.feature_projection(extract_features)
        hidden_states = self.feature_dropout(extract_features)

        if attention_mask is not None:
            # compute reduced attention_mask corresponding to feature vectors
            attention_mask = self._get_feature_vector_attention_mask(hidden_states.shape[1], attention_mask)

        hidden_states = self._mask_hidden_states(hidden_states, mask_time_indices=mask_time_indices)

        encoder_outputs = self.encoder(
            hidden_states,
            attention_mask=attention_mask,
            output_attentions=output_attentions,
            output_hidden_states=output_hidden_states,
            return_dict=return_dict,
        )

        hidden_states = encoder_outputs[0]

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

        return BaseModelOutput(
            last_hidden_state=hidden_states,
            hidden_states=encoder_outputs.hidden_states,
            attentions=encoder_outputs.attentions,
        )


_HIDDEN_STATES_START_POSITION = 1


@auto_docstring(
    custom_intro="""
    SEW Model with a `language modeling` head on top for Connectionist Temporal Classification (CTC).
    """
)
class SEWForCTC(SEWPreTrainedModel):
    def __init__(self, config, target_lang: str | None = None):
        r"""
        target_lang (`str`, *optional*):
            Language id of adapter weights. Adapter weights are stored in the format adapter.<lang>.safetensors or
            adapter.<lang>.bin. Only relevant when using an instance of [`SEWForCTC`] with adapters. Uses 'eng' by
            default.
        """
        super().__init__(config)

        self.sew = SEWModel(config)
        self.dropout = nn.Dropout(config.final_dropout)

        self.target_lang = target_lang

        if config.vocab_size is None:
            raise ValueError(
                f"You are trying to instantiate {self.__class__} with a configuration that "
                "does not define the vocabulary size of the language model head. Please "
                "instantiate the model as follows: `SEWForCTC.from_pretrained(..., vocab_size=vocab_size)`. "
                "or define `vocab_size` of your model's configuration."
            )
        output_hidden_size = (
            config.output_hidden_size if hasattr(config, "add_adapter") and config.add_adapter else config.hidden_size
        )
        self.lm_head = nn.Linear(output_hidden_size, config.vocab_size)

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

    def tie_weights(self, **kwargs):
        """
        This method overwrites [`~PreTrainedModel.tie_weights`] so that adapter weights can be correctly loaded when
        passing `target_lang=...` to `from_pretrained(...)`.

        This method is **not** supposed to be called by the user and is prone to be changed in the future.
        """

        if get_torch_context_manager_or_global_device() == torch.device("meta"):
            return

        # Note that `tie_weights` is usually used to tie input and output embedding weights. The method is re-purposed to
        # correctly load adapter layers for SEW so that we do not have to introduce a new API to
        # [`PreTrainedModel`]. While slightly hacky, SEW never has to tie input and output embeddings, so that it is
        # ok to repurpose this function here.
        target_lang = self.target_lang

        if target_lang is not None and getattr(self.config, "adapter_attn_dim", None) is None:
            raise ValueError(f"Cannot pass `target_lang`: {target_lang} if `config.adapter_attn_dim` is not defined.")
        elif target_lang is None and getattr(self.config, "adapter_attn_dim", None) is not None:
            logger.info("By default `target_lang` is set to 'eng'.")
        elif target_lang is not None:
            self.load_adapter(target_lang, force_load=True)

    def freeze_feature_encoder(self):
        """
        Calling this function will disable the gradient computation for the feature encoder so that its parameter will
        not be updated during training.
        """
        self.sew.feature_extractor._freeze_parameters()

    def freeze_base_model(self):
        """
        Calling this function will disable the gradient computation for the base model so that its parameters will not
        be updated during training. Only the classification head will be updated.
        """
        for param in self.sew.parameters():
            param.requires_grad = False

    @auto_docstring
    def forward(
        self,
        input_values: torch.Tensor | None,
        attention_mask: torch.Tensor | None = None,
        output_attentions: bool | None = None,
        output_hidden_states: bool | None = None,
        return_dict: bool | None = None,
        labels: torch.Tensor | None = None,
        **kwargs,
    ) -> tuple | CausalLMOutput:
        r"""
        labels (`torch.LongTensor` of shape `(batch_size, target_length)`, *optional*):
            Labels for connectionist temporal classification. Note that `target_length` has to be smaller or equal to
            the sequence length of the output logits. Indices are selected in `[-100, 0, ..., config.vocab_size - 1]`.
            All labels set to `-100` are ignored (masked), the loss is only computed for labels in `[0, ...,
            config.vocab_size - 1]`.
        """
        return_dict = return_dict if return_dict is not None else self.config.return_dict

        if labels is not None and labels.max() >= self.config.vocab_size:
            raise ValueError(f"Label values must be <= vocab_size: {self.config.vocab_size}")

        outputs = self.sew(
            input_values,
            attention_mask=attention_mask,
            output_attentions=output_attentions,
            output_hidden_states=output_hidden_states,
            return_dict=return_dict,
        )

        hidden_states = outputs[0]
        hidden_states = self.dropout(hidden_states)

        logits = self.lm_head(hidden_states)

        loss = None
        if labels is not None:
            # retrieve loss input_lengths from attention_mask
            attention_mask = (
                attention_mask if attention_mask is not None else torch.ones_like(input_values, dtype=torch.long)
            )
            input_lengths = self._get_feat_extract_output_lengths(attention_mask.sum(-1)).to(torch.long)

            # assuming that padded tokens are filled with -100
            # when not being attended to
            labels_mask = labels >= 0
            target_lengths = labels_mask.sum(-1)
            flattened_targets = labels.masked_select(labels_mask)

            # ctc_loss doesn't support fp16
            log_probs = nn.functional.log_softmax(logits, dim=-1, dtype=torch.float32).transpose(0, 1)

            with torch.backends.cudnn.flags(enabled=False):
                loss = nn.functional.ctc_loss(
                    log_probs,
                    flattened_targets,
                    input_lengths,
                    target_lengths,
                    blank=self.config.pad_token_id,
                    reduction=self.config.ctc_loss_reduction,
                    zero_infinity=self.config.ctc_zero_infinity,
                )

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

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


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

        if hasattr(config, "add_adapter") and config.add_adapter:
            raise ValueError(
                "Sequence classification does not support the use of SEW adapters (config.add_adapter=True)"
            )
        self.sew = SEWModel(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_feature_encoder(self):
        """
        Calling this function will disable the gradient computation for the feature encoder so that its parameter will
        not be updated during training.
        """
        self.sew.feature_extractor._freeze_parameters()

    def freeze_base_model(self):
        """
        Calling this function will disable the gradient computation for the base model so that its parameters will not
        be updated during training. Only the classification head will be updated.
        """
        for param in self.sew.parameters():
            param.requires_grad = False

    @auto_docstring
    def forward(
        self,
        input_values: torch.Tensor | None,
        attention_mask: torch.Tensor | None = None,
        output_attentions: bool | None = None,
        output_hidden_states: bool | None = None,
        return_dict: bool | None = None,
        labels: torch.Tensor | None = None,
        **kwargs,
    ) -> tuple | SequenceClassifierOutput:
        r"""
        input_values (`torch.FloatTensor` of shape `(batch_size, sequence_length)`):
            Float values of input raw speech waveform. Values 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_values`, the [`AutoProcessor`] should be used for padding and conversion
            into a tensor of type `torch.FloatTensor`. See [`SEWProcessor.__call__`] for details.
        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).
        """

        return_dict = return_dict if return_dict is not None else self.config.return_dict
        output_hidden_states = True if self.config.use_weighted_layer_sum else output_hidden_states

        outputs = self.sew(
            input_values,
            attention_mask=attention_mask,
            output_attentions=output_attentions,
            output_hidden_states=output_hidden_states,
            return_dict=return_dict,
        )

        if self.config.use_weighted_layer_sum:
            hidden_states = 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 = outputs[0]

        hidden_states = self.projector(hidden_states)
        if attention_mask is None:
            pooled_output = hidden_states.mean(dim=1)
        else:
            padding_mask = self._get_feature_vector_attention_mask(hidden_states.shape[1], attention_mask)
            expand_padding_mask = padding_mask.unsqueeze(-1).repeat(1, 1, hidden_states.shape[2])
            hidden_states[~expand_padding_mask] = 0.0
            pooled_output = hidden_states.sum(dim=1) / padding_mask.sum(dim=1).view(-1, 1)

        logits = self.classifier(pooled_output)

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

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

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


__all__ = ["SEWForCTC", "SEWForSequenceClassification", "SEWModel", "SEWPreTrainedModel"]