transformers/src/transformers/models/hubert/modeling_hubert.py

#                🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨
#           This file was automatically generated from src/transformers/models/hubert/modular_hubert.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_hubert.py file directly. One of our CI enforces this.
#                🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨
import warnings
from typing import Optional, Tuple, Union

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

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 flash_attn_supports_top_left_mask, is_flash_attn_available
from ...modeling_outputs import BaseModelOutput, CausalLMOutput, SequenceClassifierOutput
from ...modeling_utils import PreTrainedModel
from ...utils import (
    add_code_sample_docstrings,
    add_start_docstrings,
    add_start_docstrings_to_model_forward,
    logging,
    replace_return_docstrings,
)
from .configuration_hubert import HubertConfig


if is_flash_attn_available():
    from ...modeling_flash_attention_utils import _flash_attention_forward


logger = logging.get_logger(__name__)

# Base docstring
_CHECKPOINT_FOR_DOC = "facebook/hubert-large-ls960-ft"

# General docstring
_CONFIG_FOR_DOC = "HubertConfig"


class HubertPositionalConvEmbedding(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,
        )

        self.batch_norm = None
        if config.conv_pos_batch_norm:
            self.batch_norm = nn.BatchNorm1d(config.hidden_size)
        else:
            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 = HubertSamePadLayer(config.num_conv_pos_embeddings)
        self.activation = ACT2FN[config.feat_extract_activation]

    def forward(self, hidden_states):
        hidden_states = hidden_states.transpose(1, 2)
        if self.batch_norm is not None:
            hidden_states = self.batch_norm(hidden_states)
        hidden_states = self.conv(hidden_states)
        hidden_states = self.padding(hidden_states)
        hidden_states = self.activation(hidden_states)

        hidden_states = hidden_states.transpose(1, 2)
        return hidden_states


class HubertSamePadLayer(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 HubertNoLayerNormConvLayer(nn.Module):
    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 HubertLayerNormConvLayer(nn.Module):
    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 HubertGroupNormConvLayer(nn.Module):
    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 HubertFeatureEncoder(nn.Module):
    """Construct the features from raw audio waveform"""

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

        if config.feat_extract_norm == "group":
            conv_layers = [HubertGroupNormConvLayer(config, layer_id=0)] + [
                HubertNoLayerNormConvLayer(config, layer_id=i + 1) for i in range(config.num_feat_extract_layers - 1)
            ]
        elif config.feat_extract_norm == "layer":
            conv_layers = [HubertLayerNormConvLayer(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:
            if self._requires_grad and self.gradient_checkpointing and self.training:
                hidden_states = self._gradient_checkpointing_func(
                    conv_layer.__call__,
                    hidden_states,
                )
            else:
                hidden_states = conv_layer(hidden_states)

        return hidden_states


class HubertFeatureProjection(nn.Module):
    def __init__(self, config):
        super().__init__()
        self.feat_proj_layer_norm = config.feat_proj_layer_norm
        if self.feat_proj_layer_norm:
            self.layer_norm = nn.LayerNorm(config.conv_dim[-1], eps=config.layer_norm_eps)
        self.projection = nn.Linear(config.conv_dim[-1], config.hidden_size)
        self.dropout = nn.Dropout(config.feat_proj_dropout)

    def forward(self, hidden_states):
        # non-projected hidden states are needed for quantization
        if self.feat_proj_layer_norm:
            hidden_states = self.layer_norm(hidden_states)
        hidden_states = self.projection(hidden_states)
        hidden_states = self.dropout(hidden_states)
        return hidden_states


class HubertAttention(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: Optional[HubertConfig] = 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 _shape(self, tensor: torch.Tensor, seq_len: int, bsz: int):
        return tensor.view(bsz, seq_len, self.num_heads, self.head_dim).transpose(1, 2).contiguous()

    def forward(
        self,
        hidden_states: torch.Tensor,
        key_value_states: Optional[torch.Tensor] = None,
        past_key_value: Optional[Tuple[torch.Tensor]] = None,
        attention_mask: Optional[torch.Tensor] = None,
        layer_head_mask: Optional[torch.Tensor] = None,
        output_attentions: bool = False,
    ) -> 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

        bsz, tgt_len, _ = hidden_states.size()

        # get query proj
        query_states = self.q_proj(hidden_states) * self.scaling
        # get key, value proj
        # `past_key_value[0].shape[2] == key_value_states.shape[1]`
        # is checking that the `sequence_length` of the `past_key_value` is the same as
        # the provided `key_value_states` to support prefix tuning
        if (
            is_cross_attention
            and past_key_value is not None
            and past_key_value[0].shape[2] == key_value_states.shape[1]
        ):
            # reuse k,v, cross_attentions
            key_states = past_key_value[0]
            value_states = past_key_value[1]
        elif is_cross_attention:
            # cross_attentions
            key_states = self._shape(self.k_proj(key_value_states), -1, bsz)
            value_states = self._shape(self.v_proj(key_value_states), -1, bsz)
        elif past_key_value is not None:
            # reuse k, v, self_attention
            key_states = self._shape(self.k_proj(hidden_states), -1, bsz)
            value_states = self._shape(self.v_proj(hidden_states), -1, bsz)
            key_states = torch.cat([past_key_value[0], key_states], dim=2)
            value_states = torch.cat([past_key_value[1], value_states], dim=2)
        else:
            # self_attention
            key_states = self._shape(self.k_proj(hidden_states), -1, bsz)
            value_states = self._shape(self.v_proj(hidden_states), -1, bsz)

        if self.is_decoder:
            # if cross_attention save Tuple(torch.Tensor, torch.Tensor) of all cross attention key/value_states.
            # Further calls to cross_attention layer can then reuse all cross-attention
            # key/value_states (first "if" case)
            # if uni-directional self-attention (decoder) save Tuple(torch.Tensor, torch.Tensor) of
            # all previous decoder key/value_states. Further calls to uni-directional self-attention
            # can concat previous decoder key/value_states to current projected key/value_states (third "elif" case)
            # if encoder bi-directional self-attention `past_key_value` is always `None`
            past_key_value = (key_states, value_states)

        proj_shape = (bsz * self.num_heads, -1, self.head_dim)
        query_states = self._shape(query_states, tgt_len, bsz).view(*proj_shape)
        key_states = key_states.reshape(*proj_shape)
        value_states = value_states.reshape(*proj_shape)

        src_len = key_states.size(1)
        attn_weights = torch.bmm(query_states, key_states.transpose(1, 2))

        if attn_weights.size() != (bsz * self.num_heads, tgt_len, src_len):
            raise ValueError(
                f"Attention weights should be of size {(bsz * self.num_heads, tgt_len, src_len)}, but is"
                f" {attn_weights.size()}"
            )

        if attention_mask is not None:
            if attention_mask.size() != (bsz, 1, tgt_len, src_len):
                raise ValueError(
                    f"Attention mask should be of size {(bsz, 1, tgt_len, src_len)}, but is {attention_mask.size()}"
                )
            attn_weights = attn_weights.view(bsz, self.num_heads, tgt_len, src_len) + attention_mask
            attn_weights = attn_weights.view(bsz * self.num_heads, tgt_len, src_len)

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

        if layer_head_mask is not None:
            if layer_head_mask.size() != (self.num_heads,):
                raise ValueError(
                    f"Head mask for a single layer should be of size {(self.num_heads,)}, but is"
                    f" {layer_head_mask.size()}"
                )
            attn_weights = layer_head_mask.view(1, -1, 1, 1) * attn_weights.view(bsz, self.num_heads, tgt_len, src_len)
            attn_weights = attn_weights.view(bsz * self.num_heads, tgt_len, src_len)

        if output_attentions:
            # this operation is a bit awkward, but it's required to
            # make sure that attn_weights keeps its gradient.
            # In order to do so, attn_weights have to be reshaped
            # twice and have to be reused in the following
            attn_weights_reshaped = attn_weights.view(bsz, self.num_heads, tgt_len, src_len)
            attn_weights = attn_weights_reshaped.view(bsz * self.num_heads, tgt_len, src_len)
        else:
            attn_weights_reshaped = None

        attn_probs = nn.functional.dropout(attn_weights, p=self.dropout, training=self.training)

        attn_output = torch.bmm(attn_probs, value_states)

        if attn_output.size() != (bsz * self.num_heads, tgt_len, self.head_dim):
            raise ValueError(
                f"`attn_output` should be of size {(bsz * self.num_heads, tgt_len, self.head_dim)}, but is"
                f" {attn_output.size()}"
            )

        attn_output = attn_output.view(bsz, self.num_heads, tgt_len, self.head_dim)
        attn_output = attn_output.transpose(1, 2)

        # Use the `embed_dim` from the config (stored in the class) rather than `hidden_state` because `attn_output` can be
        # partitioned across GPUs when using tensor-parallelism.
        attn_output = attn_output.reshape(bsz, tgt_len, self.embed_dim)

        attn_output = self.out_proj(attn_output)

        return attn_output, attn_weights_reshaped, past_key_value


class HubertFlashAttention2(HubertAttention):
    """
    Hubert flash attention module. This module inherits from `HubertAttention` as the weights of the module stays
    untouched. The only required change would be on the forward pass where it needs to correctly call the public API of
    flash attention and deal with padding tokens in case the input contains any of them.
    """

    def __init__(self, *args, **kwargs):
        super().__init__(*args, **kwargs)

        # TODO: Should be removed once Flash Attention for RoCm is bumped to 2.1.
        # flash_attn<2.1 generates top-left aligned causal mask, while what is needed here is bottom-right alignment, that was made default for flash_attn>=2.1. This attribute is used to handle this difference. Reference: https://github.com/Dao-AILab/flash-attention/releases/tag/v2.1.0.
        # Beware that with flash_attn<2.1, using q_seqlen != k_seqlen (except for the case q_seqlen == 1) produces a wrong mask (top-left).
        self._flash_attn_uses_top_left_mask = flash_attn_supports_top_left_mask()

    def _reshape(self, tensor: torch.Tensor, seq_len: int, bsz: int):
        return tensor.view(bsz, seq_len, self.num_heads, self.head_dim)

    def forward(
        self,
        hidden_states: torch.Tensor,
        key_value_states: Optional[torch.Tensor] = None,
        past_key_value: Optional[Tuple[torch.Tensor]] = None,
        attention_mask: Optional[torch.Tensor] = None,
        layer_head_mask: Optional[torch.Tensor] = None,
        output_attentions: bool = False,
    ) -> Tuple[torch.Tensor, Optional[torch.Tensor], Optional[Tuple[torch.Tensor]]]:
        # HubertFlashAttention2 attention does not support output_attentions
        if output_attentions:
            raise ValueError("HubertFlashAttention2 attention does not support output_attentions")

        # if key_value_states are provided this layer is used as a cross-attention layer
        # for the decoder
        is_cross_attention = key_value_states is not None

        bsz, q_len, _ = hidden_states.size()

        # get query proj
        query_states = self._reshape(self.q_proj(hidden_states), -1, bsz)
        # get key, value proj
        # `past_key_value[0].shape[2] == key_value_states.shape[1]`
        # is checking that the `sequence_length` of the `past_key_value` is the same as
        # the provided `key_value_states` to support prefix tuning
        if (
            is_cross_attention
            and past_key_value is not None
            and past_key_value[0].shape[2] == key_value_states.shape[1]
        ):
            # reuse k,v, cross_attentions
            key_states = past_key_value[0].transpose(1, 2)
            value_states = past_key_value[1].transpose(1, 2)
        elif is_cross_attention:
            # cross_attentions
            key_states = self._reshape(self.k_proj(key_value_states), -1, bsz)
            value_states = self._reshape(self.v_proj(key_value_states), -1, bsz)
        elif past_key_value is not None:
            # reuse k, v, self_attention
            key_states = self._reshape(self.k_proj(hidden_states), -1, bsz)
            value_states = self._reshape(self.v_proj(hidden_states), -1, bsz)
            key_states = torch.cat([past_key_value[0].transpose(1, 2), key_states], dim=1)
            value_states = torch.cat([past_key_value[1].transpose(1, 2), value_states], dim=1)
        else:
            # self_attention
            key_states = self._reshape(self.k_proj(hidden_states), -1, bsz)
            value_states = self._reshape(self.v_proj(hidden_states), -1, bsz)

        if self.is_decoder:
            # if cross_attention save Tuple(torch.Tensor, torch.Tensor) of all cross attention key/value_states.
            # Further calls to cross_attention layer can then reuse all cross-attention
            # key/value_states (first "if" case)
            # if uni-directional self-attention (decoder) save Tuple(torch.Tensor, torch.Tensor) of
            # all previous decoder key/value_states. Further calls to uni-directional self-attention
            # can concat previous decoder key/value_states to current projected key/value_states (third "elif" case)
            # if encoder bi-directional self-attention `past_key_value` is always `None`
            past_key_value = (key_states.transpose(1, 2), value_states.transpose(1, 2))

        kv_seq_len = key_states.shape[-2]
        if past_key_value is not None:
            kv_seq_len += past_key_value[0].shape[-2]

        # In PEFT, usually we cast the layer norms in float32 for training stability reasons
        # therefore the input hidden states gets silently casted in float32. Hence, we need
        # cast them back in the correct dtype just to be sure everything works as expected.
        # This might slowdown training & inference so it is recommended to not cast the LayerNorms
        # in fp32. (LlamaRMSNorm handles it correctly)

        input_dtype = query_states.dtype
        if input_dtype == torch.float32:
            if torch.is_autocast_enabled():
                target_dtype = torch.get_autocast_gpu_dtype()
            # Handle the case where the model is quantized
            elif hasattr(self.config, "_pre_quantization_dtype"):
                target_dtype = self.config._pre_quantization_dtype
            else:
                target_dtype = self.q_proj.weight.dtype

            logger.warning_once(
                f"The input hidden states seems to be silently casted in float32, this might be related to"
                f" the fact you have upcasted embedding or layer norm layers in float32. We will cast back the input in"
                f" {target_dtype}."
            )

            query_states = query_states.to(target_dtype)
            key_states = key_states.to(target_dtype)
            value_states = value_states.to(target_dtype)

        attn_output = _flash_attention_forward(
            query_states,
            key_states,
            value_states,
            attention_mask,
            q_len,
            dropout=self.dropout if self.training else 0.0,
            is_causal=self.is_causal,
            use_top_left_mask=self._flash_attn_uses_top_left_mask,
        )

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

        if not output_attentions:
            attn_weights = None

        return attn_output, attn_weights, past_key_value


class HubertSdpaAttention(HubertAttention):
    def forward(
        self,
        hidden_states: torch.Tensor,
        key_value_states: Optional[torch.Tensor] = None,
        past_key_value: Optional[Tuple[torch.Tensor]] = None,
        attention_mask: Optional[torch.Tensor] = None,
        layer_head_mask: Optional[torch.Tensor] = None,
        output_attentions: bool = False,
    ) -> Tuple[torch.Tensor, Optional[torch.Tensor], Optional[Tuple[torch.Tensor]]]:
        """Input shape: Batch x Time x Channel"""
        if output_attentions or layer_head_mask is not None:
            # TODO: Improve this warning with e.g. `model.config._attn_implementation = "manual"` once this is implemented.
            logger.warning_once(
                "HubertModel is using HubertSdpaAttention, but `torch.nn.functional.scaled_dot_product_attention` does not support `output_attentions=True` or `layer_head_mask` not None. Falling back to the manual attention"
                ' implementation, but specifying the manual implementation will be required from Transformers version v5.0.0 onwards. This warning can be removed using the argument `attn_implementation="eager"` when loading the model.'
            )
            return super().forward(
                hidden_states,
                key_value_states=key_value_states,
                past_key_value=past_key_value,
                attention_mask=attention_mask,
                layer_head_mask=layer_head_mask,
                output_attentions=output_attentions,
            )

        # if key_value_states are provided this layer is used as a cross-attention layer
        # for the decoder
        is_cross_attention = key_value_states is not None

        bsz, tgt_len, _ = hidden_states.size()

        # get query proj
        query_states = self.q_proj(hidden_states)
        # get key, value proj
        # `past_key_value[0].shape[2] == key_value_states.shape[1]`
        # is checking that the `sequence_length` of the `past_key_value` is the same as
        # the provided `key_value_states` to support prefix tuning
        if (
            is_cross_attention
            and past_key_value is not None
            and past_key_value[0].shape[2] == key_value_states.shape[1]
        ):
            # reuse k,v, cross_attentions
            key_states = past_key_value[0]
            value_states = past_key_value[1]
        elif is_cross_attention:
            # cross_attentions
            key_states = self._shape(self.k_proj(key_value_states), -1, bsz)
            value_states = self._shape(self.v_proj(key_value_states), -1, bsz)
        elif past_key_value is not None:
            # reuse k, v, self_attention
            key_states = self._shape(self.k_proj(hidden_states), -1, bsz)
            value_states = self._shape(self.v_proj(hidden_states), -1, bsz)
            key_states = torch.cat([past_key_value[0], key_states], dim=2)
            value_states = torch.cat([past_key_value[1], value_states], dim=2)
        else:
            # self_attention
            key_states = self._shape(self.k_proj(hidden_states), -1, bsz)
            value_states = self._shape(self.v_proj(hidden_states), -1, bsz)

        if self.is_decoder:
            # if cross_attention save Tuple(torch.Tensor, torch.Tensor) of all cross attention key/value_states.
            # Further calls to cross_attention layer can then reuse all cross-attention
            # key/value_states (first "if" case)
            # if uni-directional self-attention (decoder) save Tuple(torch.Tensor, torch.Tensor) of
            # all previous decoder key/value_states. Further calls to uni-directional self-attention
            # can concat previous decoder key/value_states to current projected key/value_states (third "elif" case)
            # if encoder bi-directional self-attention `past_key_value` is always `None`
            past_key_value = (key_states, value_states)

        query_states = self._shape(query_states, tgt_len, bsz)

        # We dispatch to SDPA's Flash Attention or Efficient kernels via this `is_causal` if statement instead of an inline conditional assignment
        # in SDPA to support both torch.compile's dynamic shapes and full graph options. An inline conditional prevents dynamic shapes from compiling.
        # The tgt_len > 1 is necessary to match with AttentionMaskConverter.to_causal_4d that does not create a causal mask in case tgt_len == 1.
        is_causal = True if self.is_causal and attention_mask is None and tgt_len > 1 else False

        # NOTE: SDPA with memory-efficient backend is currently (torch==2.1.2) bugged when using non-contiguous inputs and a custom attn_mask,
        # but we are fine here as `_shape` do call `.contiguous()`. Reference: https://github.com/pytorch/pytorch/issues/112577
        attn_output = torch.nn.functional.scaled_dot_product_attention(
            query_states,
            key_states,
            value_states,
            attn_mask=attention_mask,
            dropout_p=self.dropout if self.training else 0.0,
            is_causal=is_causal,
        )

        if attn_output.size() != (bsz, self.num_heads, tgt_len, self.head_dim):
            raise ValueError(
                f"`attn_output` should be of size {(bsz, self.num_heads, tgt_len, self.head_dim)}, but is"
                f" {attn_output.size()}"
            )

        attn_output = attn_output.transpose(1, 2)

        # Use the `embed_dim` from the config (stored in the class) rather than `hidden_state` because `attn_output` can be
        # partitioned across GPUs when using tensor-parallelism.
        attn_output = attn_output.reshape(bsz, tgt_len, self.embed_dim)

        attn_output = self.out_proj(attn_output)

        return attn_output, None, past_key_value


class HubertFeedForward(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


HUBERT_ATTENTION_CLASSES = {
    "eager": HubertAttention,
    "sdpa": HubertSdpaAttention,
    "flash_attention_2": HubertFlashAttention2,
}


class HubertEncoderLayer(nn.Module):
    def __init__(self, config):
        super().__init__()
        self.attention = HUBERT_ATTENTION_CLASSES[config._attn_implementation](
            embed_dim=config.hidden_size,
            num_heads=config.num_attention_heads,
            dropout=config.attention_dropout,
            is_decoder=False,
        )

        self.dropout = nn.Dropout(config.hidden_dropout)
        self.layer_norm = nn.LayerNorm(config.hidden_size, eps=config.layer_norm_eps)
        self.feed_forward = HubertFeedForward(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 HubertEncoder(nn.Module):
    def __init__(self, config):
        super().__init__()
        self.config = config
        self.pos_conv_embed = HubertPositionalConvEmbedding(config)
        self.layer_norm = nn.LayerNorm(config.hidden_size, eps=config.layer_norm_eps)
        self.dropout = nn.Dropout(config.hidden_dropout)
        self.layers = nn.ModuleList([HubertEncoderLayer(config) for _ in range(config.num_hidden_layers)])
        self.gradient_checkpointing = False
        self._use_flash_attention_2 = config._attn_implementation == "flash_attention_2"

    def forward(
        self,
        hidden_states: torch.tensor,
        attention_mask: Optional[torch.Tensor] = None,
        output_attentions: bool = False,
        output_hidden_states: bool = False,
        return_dict: bool = True,
    ):
        all_hidden_states = () if output_hidden_states else None
        all_self_attentions = () if output_attentions else None

        if attention_mask is not None:
            # make sure padded tokens output 0
            expand_attention_mask = attention_mask.unsqueeze(-1).repeat(1, 1, hidden_states.shape[2])
            hidden_states[~expand_attention_mask] = 0
            if self._use_flash_attention_2:
                # 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:
                # 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]
                )

        position_embeddings = self.pos_conv_embed(hidden_states)
        hidden_states = hidden_states + position_embeddings
        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://arxiv.org/abs/1909.11556 for description)
            dropout_probability = torch.rand([])

            skip_the_layer = True if self.training and (dropout_probability < self.config.layerdrop) else False
            if not skip_the_layer or synced_gpus:
                # under fsdp or deepspeed zero3 all gpus must run in sync
                if self.gradient_checkpointing and self.training:
                    layer_outputs = self._gradient_checkpointing_func(
                        layer.__call__,
                        hidden_states,
                        attention_mask,
                        output_attentions,
                    )
                else:
                    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,)

        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,
        )


class HubertAttnAdapterLayer(nn.Module):
    def __init__(self, config):
        """
        Implements adapter modules directly with 3D tensor weight as parameters and without using ModuleList to speed
        up training throughput.
        """
        super().__init__()
        self.input_dim = config.adapter_attn_dim
        self.hidden_dim = config.hidden_size

        self.norm = nn.LayerNorm(self.hidden_dim)
        self.linear_1 = nn.Linear(self.hidden_dim, self.input_dim)
        self.act_fn = nn.ReLU()
        self.linear_2 = nn.Linear(self.input_dim, self.hidden_dim)

    def forward(self, hidden_states: torch.FloatTensor):
        hidden_states = self.norm(hidden_states)

        hidden_states = self.linear_1(hidden_states)
        hidden_states = self.act_fn(hidden_states)
        hidden_states = self.linear_2(hidden_states)

        return hidden_states


class HubertEncoderLayerStableLayerNorm(nn.Module):
    def __init__(self, config):
        super().__init__()
        self.attention = HUBERT_ATTENTION_CLASSES[config._attn_implementation](
            embed_dim=config.hidden_size,
            num_heads=config.num_attention_heads,
            dropout=config.attention_dropout,
            is_decoder=False,
        )
        self.dropout = nn.Dropout(config.hidden_dropout)
        self.layer_norm = nn.LayerNorm(config.hidden_size, eps=config.layer_norm_eps)
        self.feed_forward = HubertFeedForward(config)
        self.final_layer_norm = nn.LayerNorm(config.hidden_size, eps=config.layer_norm_eps)

        if getattr(config, "adapter_attn_dim", None) is not None:
            self.adapter_layer = HubertAttnAdapterLayer(config)
        else:
            self.adapter_layer = None

    def forward(
        self,
        hidden_states: torch.Tensor,
        attention_mask: Optional[torch.Tensor] = None,
        output_attentions: bool = False,
    ):
        attn_residual = hidden_states
        hidden_states = self.layer_norm(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 = hidden_states + self.feed_forward(self.final_layer_norm(hidden_states))

        if self.adapter_layer is not None:
            hidden_states = hidden_states + self.adapter_layer(hidden_states)

        outputs = (hidden_states,)

        if output_attentions:
            outputs += (attn_weights,)

        return outputs


class HubertEncoderStableLayerNorm(nn.Module):
    def __init__(self, config):
        super().__init__()
        self.config = config
        self.pos_conv_embed = HubertPositionalConvEmbedding(config)
        self.layer_norm = nn.LayerNorm(config.hidden_size, eps=config.layer_norm_eps)
        self.dropout = nn.Dropout(config.hidden_dropout)
        self.layers = nn.ModuleList(
            [HubertEncoderLayerStableLayerNorm(config) for _ in range(config.num_hidden_layers)]
        )
        self.gradient_checkpointing = False
        self._use_flash_attention_2 = config._attn_implementation == "flash_attention_2"

    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:
            # make sure padded tokens are not attended to
            expand_attention_mask = attention_mask.unsqueeze(-1).repeat(1, 1, hidden_states.shape[2])
            hidden_states = hidden_states * expand_attention_mask.to(dtype=hidden_states.dtype)
            if self._use_flash_attention_2:
                # 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:
                # 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]
                )

        position_embeddings = self.pos_conv_embed(hidden_states)
        hidden_states = hidden_states + position_embeddings
        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://arxiv.org/abs/1909.11556 for description)
            dropout_probability = torch.rand([])

            skip_the_layer = True if self.training and (dropout_probability < self.config.layerdrop) else False
            if not skip_the_layer or synced_gpus:
                # under fsdp or deepspeed zero3 all gpus must run in sync
                # XXX: could optimize this like synced_gpus in generate_utils but not sure if it's worth the code complication
                if self.gradient_checkpointing and self.training:
                    layer_outputs = self._gradient_checkpointing_func(
                        layer.__call__,
                        hidden_states,
                        attention_mask,
                        output_attentions,
                    )
                else:
                    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],)

        hidden_states = self.layer_norm(hidden_states)

        if output_hidden_states:
            all_hidden_states = all_hidden_states + (hidden_states,)

        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,
        )


class HubertPreTrainedModel(PreTrainedModel):
    """
    An abstract class to handle weights initialization and a simple interface for downloading and loading pretrained
    models.
    """

    config_class = HubertConfig
    base_model_prefix = "hubert"
    main_input_name = "input_values"
    supports_gradient_checkpointing = True
    _supports_flash_attn_2 = True
    _supports_sdpa = True

    def _init_weights(self, module):
        """Initialize the weights"""
        if isinstance(module, nn.Linear):
            # Slightly different from the TF version which uses truncated_normal for initialization
            # cf https://github.com/pytorch/pytorch/pull/5617
            module.weight.data.normal_(mean=0.0, std=self.config.initializer_range)
        elif isinstance(module, (nn.LayerNorm, nn.GroupNorm)):
            module.bias.data.zero_()
            module.weight.data.fill_(1.0)
        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):
                        nn.init.kaiming_normal_(module.weight.data)
                else:
                    with deepspeed.zero.GatheredParameters(module.weight, modifier_rank=0):
                        nn.init.kaiming_normal_(module.weight.data)
            else:
                nn.init.kaiming_normal_(module.weight.data)

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

    def _get_feat_extract_output_lengths(self, input_lengths: Union[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: 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://arxiv.org/abs/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


HUBERT_START_DOCSTRING = r"""
    Hubert was proposed in [HuBERT: Self-Supervised Speech Representation Learning by Masked Prediction of Hidden
    Units](https://arxiv.org/abs/2106.07447) by Wei-Ning Hsu, Benjamin Bolte, Yao-Hung Hubert Tsai, Kushal Lakhotia,
    Ruslan Salakhutdinov, Abdelrahman Mohamed.

    This model inherits from [`PreTrainedModel`]. Check the superclass documentation for the generic methods the
    library implements for all its model (such as downloading or saving etc.).

    This model is a PyTorch [torch.nn.Module](https://pytorch.org/docs/stable/nn.html#torch.nn.Module) sub-class. Use
    it as a regular PyTorch Module and refer to the PyTorch documentation for all matter related to general usage and
    behavior.

    Parameters:
        config ([`HubertConfig`]): Model configuration class with all the parameters of the model.
            Initializing with a config file does not load the weights associated with the model, only the
            configuration. Check out the [`~PreTrainedModel.from_pretrained`] method to load the model weights.
"""


HUBERT_INPUTS_DOCSTRING = r"""
    Args:
        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]` or a `numpy.ndarray`, *e.g.* via 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 [`Wav2Vec2Processor.__call__`] for details.
        attention_mask (`torch.LongTensor` of shape `(batch_size, sequence_length)`, *optional*):
            Mask to avoid performing convolution and 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)

            <Tip warning={true}>

            `attention_mask` should only be passed if the corresponding processor has `config.return_attention_mask ==
            True`. For all models whose processor has `config.return_attention_mask == False`, such as
            [hubert-base](https://huggingface.co/facebook/hubert-base-ls960), `attention_mask` should **not** be passed
            to avoid degraded performance when doing batched inference. For such models `input_values` should simply be
            padded with 0 and passed without `attention_mask`. Be aware that these models also yield slightly different
            results depending on whether `input_values` is padded or not.

            </Tip>

        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.
"""


@add_start_docstrings(
    "The bare Hubert Model transformer outputting raw hidden-states without any specific head on top.",
    HUBERT_START_DOCSTRING,
)
class HubertModel(HubertPreTrainedModel):
    def __init__(self, config: HubertConfig):
        super().__init__(config)
        self.config = config
        self.feature_extractor = HubertFeatureEncoder(config)
        self.feature_projection = HubertFeatureProjection(config)

        # model only needs masking vector if mask prob is > 0.0
        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_())

        if config.do_stable_layer_norm:
            self.encoder = HubertEncoderStableLayerNorm(config)
        else:
            self.encoder = HubertEncoder(config)

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

    def _mask_hidden_states(
        self,
        hidden_states: torch.FloatTensor,
        mask_time_indices: Optional[torch.FloatTensor] = None,
        attention_mask: Optional[torch.LongTensor] = None,
    ):
        """
        Masks extracted features along time axis and/or along feature axis according to
        [SpecAugment](https://arxiv.org/abs/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

    @add_start_docstrings_to_model_forward(HUBERT_INPUTS_DOCSTRING)
    @replace_return_docstrings(output_type=BaseModelOutput, config_class=_CONFIG_FOR_DOC)
    def forward(
        self,
        input_values: Optional[torch.Tensor],
        attention_mask: Optional[torch.Tensor] = None,
        mask_time_indices: Optional[torch.FloatTensor] = None,
        output_attentions: Optional[bool] = None,
        output_hidden_states: Optional[bool] = None,
        return_dict: Optional[bool] = None,
    ) -> Union[Tuple, BaseModelOutput]:
        """

        Returns:

        Example:

        ```python
        >>> from transformers import AutoProcessor, HubertModel
        >>> from datasets import load_dataset
        >>> import soundfile as sf

        >>> processor = AutoProcessor.from_pretrained("facebook/hubert-large-ls960-ft")
        >>> model = HubertModel.from_pretrained("facebook/hubert-large-ls960-ft")


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


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

        >>> input_values = processor(ds["speech"][0], return_tensors="pt").input_values  # Batch size 1
        >>> hidden_states = model(input_values).last_hidden_state
        ```"""
        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

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

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

        hidden_states = self.feature_projection(extract_features)
        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


_CTC_EXPECTED_OUTPUT = "'MISTER QUILTER IS THE APOSTLE OF THE MIDDLE CLASSES AND WE ARE GLAD TO WELCOME HIS GOSPEL'"
_CTC_EXPECTED_LOSS = 22.68


@add_start_docstrings(
    """Hubert Model with a `language modeling` head on top for Connectionist Temporal Classification (CTC).""",
    HUBERT_START_DOCSTRING,
)
class HubertForCTC(HubertPreTrainedModel):
    def __init__(self, config, target_lang: Optional[str] = None):
        super().__init__(config)

        self.hubert = HubertModel(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: `HubertForCTC.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):
        """
        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.
        """

        # 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 Hubert so that we do not have to introduce a new API to
        # [`PreTrainedModel`]. While slightly hacky, Hubert 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_extractor(self):
        """
        Calling this function will disable the gradient computation for the feature encoder so that its parameter will
        not be updated during training.
        """
        warnings.warn(
            "The method `freeze_feature_extractor` is deprecated and will be removed in Transformers v5. "
            "Please use the equivalent `freeze_feature_encoder` method instead.",
            FutureWarning,
        )
        self.freeze_feature_encoder()

    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.hubert.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.hubert.parameters():
            param.requires_grad = False

    @add_start_docstrings_to_model_forward(HUBERT_INPUTS_DOCSTRING)
    @add_code_sample_docstrings(
        checkpoint=_CHECKPOINT_FOR_DOC,
        output_type=CausalLMOutput,
        config_class=_CONFIG_FOR_DOC,
        expected_output=_CTC_EXPECTED_OUTPUT,
        expected_loss=_CTC_EXPECTED_LOSS,
    )
    def forward(
        self,
        input_values: Optional[torch.Tensor],
        attention_mask: Optional[torch.Tensor] = None,
        output_attentions: Optional[bool] = None,
        output_hidden_states: Optional[bool] = None,
        return_dict: Optional[bool] = None,
        labels: Optional[torch.Tensor] = None,
    ) -> Union[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.use_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.hubert(
            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
        )


_SEQ_CLASS_CHECKPOINT = "superb/hubert-base-superb-ks"
_SEQ_CLASS_EXPECTED_OUTPUT = "'_unknown_'"
_SEQ_CLASS_EXPECTED_LOSS = 8.53


@add_start_docstrings(
    """
    Hubert Model with a sequence classification head on top (a linear layer over the pooled output) for tasks like
    SUPERB Keyword Spotting.
    """,
    HUBERT_START_DOCSTRING,
)
class HubertForSequenceClassification(HubertPreTrainedModel):
    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 Hubert adapters (config.add_adapter=True)"
            )
        self.hubert = HubertModel(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_extractor(self):
        """
        Calling this function will disable the gradient computation for the feature encoder so that its parameters will
        not be updated during training.
        """
        warnings.warn(
            "The method `freeze_feature_extractor` is deprecated and will be removed in Transformers v5. "
            "Please use the equivalent `freeze_feature_encoder` method instead.",
            FutureWarning,
        )
        self.freeze_feature_encoder()

    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.hubert.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.hubert.parameters():
            param.requires_grad = False

    @add_start_docstrings_to_model_forward(HUBERT_INPUTS_DOCSTRING)
    @add_code_sample_docstrings(
        checkpoint=_SEQ_CLASS_CHECKPOINT,
        output_type=SequenceClassifierOutput,
        config_class=_CONFIG_FOR_DOC,
        modality="audio",
        expected_output=_SEQ_CLASS_EXPECTED_OUTPUT,
        expected_loss=_SEQ_CLASS_EXPECTED_LOSS,
    )
    def forward(
        self,
        input_values: Optional[torch.Tensor],
        attention_mask: Optional[torch.Tensor] = None,
        output_attentions: Optional[bool] = None,
        output_hidden_states: Optional[bool] = None,
        return_dict: Optional[bool] = None,
        labels: Optional[torch.Tensor] = None,
    ) -> Union[Tuple, 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).
        """

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

        outputs = self.hubert(
            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__ = ["HubertForCTC", "HubertForSequenceClassification", "HubertModel", "HubertPreTrainedModel"]