# coding=utf-8 # Copyright 2021 The Fairseq Authors and The HuggingFace Inc. team. All rights reserved. # # Licensed under the Apache License, Version 2.0 (the "License"); # you may not use this file except in compliance with the License. # You may obtain a copy of the License at # # http://www.apache.org/licenses/LICENSE-2.0 # # Unless required by applicable law or agreed to in writing, software # distributed under the License is distributed on an "AS IS" BASIS, # WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. # See the License for the specific language governing permissions and # limitations under the License. """PyTorch Speech2Text model.""" import math from typing import Callable, Optional, Union import torch from torch import nn from torch.nn import CrossEntropyLoss from ...activations import ACT2FN from ...cache_utils import Cache, EncoderDecoderCache from ...generation import GenerationMixin from ...modeling_attn_mask_utils import ( _prepare_4d_attention_mask, _prepare_4d_attention_mask_for_sdpa, _prepare_4d_causal_attention_mask, _prepare_4d_causal_attention_mask_for_sdpa, ) from ...modeling_flash_attention_utils import FlashAttentionKwargs from ...modeling_layers import GradientCheckpointingLayer from ...modeling_outputs import ( BaseModelOutput, BaseModelOutputWithPastAndCrossAttentions, Seq2SeqLMOutput, Seq2SeqModelOutput, ) from ...modeling_utils import ALL_ATTENTION_FUNCTIONS, PreTrainedModel from ...processing_utils import Unpack from ...utils import ( auto_docstring, is_torch_flex_attn_available, logging, ) from .configuration_speech_to_text import Speech2TextConfig if is_torch_flex_attn_available(): from ...integrations.flex_attention import make_flex_block_causal_mask logger = logging.get_logger(__name__) # Copied from transformers.models.bart.modeling_bart.shift_tokens_right def shift_tokens_right(input_ids: torch.Tensor, pad_token_id: int, decoder_start_token_id: int): """ Shift input ids one token to the right. """ shifted_input_ids = input_ids.new_zeros(input_ids.shape) shifted_input_ids[:, 1:] = input_ids[:, :-1].clone() shifted_input_ids[:, 0] = decoder_start_token_id if pad_token_id is None: raise ValueError("self.model.config.pad_token_id has to be defined.") # replace possible -100 values in labels by `pad_token_id` shifted_input_ids.masked_fill_(shifted_input_ids == -100, pad_token_id) return shifted_input_ids class Conv1dSubsampler(nn.Module): """ Convolutional subsampler: a stack of 1D convolution (along temporal dimension) followed by non-linear activation via gated linear units (https://huggingface.co/papers/1911.08460) """ def __init__(self, config): super().__init__() self.config = config self.num_layers = config.num_conv_layers self.in_channels = config.input_feat_per_channel * config.input_channels self.mid_channels = config.conv_channels self.out_channels = config.d_model self.kernel_sizes = config.conv_kernel_sizes self.conv_layers = nn.ModuleList( nn.Conv1d( self.in_channels if i == 0 else self.mid_channels // 2, self.mid_channels if i < self.num_layers - 1 else self.out_channels * 2, kernel_size=k, stride=2, padding=k // 2, ) for i, k in enumerate(self.kernel_sizes) ) def forward(self, input_features): hidden_states = input_features.transpose(1, 2).contiguous() # -> B x (C x D) x T for conv in self.conv_layers: hidden_states = conv(hidden_states) hidden_states = nn.functional.glu(hidden_states, dim=1) hidden_states = hidden_states.transpose(1, 2).contiguous() # -> T x B x (C x D) return hidden_states class Speech2TextSinusoidalPositionalEmbedding(nn.Module): """This module produces sinusoidal positional embeddings of any length.""" def __init__(self, num_positions: int, embedding_dim: int, padding_idx: Optional[int] = None): super().__init__() self.offset = 2 self.embedding_dim = embedding_dim self.padding_idx = padding_idx self.make_weights(num_positions + self.offset, embedding_dim, padding_idx) def make_weights(self, num_embeddings: int, embedding_dim: int, padding_idx: Optional[int] = None): emb_weights = self.get_embedding(num_embeddings, embedding_dim, padding_idx) if hasattr(self, "weights"): # in forward put the weights on the correct dtype and device of the param emb_weights = emb_weights.to(dtype=self.weights.dtype, device=self.weights.device) self.register_buffer("weights", emb_weights, persistent=False) @staticmethod def get_embedding(num_embeddings: int, embedding_dim: int, padding_idx: Optional[int] = None): """ Build sinusoidal embeddings. This matches the implementation in tensor2tensor, but differs slightly from the description in Section 3.5 of "Attention Is All You Need". """ half_dim = embedding_dim // 2 emb = math.log(10000) / (half_dim - 1) emb = torch.exp(torch.arange(half_dim, dtype=torch.int64).float() * -emb) emb = torch.arange(num_embeddings, dtype=torch.int64).float().unsqueeze(1) * emb.unsqueeze(0) emb = torch.cat([torch.sin(emb), torch.cos(emb)], dim=1).view(num_embeddings, -1) if embedding_dim % 2 == 1: # zero pad emb = torch.cat([emb, torch.zeros(num_embeddings, 1)], dim=1) if padding_idx is not None: emb[padding_idx, :] = 0 return emb.to(torch.get_default_dtype()) @torch.no_grad() def forward(self, input_ids: torch.Tensor, past_key_values_length: int = 0): bsz, seq_len = input_ids.size() # Create the position ids from the input token ids. Any padded tokens remain padded. position_ids = self.create_position_ids_from_input_ids(input_ids, self.padding_idx, past_key_values_length).to( input_ids.device ) # expand embeddings if needed max_pos = self.padding_idx + 1 + seq_len if max_pos > self.weights.size(0): self.make_weights(max_pos + self.offset, self.embedding_dim, self.padding_idx) return self.weights.index_select(0, position_ids.view(-1)).view(bsz, seq_len, -1).detach() def create_position_ids_from_input_ids( self, input_ids: torch.Tensor, padding_idx: int, past_key_values_length: Optional[int] = 0 ): """ Replace non-padding symbols with their position numbers. Position numbers begin at padding_idx+1. Padding symbols are ignored. This is modified from fairseq's `utils.make_positions`. Args: x: torch.Tensor x: Returns: torch.Tensor """ # The series of casts and type-conversions here are carefully balanced to both work with ONNX export and XLA. mask = input_ids.ne(padding_idx).int() incremental_indices = (torch.cumsum(mask, dim=1).type_as(mask) + past_key_values_length) * mask return incremental_indices.long() + padding_idx # Copied from transformers.models.bart.modeling_bart.eager_attention_forward def eager_attention_forward( module: nn.Module, query: torch.Tensor, key: torch.Tensor, value: torch.Tensor, attention_mask: Optional[torch.Tensor], scaling: Optional[float] = None, dropout: float = 0.0, head_mask: Optional[torch.Tensor] = None, **kwargs, ): if scaling is None: scaling = query.size(-1) ** -0.5 attn_weights = torch.matmul(query, key.transpose(2, 3)) * scaling if attention_mask is not None: attn_weights = attn_weights + attention_mask attn_weights = nn.functional.softmax(attn_weights, dim=-1) if head_mask is not None: attn_weights = attn_weights * head_mask.view(1, -1, 1, 1) attn_weights = nn.functional.dropout(attn_weights, p=dropout, training=module.training) attn_output = torch.matmul(attn_weights, value) attn_output = attn_output.transpose(1, 2).contiguous() return attn_output, attn_weights # Copied from transformers.models.musicgen.modeling_musicgen.MusicgenAttention with Musicgen->Speech2Text class Speech2TextAttention(nn.Module): """Multi-headed attention from 'Attention Is All You Need' paper""" def __init__( self, embed_dim: int, num_heads: int, dropout: Optional[float] = 0.0, is_decoder: Optional[bool] = False, bias: Optional[bool] = True, is_causal: Optional[bool] = False, config: Optional[Speech2TextConfig] = None, layer_idx: Optional[int] = 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.layer_idx = layer_idx 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: Optional[torch.Tensor] = None, past_key_value: Optional[Cache] = None, attention_mask: Optional[torch.Tensor] = None, layer_head_mask: Optional[torch.Tensor] = None, output_attentions: Optional[bool] = False, cache_position: Optional[torch.Tensor] = None, # TODO: we need a refactor so that the different attention modules can get their specific kwargs # ATM, we have mixed things encoder, decoder, and encoder-decoder attn **kwargs: Unpack[FlashAttentionKwargs], ) -> tuple[torch.Tensor, Optional[torch.Tensor], Optional[tuple[torch.Tensor]]]: """Input shape: Batch x Time x Channel""" # if key_value_states are provided this layer is used as a cross-attention layer # for the decoder is_cross_attention = key_value_states is not None # determine input shapes bsz, tgt_len = hidden_states.shape[:-1] src_len = key_value_states.shape[1] if is_cross_attention else tgt_len q_input_shape = (bsz, tgt_len, -1, self.head_dim) kv_input_shape = (bsz, src_len, -1, self.head_dim) # get query proj query_states = self.q_proj(hidden_states).view(*q_input_shape).transpose(1, 2) if past_key_value is not None: if isinstance(past_key_value, EncoderDecoderCache): is_updated = past_key_value.is_updated.get(self.layer_idx) if is_cross_attention: # after the first generated id, we can subsequently re-use all key/value_layer from cache curr_past_key_value = past_key_value.cross_attention_cache else: curr_past_key_value = past_key_value.self_attention_cache else: curr_past_key_value = past_key_value current_states = key_value_states if is_cross_attention else hidden_states if is_cross_attention and past_key_value is not None and is_updated: # reuse k,v, cross_attentions key_states = curr_past_key_value.layers[self.layer_idx].keys value_states = curr_past_key_value.layers[self.layer_idx].values else: key_states = self.k_proj(current_states).view(*kv_input_shape).transpose(1, 2) value_states = self.v_proj(current_states).view(*kv_input_shape).transpose(1, 2) if past_key_value is not None: # save all key/value_states to cache to be re-used for fast auto-regressive generation cache_position = cache_position if not is_cross_attention else None key_states, value_states = curr_past_key_value.update( key_states, value_states, self.layer_idx, {"cache_position": cache_position} ) # set flag that curr layer for cross-attn is already updated so we can re-use in subsequent calls if is_cross_attention: past_key_value.is_updated[self.layer_idx] = True attention_interface: Callable = eager_attention_forward if self.config._attn_implementation != "eager": attention_interface = ALL_ATTENTION_FUNCTIONS[self.config._attn_implementation] attn_output, attn_weights = attention_interface( self, query_states, key_states, value_states, attention_mask, dropout=0.0 if not self.training else self.dropout, scaling=self.scaling, output_attentions=output_attentions, head_mask=layer_head_mask, **kwargs, ) attn_output = attn_output.reshape(bsz, tgt_len, -1).contiguous() attn_output = self.out_proj(attn_output) return attn_output, attn_weights # Copied from transformers.models.mbart.modeling_mbart.MBartEncoderLayer with MBart->Speech2Text, MBART->SPEECH_TO_TEXT class Speech2TextEncoderLayer(GradientCheckpointingLayer): def __init__(self, config: Speech2TextConfig): super().__init__() self.embed_dim = config.d_model self.self_attn = Speech2TextAttention( embed_dim=self.embed_dim, num_heads=config.encoder_attention_heads, dropout=config.attention_dropout, config=config, ) self.self_attn_layer_norm = nn.LayerNorm(self.embed_dim) self.dropout = config.dropout self.activation_fn = ACT2FN[config.activation_function] self.activation_dropout = config.activation_dropout self.fc1 = nn.Linear(self.embed_dim, config.encoder_ffn_dim) self.fc2 = nn.Linear(config.encoder_ffn_dim, self.embed_dim) self.final_layer_norm = nn.LayerNorm(self.embed_dim) def forward( self, hidden_states: torch.Tensor, attention_mask: torch.Tensor, layer_head_mask: torch.Tensor, output_attentions: bool = False, ) -> torch.Tensor: """ Args: hidden_states (`torch.FloatTensor`): input to the layer of shape `(batch, seq_len, embed_dim)` attention_mask (`torch.FloatTensor`): attention mask of size `(batch, 1, tgt_len, src_len)` where padding elements are indicated by very large negative values. layer_head_mask (`torch.FloatTensor`): mask for attention heads in a given layer of size `(encoder_attention_heads,)`. output_attentions (`bool`, *optional*): Whether or not to return the attentions tensors of all attention layers. See `attentions` under returned tensors for more detail. """ residual = hidden_states hidden_states = self.self_attn_layer_norm(hidden_states) hidden_states, attn_weights = self.self_attn( hidden_states=hidden_states, attention_mask=attention_mask, layer_head_mask=layer_head_mask, output_attentions=output_attentions, ) hidden_states = nn.functional.dropout(hidden_states, p=self.dropout, training=self.training) hidden_states = residual + hidden_states residual = hidden_states hidden_states = self.final_layer_norm(hidden_states) hidden_states = self.activation_fn(self.fc1(hidden_states)) hidden_states = nn.functional.dropout(hidden_states, p=self.activation_dropout, training=self.training) hidden_states = self.fc2(hidden_states) hidden_states = nn.functional.dropout(hidden_states, p=self.dropout, training=self.training) hidden_states = residual + hidden_states if hidden_states.dtype == torch.float16: clamp_value = torch.finfo(hidden_states.dtype).max - 1000 hidden_states = torch.clamp(hidden_states, min=-clamp_value, max=clamp_value) return hidden_states, attn_weights # copied from transformers.models.mbart.modeling_mbart.MBartDecoderLayer with MBart->Speech2Text, MBART->SPEECH_TO_TEXT # TODO: change copy when applying cache class class Speech2TextDecoderLayer(GradientCheckpointingLayer): def __init__(self, config: Speech2TextConfig, layer_idx=None): super().__init__() self.embed_dim = config.d_model self.self_attn = Speech2TextAttention( embed_dim=self.embed_dim, num_heads=config.decoder_attention_heads, dropout=config.attention_dropout, is_decoder=True, is_causal=True, config=config, layer_idx=layer_idx, ) self.dropout = config.dropout self.activation_fn = ACT2FN[config.activation_function] self.activation_dropout = config.activation_dropout self.self_attn_layer_norm = nn.LayerNorm(self.embed_dim) self.encoder_attn = Speech2TextAttention( self.embed_dim, config.decoder_attention_heads, dropout=config.attention_dropout, is_decoder=True, config=config, layer_idx=layer_idx, ) self.encoder_attn_layer_norm = nn.LayerNorm(self.embed_dim) self.fc1 = nn.Linear(self.embed_dim, config.decoder_ffn_dim) self.fc2 = nn.Linear(config.decoder_ffn_dim, self.embed_dim) self.final_layer_norm = nn.LayerNorm(self.embed_dim) # Copied from transformers.models.musicgen.modeling_musicgen.MusicgenDecoderLayer.forward def forward( self, hidden_states: torch.Tensor, attention_mask: Optional[torch.Tensor] = None, encoder_hidden_states: Optional[torch.Tensor] = None, encoder_attention_mask: Optional[torch.Tensor] = None, layer_head_mask: Optional[torch.Tensor] = None, cross_attn_layer_head_mask: Optional[torch.Tensor] = None, past_key_value: Optional[Cache] = None, output_attentions: Optional[bool] = False, use_cache: Optional[bool] = True, cache_position: Optional[torch.Tensor] = None, ) -> torch.Tensor: """ Args: hidden_states (`torch.FloatTensor`): input to the layer of shape `(batch, seq_len, embed_dim)` attention_mask (`torch.FloatTensor`): attention mask of size `(batch, 1, tgt_len, src_len)` where padding elements are indicated by very large negative values. encoder_hidden_states (`torch.FloatTensor`): cross attention input to the layer of shape `(batch, seq_len, embed_dim)` encoder_attention_mask (`torch.FloatTensor`): encoder attention mask of size `(batch, 1, tgt_len, src_len)` where padding elements are indicated by very large negative values. layer_head_mask (`torch.FloatTensor`): mask for attention heads in a given layer of size `(encoder_attention_heads,)`. cross_attn_layer_head_mask (`torch.FloatTensor`): mask for cross-attention heads in a given layer of size `(decoder_attention_heads,)`. past_key_value (`Tuple(torch.FloatTensor)`): cached past key and value projection states output_attentions (`bool`, *optional*): Whether or not to return the attentions tensors of all attention layers. See `attentions` under returned tensors for more detail. """ residual = hidden_states hidden_states = self.self_attn_layer_norm(hidden_states) # Self Attention hidden_states, self_attn_weights = self.self_attn( hidden_states=hidden_states, past_key_value=past_key_value, attention_mask=attention_mask, layer_head_mask=layer_head_mask, output_attentions=output_attentions, cache_position=cache_position, ) hidden_states = nn.functional.dropout(hidden_states, p=self.dropout, training=self.training) hidden_states = residual + hidden_states # Cross-Attention Block cross_attn_weights = None if encoder_hidden_states is not None: residual = hidden_states hidden_states = self.encoder_attn_layer_norm(hidden_states) hidden_states, cross_attn_weights = self.encoder_attn( hidden_states=hidden_states, key_value_states=encoder_hidden_states, attention_mask=encoder_attention_mask, layer_head_mask=cross_attn_layer_head_mask, past_key_value=past_key_value, output_attentions=output_attentions, cache_position=cache_position, ) hidden_states = nn.functional.dropout(hidden_states, p=self.dropout, training=self.training) hidden_states = residual + hidden_states # Fully Connected residual = hidden_states hidden_states = self.final_layer_norm(hidden_states) hidden_states = self.activation_fn(self.fc1(hidden_states)) hidden_states = nn.functional.dropout(hidden_states, p=self.activation_dropout, training=self.training) hidden_states = self.fc2(hidden_states) hidden_states = nn.functional.dropout(hidden_states, p=self.dropout, training=self.training) hidden_states = residual + hidden_states outputs = (hidden_states,) if output_attentions: outputs += (self_attn_weights, cross_attn_weights) return outputs @auto_docstring class Speech2TextPreTrainedModel(PreTrainedModel): config: Speech2TextConfig base_model_prefix = "model" main_input_name = "input_features" supports_gradient_checkpointing = True # TODO: tests would need a rewrite to check for correct implementation # Current tests always assume certain inputs to be passed _supports_flash_attn = False _supports_sdpa = False _supports_flex_attn = False def _init_weights(self, module): std = self.config.init_std if isinstance(module, (nn.Linear, nn.Conv1d)): module.weight.data.normal_(mean=0.0, std=std) if module.bias is not None: module.bias.data.zero_() elif isinstance(module, nn.Embedding): module.weight.data.normal_(mean=0.0, std=std) if module.padding_idx is not None: module.weight.data[module.padding_idx].zero_() def _get_feat_extract_output_lengths(self, input_lengths: torch.LongTensor): """ Computes the output length of the convolutional layers """ for i in range(self.config.num_conv_layers): input_lengths = (input_lengths - 1) // 2 + 1 return input_lengths def _get_feature_vector_attention_mask(self, feature_vector_length, attention_mask): # generate creates 3D attention mask, because of the shape of input_features # convert it to 2D if that's the case if len(attention_mask.shape) > 2: attention_mask = attention_mask[:, :, -1] subsampled_lengths = self._get_feat_extract_output_lengths(attention_mask.sum(-1)) bsz = attention_mask.size()[0] attention_mask = torch.zeros( (bsz, feature_vector_length), dtype=attention_mask.dtype, device=attention_mask.device ) # these two operations makes sure that all values # before the output lengths indices are attended to attention_mask[(torch.arange(bsz, device=attention_mask.device), subsampled_lengths - 1)] = 1 attention_mask = attention_mask.flip([-1]).cumsum(-1).flip([-1]).long() return attention_mask class Speech2TextEncoder(Speech2TextPreTrainedModel): """ Transformer encoder consisting of *config.encoder_layers* self attention layers. Each layer is a [`Speech2TextEncoderLayer`]. Args: config: Speech2TextConfig embed_tokens (nn.Embedding): output embedding """ def __init__(self, config: Speech2TextConfig): super().__init__(config) self.dropout = config.dropout self.layerdrop = config.encoder_layerdrop embed_dim = config.d_model self.padding_idx = config.pad_token_id self.max_source_positions = config.max_source_positions self.embed_scale = math.sqrt(embed_dim) if config.scale_embedding else 1.0 self.conv = Conv1dSubsampler(config) self.embed_positions = Speech2TextSinusoidalPositionalEmbedding( self.max_source_positions, embed_dim, self.padding_idx, ) self.layers = nn.ModuleList([Speech2TextEncoderLayer(config) for _ in range(config.encoder_layers)]) self.layer_norm = nn.LayerNorm(config.d_model) self.gradient_checkpointing = False # Initialize weights and apply final processing self.post_init() def forward( self, input_features, attention_mask=None, head_mask=None, output_attentions=None, output_hidden_states=None, return_dict=None, ): r""" Args: input_features (`torch.LongTensor` of shape `(batch_size, sequence_length, feature_size)`): Float values of fbank features extracted from the raw speech waveform. Raw speech waveform can be obtained by loading a `.flac` or `.wav` audio file into an array of type `list[float]`, a `numpy.ndarray` or a `torch.Tensor`, *e.g.* via the torchcodec libary (`pip install torchcodec`) or the soundfile library (`pip install soundfile`). To prepare the array into `input_features`, the [`AutoFeatureExtractor`] should be used for extracting the fbank features, padding and conversion into a tensor of type `torch.FloatTensor`. See [`~Speech2TextFeatureExtractor.__call__`] attention_mask (`torch.Tensor` 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) head_mask (`torch.Tensor` of shape `(encoder_layers, encoder_attention_heads)`, *optional*): Mask to nullify selected heads of the attention modules. Mask values selected in `[0, 1]`: - 1 indicates the head is **not masked**, - 0 indicates the head is **masked**. output_attentions (`bool`, *optional*): Whether or not to return the attentions tensors of all attention layers. See `attentions` under returned tensors for more detail. output_hidden_states (`bool`, *optional*): Whether or not to return the hidden states of all layers. See `hidden_states` under returned tensors for more detail. return_dict (`bool`, *optional*): Whether or not to return a [`~utils.ModelOutput`] instead of a plain tuple. """ output_attentions = output_attentions if output_attentions is not None else self.config.output_attentions output_hidden_states = ( output_hidden_states if output_hidden_states is not None else self.config.output_hidden_states ) return_dict = return_dict if return_dict is not None else self.config.use_return_dict inputs_embeds = self.conv(input_features) inputs_embeds = self.embed_scale * inputs_embeds # subsample attention mask if necessary if attention_mask is not None: attention_mask = self._get_feature_vector_attention_mask(inputs_embeds.shape[1], attention_mask) padding_mask = attention_mask.ne(1).long() else: padding_mask = torch.zeros(inputs_embeds.shape[:2], dtype=torch.long, device=inputs_embeds.device) embed_pos = self.embed_positions(padding_mask) hidden_states = inputs_embeds + embed_pos hidden_states = nn.functional.dropout(hidden_states, p=self.dropout, training=self.training) attention_mask = self._update_full_mask( attention_mask, inputs_embeds, ) encoder_states = () if output_hidden_states else None all_attentions = () if output_attentions else None # check if head_mask has a correct number of layers specified if desired if head_mask is not None: assert head_mask.size()[0] == (len(self.layers)), ( f"The head_mask should be specified for {len(self.layers)} layers, but it is for {head_mask.size()[0]}." ) for idx, encoder_layer in enumerate(self.layers): if output_hidden_states: encoder_states = encoder_states + (hidden_states,) # add LayerDrop (see https://huggingface.co/papers/1909.11556 for description) to_drop = False if self.training: dropout_probability = torch.rand([]) if dropout_probability < self.layerdrop: # skip the layer to_drop = True if to_drop: layer_outputs = (None, None) else: layer_outputs = encoder_layer( hidden_states, attention_mask, layer_head_mask=(head_mask[idx] if head_mask is not None else None), output_attentions=output_attentions, ) hidden_states = layer_outputs[0] if output_attentions: all_attentions = all_attentions + (layer_outputs[1],) hidden_states = self.layer_norm(hidden_states) if output_hidden_states: encoder_states = encoder_states + (hidden_states,) if not return_dict: return tuple(v for v in [hidden_states, encoder_states, all_attentions] if v is not None) return BaseModelOutput( last_hidden_state=hidden_states, hidden_states=encoder_states, attentions=all_attentions ) # Copied from transformers.models.bart.modeling_bart.BartPreTrainedModel._update_full_mask def _update_full_mask( self, attention_mask: Union[torch.Tensor, None], inputs_embeds: torch.Tensor, ): if attention_mask is not None: if self.config._attn_implementation == "flash_attention_2": attention_mask = attention_mask if 0 in attention_mask else None elif self.config._attn_implementation == "sdpa": # output_attentions=True & head_mask can not be supported when using SDPA, fall back to # the manual implementation that requires a 4D causal mask in all cases. # [bsz, seq_len] -> [bsz, 1, tgt_seq_len, src_seq_len] attention_mask = _prepare_4d_attention_mask_for_sdpa(attention_mask, inputs_embeds.dtype) elif self.config._attn_implementation == "flex_attention": if isinstance(attention_mask, torch.Tensor): attention_mask = make_flex_block_causal_mask(attention_mask, is_causal=False) else: # [bsz, seq_len] -> [bsz, 1, tgt_seq_len, src_seq_len] attention_mask = _prepare_4d_attention_mask(attention_mask, inputs_embeds.dtype) return attention_mask class Speech2TextDecoder(Speech2TextPreTrainedModel): """ Transformer decoder consisting of *config.decoder_layers* layers. Each layer is a [`Speech2TextDecoderLayer`] Args: config: Speech2TextConfig embed_tokens (nn.Embedding): output embedding """ def __init__(self, config: Speech2TextConfig): super().__init__(config) self.dropout = config.dropout self.layerdrop = config.decoder_layerdrop self.padding_idx = config.pad_token_id self.max_target_positions = config.max_target_positions self.embed_scale = math.sqrt(config.d_model) if config.scale_embedding else 1.0 self.embed_tokens = nn.Embedding(config.vocab_size, config.d_model, self.padding_idx) self.embed_positions = Speech2TextSinusoidalPositionalEmbedding( self.max_target_positions, config.d_model, self.padding_idx, ) self.layers = nn.ModuleList( [Speech2TextDecoderLayer(config, layer_idx=i) for i in range(config.decoder_layers)] ) self.layer_norm = nn.LayerNorm(config.d_model) self.gradient_checkpointing = False # Initialize weights and apply final processing self.post_init() def forward( self, input_ids=None, attention_mask=None, encoder_hidden_states=None, encoder_attention_mask=None, head_mask=None, cross_attn_head_mask=None, past_key_values=None, inputs_embeds=None, use_cache=None, output_attentions=None, output_hidden_states=None, return_dict=None, cache_position=None, ): r""" Args: input_ids (`torch.LongTensor` of shape `(batch_size, sequence_length)`): Indices of input sequence tokens in the vocabulary. Padding will be ignored by default should you provide it. Indices can be obtained using [`Speech2TextTokenizer`]. See [`PreTrainedTokenizer.encode`] and [`PreTrainedTokenizer.__call__`] for details. [What are input IDs?](../glossary#input-ids) attention_mask (`torch.Tensor` of shape `(batch_size, sequence_length)`, *optional*): Mask to avoid performing attention on padding token indices. Mask values selected in `[0, 1]`: - 1 for tokens that are **not masked**, - 0 for tokens that are **masked**. [What are attention masks?](../glossary#attention-mask) encoder_hidden_states (`torch.FloatTensor` of shape `(batch_size, encoder_sequence_length, hidden_size)`, *optional*): Sequence of hidden-states at the output of the last layer of the encoder. Used in the cross-attention of the decoder. encoder_attention_mask (`torch.LongTensor` of shape `(batch_size, encoder_sequence_length)`, *optional*): Mask to avoid performing cross-attention on padding tokens indices of encoder input_ids. Mask values selected in `[0, 1]`: - 1 for tokens that are **not masked**, - 0 for tokens that are **masked**. [What are attention masks?](../glossary#attention-mask) head_mask (`torch.Tensor` of shape `(decoder_layers, decoder_attention_heads)`, *optional*): Mask to nullify selected heads of the attention modules. Mask values selected in `[0, 1]`: - 1 indicates the head is **not masked**, - 0 indicates the head is **masked**. cross_attn_head_mask (`torch.Tensor` of shape `(decoder_layers, decoder_attention_heads)`, *optional*): Mask to nullify selected heads of the attention modules in encoder to avoid performing cross-attention on hidden heads. Mask values selected in `[0, 1]`: - 1 indicates the head is **not masked**, - 0 indicates the head is **masked**. past_key_values (`tuple(tuple(torch.FloatTensor))`, *optional*, returned when `use_cache=True` is passed or when `config.use_cache=True`): Tuple of `tuple(torch.FloatTensor)` of length `config.n_layers`, with each tuple having 2 tensors of shape `(batch_size, num_heads, sequence_length, embed_size_per_head)`) and 2 additional tensors of shape `(batch_size, num_heads, encoder_sequence_length, embed_size_per_head)`. Contains pre-computed hidden-states (key and values in the self-attention blocks and in the cross-attention blocks) that can be used (see `past_key_values` input) to speed up sequential decoding. If `past_key_values` are used, the user can optionally input only the last `decoder_input_ids` (those that don't have their past key value states given to this model) of shape `(batch_size, 1)` instead of all `decoder_input_ids` of shape `(batch_size, sequence_length)`. inputs_embeds (`torch.FloatTensor` of shape `(batch_size, sequence_length, hidden_size)`, *optional*): Optionally, instead of passing `input_ids` you can choose to directly pass an embedded representation. This is useful if you want more control over how to convert `input_ids` indices into associated vectors than the model's internal embedding lookup matrix. output_attentions (`bool`, *optional*): Whether or not to return the attentions tensors of all attention layers. See `attentions` under returned tensors for more detail. output_hidden_states (`bool`, *optional*): Whether or not to return the hidden states of all layers. See `hidden_states` under returned tensors for more detail. return_dict (`bool`, *optional*): Whether or not to return a [`~utils.ModelOutput`] instead of a plain tuple. """ output_attentions = output_attentions if output_attentions is not None else self.config.output_attentions output_hidden_states = ( output_hidden_states if output_hidden_states is not None else self.config.output_hidden_states ) use_cache = use_cache if use_cache is not None else self.config.use_cache return_dict = return_dict if return_dict is not None else self.config.use_return_dict # retrieve input_ids and inputs_embeds if input_ids is not None and inputs_embeds is not None: raise ValueError("You cannot specify both decoder_input_ids and decoder_inputs_embeds at the same time") elif input_ids is not None: input_shape = input_ids.size() input_ids = input_ids.view(-1, input_shape[-1]) elif inputs_embeds is not None: input_shape = inputs_embeds.size()[:-1] else: raise ValueError("You have to specify either decoder_input_ids or decoder_inputs_embeds") if inputs_embeds is None: inputs_embeds = self.embed_tokens(input_ids) * self.embed_scale if self.gradient_checkpointing and self.training: if use_cache: logger.warning_once( "`use_cache = True` is incompatible with gradient checkpointing. Setting `use_cache = False`..." ) use_cache = False return_legacy_cache = False if use_cache and not isinstance(past_key_values, Cache): return_legacy_cache = True logger.warning_once( "Passing a tuple of `past_key_values` is deprecated and will be removed in Transformers v4.58.0. " "You should pass an instance of `EncoderDecoderCache` instead, e.g. " "`past_key_values=EncoderDecoderCache.from_legacy_cache(past_key_values)`." ) past_key_values = EncoderDecoderCache.from_legacy_cache(past_key_values) past_key_values_length = past_key_values.get_seq_length() if past_key_values is not None else 0 attention_mask = self._update_causal_mask( attention_mask, input_shape, inputs_embeds, past_key_values_length, ) encoder_attention_mask = self._update_cross_attn_mask( encoder_hidden_states, encoder_attention_mask, input_shape, inputs_embeds, ) # embed positions positions = self.embed_positions(input_ids, past_key_values_length=past_key_values_length) hidden_states = inputs_embeds + positions hidden_states = nn.functional.dropout(hidden_states, p=self.dropout, training=self.training) # decoder layers all_hidden_states = () if output_hidden_states else None all_self_attns = () if output_attentions else None all_cross_attentions = () if (output_attentions and encoder_hidden_states is not None) else None # check if head_mask/cross_attn_head_mask has a correct number of layers specified if desired for attn_mask, mask_name in zip([head_mask, cross_attn_head_mask], ["head_mask", "cross_attn_head_mask"]): if attn_mask is not None: assert attn_mask.size()[0] == (len(self.layers)), ( f"The `{mask_name}` should be specified for {len(self.layers)} layers, but it is for" f" {head_mask.size()[0]}." ) for idx, decoder_layer in enumerate(self.layers): # add LayerDrop (see https://huggingface.co/papers/1909.11556 for description) if output_hidden_states: all_hidden_states += (hidden_states,) if self.training: dropout_probability = torch.rand([]) if dropout_probability < self.layerdrop: continue layer_outputs = decoder_layer( hidden_states, attention_mask, encoder_hidden_states, # as a positional argument for gradient checkpointing encoder_attention_mask=encoder_attention_mask, layer_head_mask=(head_mask[idx] if head_mask is not None else None), cross_attn_layer_head_mask=(cross_attn_head_mask[idx] if cross_attn_head_mask is not None else None), past_key_value=past_key_values, output_attentions=output_attentions, use_cache=use_cache, cache_position=cache_position, ) hidden_states = layer_outputs[0] if output_attentions: all_self_attns += (layer_outputs[1],) if encoder_hidden_states is not None: all_cross_attentions += (layer_outputs[2],) hidden_states = self.layer_norm(hidden_states) # add hidden states from the last decoder layer if output_hidden_states: all_hidden_states += (hidden_states,) if return_legacy_cache: past_key_values = past_key_values.to_legacy_cache() if not return_dict: return tuple( v for v in [hidden_states, past_key_values, all_hidden_states, all_self_attns, all_cross_attentions] if v is not None ) return BaseModelOutputWithPastAndCrossAttentions( last_hidden_state=hidden_states, past_key_values=past_key_values, hidden_states=all_hidden_states, attentions=all_self_attns, cross_attentions=all_cross_attentions, ) # Copied from transformers.models.musicgen.modeling_musicgen.MusicgenDecoder._update_causal_mask def _update_causal_mask( self, attention_mask: Union[torch.Tensor, None], input_shape: torch.Size, inputs_embeds: torch.Tensor, past_key_values_length: int, ): if self.config._attn_implementation == "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 elif self.config._attn_implementation == "sdpa": # output_attentions=True & cross_attn_head_mask can not be supported when using SDPA, and we fall back on # the manual implementation that requires a 4D causal mask in all cases. attention_mask = _prepare_4d_causal_attention_mask_for_sdpa( attention_mask, input_shape, inputs_embeds, past_key_values_length, ) elif self.config._attn_implementation == "flex_attention": if isinstance(attention_mask, torch.Tensor): attention_mask = make_flex_block_causal_mask(attention_mask) # Other attention flavors support in-built causal (when `mask is None`) # while we need to create our specific block mask regardless elif attention_mask is None: attention_mask = make_flex_block_causal_mask( torch.ones( size=(input_shape), device=inputs_embeds.device, ) ) else: # 4d mask is passed through the layers attention_mask = _prepare_4d_causal_attention_mask( attention_mask, input_shape, inputs_embeds, past_key_values_length ) return attention_mask # Copied from transformers.models.musicgen.modeling_musicgen.MusicgenDecoder._update_cross_attn_mask def _update_cross_attn_mask( self, encoder_hidden_states: Union[torch.Tensor, None], encoder_attention_mask: Union[torch.Tensor, None], input_shape: torch.Size, inputs_embeds: torch.Tensor, ): # expand encoder attention mask if encoder_hidden_states is not None and encoder_attention_mask is not None: if self.config._attn_implementation == "flash_attention_2": encoder_attention_mask = encoder_attention_mask if 0 in encoder_attention_mask else None elif self.config._attn_implementation == "sdpa": # output_attentions=True & cross_attn_head_mask can not be supported when using SDPA, and we fall back on # the manual implementation that requires a 4D causal mask in all cases. # [bsz, seq_len] -> [bsz, 1, tgt_seq_len, src_seq_len] encoder_attention_mask = _prepare_4d_attention_mask_for_sdpa( encoder_attention_mask, inputs_embeds.dtype, tgt_len=input_shape[-1], ) elif self.config._attn_implementation == "flex_attention": if isinstance(encoder_attention_mask, torch.Tensor): encoder_attention_mask = make_flex_block_causal_mask( encoder_attention_mask, query_length=input_shape[-1], is_causal=False, ) else: # [bsz, seq_len] -> [bsz, 1, tgt_seq_len, src_seq_len] encoder_attention_mask = _prepare_4d_attention_mask( encoder_attention_mask, inputs_embeds.dtype, tgt_len=input_shape[-1] ) return encoder_attention_mask @auto_docstring class Speech2TextModel(Speech2TextPreTrainedModel): def __init__(self, config: Speech2TextConfig): super().__init__(config) self.encoder = Speech2TextEncoder(config) self.decoder = Speech2TextDecoder(config) # Initialize weights and apply final processing self.post_init() def get_input_embeddings(self): return self.decoder.embed_tokens def set_input_embeddings(self, value): self.decoder.embed_tokens = value def get_encoder(self): return self.encoder def get_decoder(self): return self.decoder @auto_docstring def forward( self, input_features: Optional[torch.LongTensor] = None, attention_mask: Optional[torch.Tensor] = None, decoder_input_ids: Optional[torch.LongTensor] = None, decoder_attention_mask: Optional[torch.LongTensor] = None, head_mask: Optional[torch.Tensor] = None, decoder_head_mask: Optional[torch.Tensor] = None, cross_attn_head_mask: Optional[torch.Tensor] = None, encoder_outputs: Optional[tuple[tuple[torch.FloatTensor]]] = None, past_key_values: Optional[tuple[tuple[torch.FloatTensor]]] = None, decoder_inputs_embeds: Optional[torch.FloatTensor] = None, use_cache: Optional[bool] = None, output_attentions: Optional[bool] = None, output_hidden_states: Optional[bool] = None, return_dict: Optional[bool] = None, cache_position: Optional[torch.Tensor] = None, ) -> Union[tuple[torch.FloatTensor], Seq2SeqLMOutput]: r""" decoder_input_ids (`torch.LongTensor` of shape `(batch_size, target_sequence_length)`, *optional*): Indices of decoder input sequence tokens in the vocabulary. Indices can be obtained using [`SpeechToTextTokenizer`]. See [`PreTrainedTokenizer.encode`] and [`PreTrainedTokenizer.__call__`] for details. [What are decoder input IDs?](../glossary#decoder-input-ids) SpeechToText uses the `eos_token_id` as the starting token for `decoder_input_ids` generation. If `past_key_values` is used, optionally only the last `decoder_input_ids` have to be input (see `past_key_values`). decoder_attention_mask (`torch.LongTensor` of shape `(batch_size, target_sequence_length)`, *optional*): Default behavior: generate a tensor that ignores pad tokens in `decoder_input_ids`. Causal mask will also be used by default. If you want to change padding behavior, you should read [`modeling_speech_to_text._prepare_decoder_attention_mask`] and modify to your needs. See diagram 1 in [the paper](https://huggingface.co/papers/1910.13461) for more information on the default strategy. cross_attn_head_mask (`torch.Tensor` of shape `(decoder_layers, decoder_attention_heads)`, *optional*): Mask to nullify selected heads of the cross-attention modules. Mask values selected in `[0, 1]`: - 1 indicates the head is **not masked**, - 0 indicates the head is **masked**. Example: ```python >>> import torch >>> from transformers import Speech2TextModel, AutoFeatureExtractor >>> from datasets import load_dataset >>> model = Speech2TextModel.from_pretrained("facebook/s2t-small-librispeech-asr") >>> feature_extractor = AutoFeatureExtractor.from_pretrained("facebook/s2t-small-librispeech-asr") >>> ds = load_dataset("hf-internal-testing/librispeech_asr_dummy", "clean", split="validation") >>> inputs = feature_extractor( ... ds[0]["audio"]["array"], sampling_rate=ds[0]["audio"]["sampling_rate"], return_tensors="pt" ... ) >>> input_features = inputs.input_features >>> decoder_input_ids = torch.tensor([[1, 1]]) * model.config.decoder_start_token_id >>> last_hidden_state = model(input_features, decoder_input_ids=decoder_input_ids).last_hidden_state >>> list(last_hidden_state.shape) [1, 2, 256] ```""" output_attentions = output_attentions if output_attentions is not None else self.config.output_attentions output_hidden_states = ( output_hidden_states if output_hidden_states is not None else self.config.output_hidden_states ) use_cache = use_cache if use_cache is not None else self.config.use_cache return_dict = return_dict if return_dict is not None else self.config.use_return_dict if encoder_outputs is None: encoder_outputs = self.encoder( input_features, attention_mask=attention_mask, head_mask=head_mask, output_attentions=output_attentions, output_hidden_states=output_hidden_states, return_dict=return_dict, ) # If the user passed a tuple for encoder_outputs, we wrap it in a BaseModelOutput when return_dict=True elif return_dict and not isinstance(encoder_outputs, BaseModelOutput): encoder_outputs = BaseModelOutput( last_hidden_state=encoder_outputs[0], hidden_states=encoder_outputs[1] if len(encoder_outputs) > 1 else None, attentions=encoder_outputs[2] if len(encoder_outputs) > 2 else None, ) # downsample encoder attention mask if attention_mask is not None: encoder_attention_mask = self._get_feature_vector_attention_mask( encoder_outputs[0].shape[1], attention_mask ) else: encoder_attention_mask = None # decoder outputs consists of (dec_features, past_key_value, dec_hidden, dec_attn) decoder_outputs = self.decoder( input_ids=decoder_input_ids, attention_mask=decoder_attention_mask, encoder_hidden_states=encoder_outputs[0], encoder_attention_mask=encoder_attention_mask, head_mask=decoder_head_mask, cross_attn_head_mask=cross_attn_head_mask, past_key_values=past_key_values, inputs_embeds=decoder_inputs_embeds, use_cache=use_cache, output_attentions=output_attentions, output_hidden_states=output_hidden_states, return_dict=return_dict, cache_position=cache_position, ) if not return_dict: return decoder_outputs + encoder_outputs return Seq2SeqModelOutput( last_hidden_state=decoder_outputs.last_hidden_state, past_key_values=decoder_outputs.past_key_values, decoder_hidden_states=decoder_outputs.hidden_states, decoder_attentions=decoder_outputs.attentions, cross_attentions=decoder_outputs.cross_attentions, encoder_last_hidden_state=encoder_outputs.last_hidden_state, encoder_hidden_states=encoder_outputs.hidden_states, encoder_attentions=encoder_outputs.attentions, ) @auto_docstring( custom_intro=""" The Speech2Text Model with a language modeling head. Can be used for summarization. """ ) class Speech2TextForConditionalGeneration(Speech2TextPreTrainedModel, GenerationMixin): base_model_prefix = "model" _tied_weights_keys = ["lm_head.weight"] def __init__(self, config: Speech2TextConfig): super().__init__(config) self.model = Speech2TextModel(config) self.lm_head = nn.Linear(config.d_model, self.config.vocab_size, bias=False) # Initialize weights and apply final processing self.post_init() def get_encoder(self): return self.model.get_encoder() def get_decoder(self): return self.model.get_decoder() @auto_docstring def forward( self, input_features: Optional[torch.LongTensor] = None, attention_mask: Optional[torch.Tensor] = None, decoder_input_ids: Optional[torch.LongTensor] = None, decoder_attention_mask: Optional[torch.LongTensor] = None, head_mask: Optional[torch.Tensor] = None, decoder_head_mask: Optional[torch.Tensor] = None, cross_attn_head_mask: Optional[torch.Tensor] = None, encoder_outputs: Optional[tuple[tuple[torch.FloatTensor]]] = None, past_key_values: Optional[tuple[tuple[torch.FloatTensor]]] = None, decoder_inputs_embeds: Optional[torch.FloatTensor] = None, labels: Optional[torch.LongTensor] = None, use_cache: Optional[bool] = None, output_attentions: Optional[bool] = None, output_hidden_states: Optional[bool] = None, return_dict: Optional[bool] = None, cache_position: Optional[torch.Tensor] = None, ) -> Union[tuple[torch.FloatTensor], Seq2SeqLMOutput]: r""" decoder_input_ids (`torch.LongTensor` of shape `(batch_size, target_sequence_length)`, *optional*): Indices of decoder input sequence tokens in the vocabulary. Indices can be obtained using [`SpeechToTextTokenizer`]. See [`PreTrainedTokenizer.encode`] and [`PreTrainedTokenizer.__call__`] for details. [What are decoder input IDs?](../glossary#decoder-input-ids) SpeechToText uses the `eos_token_id` as the starting token for `decoder_input_ids` generation. If `past_key_values` is used, optionally only the last `decoder_input_ids` have to be input (see `past_key_values`). decoder_attention_mask (`torch.LongTensor` of shape `(batch_size, target_sequence_length)`, *optional*): Default behavior: generate a tensor that ignores pad tokens in `decoder_input_ids`. Causal mask will also be used by default. If you want to change padding behavior, you should read [`modeling_speech_to_text._prepare_decoder_attention_mask`] and modify to your needs. See diagram 1 in [the paper](https://huggingface.co/papers/1910.13461) for more information on the default strategy. cross_attn_head_mask (`torch.Tensor` of shape `(decoder_layers, decoder_attention_heads)`, *optional*): Mask to nullify selected heads of the cross-attention modules. Mask values selected in `[0, 1]`: - 1 indicates the head is **not masked**, - 0 indicates the head is **masked**. labels (`torch.LongTensor` of shape `(batch_size, sequence_length)`, *optional*): Labels for computing the language modeling loss. Indices should either be in `[0, ..., config.vocab_size]` or -100 (see `input_ids` docstring). Tokens with indices set to `-100` are ignored (masked), the loss is only computed for the tokens with labels in `[0, ..., config.vocab_size]`. Example: ```python >>> import torch >>> from transformers import Speech2TextProcessor, Speech2TextForConditionalGeneration >>> from datasets import load_dataset >>> model = Speech2TextForConditionalGeneration.from_pretrained("facebook/s2t-small-librispeech-asr") >>> processor = Speech2TextProcessor.from_pretrained("facebook/s2t-small-librispeech-asr") >>> ds = load_dataset("hf-internal-testing/librispeech_asr_dummy", "clean", split="validation") >>> inputs = processor( ... ds[0]["audio"]["array"], sampling_rate=ds[0]["audio"]["sampling_rate"], return_tensors="pt" ... ) >>> input_features = inputs.input_features >>> generated_ids = model.generate(inputs=input_features) >>> transcription = processor.batch_decode(generated_ids, skip_special_tokens=True)[0] >>> transcription 'mister quilter is the apostle of the middle classes and we are glad to welcome his gospel' ```""" return_dict = return_dict if return_dict is not None else self.config.use_return_dict if labels is not None: if decoder_input_ids is None and decoder_inputs_embeds is None: decoder_input_ids = shift_tokens_right( labels, self.config.pad_token_id, self.config.decoder_start_token_id ) outputs = self.model( input_features, attention_mask=attention_mask, decoder_input_ids=decoder_input_ids, encoder_outputs=encoder_outputs, decoder_attention_mask=decoder_attention_mask, head_mask=head_mask, decoder_head_mask=decoder_head_mask, cross_attn_head_mask=cross_attn_head_mask, past_key_values=past_key_values, decoder_inputs_embeds=decoder_inputs_embeds, use_cache=use_cache, output_attentions=output_attentions, output_hidden_states=output_hidden_states, return_dict=return_dict, cache_position=cache_position, ) lm_logits = self.lm_head(outputs[0]) loss = None if labels is not None: loss_fct = CrossEntropyLoss() loss = loss_fct(lm_logits.view(-1, self.config.vocab_size), labels.view(-1)) if not return_dict: output = (lm_logits,) + outputs[1:] return ((loss,) + output) if loss is not None else output return Seq2SeqLMOutput( loss=loss, logits=lm_logits, past_key_values=outputs.past_key_values, decoder_hidden_states=outputs.decoder_hidden_states, decoder_attentions=outputs.decoder_attentions, cross_attentions=outputs.cross_attentions, encoder_last_hidden_state=outputs.encoder_last_hidden_state, encoder_hidden_states=outputs.encoder_hidden_states, encoder_attentions=outputs.encoder_attentions, ) __all__ = ["Speech2TextForConditionalGeneration", "Speech2TextModel", "Speech2TextPreTrainedModel"]