transformers/src/transformers/models/bart/modeling_bart.py

# coding=utf-8
# Copyright 2020 The Facebook AI Research Team Authors and The HuggingFace Inc. team.
#
# 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 BART model, ported from the fairseq repo."""
import math
import random
import warnings
from typing import Dict, Optional, Tuple

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

from ...activations import ACT2FN
from ...file_utils import (
    add_code_sample_docstrings,
    add_end_docstrings,
    add_start_docstrings,
    add_start_docstrings_to_model_forward,
    replace_return_docstrings,
)
from ...modeling_outputs import (
    BaseModelOutput,
    BaseModelOutputWithPastAndCrossAttentions,
    Seq2SeqLMOutput,
    Seq2SeqModelOutput,
    Seq2SeqQuestionAnsweringModelOutput,
    Seq2SeqSequenceClassifierOutput,
)
from ...modeling_utils import PreTrainedModel
from ...utils import logging
from .configuration_bart import BartConfig


logger = logging.get_logger(__name__)

_CONFIG_FOR_DOC = "BartConfig"
_TOKENIZER_FOR_DOC = "BartTokenizer"


BART_PRETRAINED_MODEL_ARCHIVE_LIST = [
    "facebook/bart-base",
    "facebook/bart-large",
    "facebook/bart-large-mnli",
    "facebook/bart-large-cnn",
    "facebook/bart-large-xsum",
    "facebook/mbart-large-en-ro",
]
# This list is incomplete. See all BART models at https://huggingface.co/models?filter=bart


def shift_tokens_right(input_ids: torch.Tensor, pad_token_id: int):
    """
    Shift input ids one token to the right, and wrap the last non pad token (usually <eos>).
    """
    prev_output_tokens = input_ids.clone()
    index_of_eos = (input_ids.ne(pad_token_id).sum(dim=1) - 1).unsqueeze(-1)
    prev_output_tokens[:, 0] = input_ids.gather(1, index_of_eos).squeeze()
    prev_output_tokens[:, 1:] = input_ids[:, :-1]
    return prev_output_tokens


def _make_causal_mask(input_ids_shape: torch.Size, dtype: torch.dtype, past_key_values_length: int = 0):
    """
    Make causal mask used for bi-directional self-attention.
    """
    bsz, tgt_len = input_ids_shape
    mask = torch.full((tgt_len, tgt_len), float("-inf"))
    mask_cond = torch.arange(mask.size(-1))
    mask.masked_fill_(mask_cond < (mask_cond + 1).view(mask.size(-1), 1), 0)
    mask = mask.to(dtype)

    if past_key_values_length > 0:
        mask = torch.cat([torch.zeros(tgt_len, past_key_values_length, dtype=dtype), mask], dim=-1)
    return mask[None, None, :, :].expand(bsz, 1, tgt_len, tgt_len + past_key_values_length)


def _expand_mask(
    mask: torch.Tensor, dtype: torch.dtype, tgt_len: Optional[int] = None, past_key_values_length: int = 0
):
    """
    Expands attention_mask from `[bsz, seq_len]` to `[bsz, 1, tgt_seq_len, src_seq_len]`.
    """
    bsz, src_len = mask.size()
    tgt_len = tgt_len if tgt_len is not None else src_len

    expanded_mask = mask[:, None, None, :].expand(bsz, 1, tgt_len, src_len).to(dtype)

    if past_key_values_length > 0:
        # concat fully attendend attention_mask to the beginning if `past_key_values` are used
        expanded_mask = torch.cat(
            [
                torch.ones(bsz, 1, tgt_len, past_key_values_length, device=expanded_mask.device, dtype=dtype),
                expanded_mask,
            ],
            dim=-1,
        )

    inverted_mask = 1.0 - expanded_mask

    return inverted_mask.masked_fill(inverted_mask.bool(), torch.finfo(dtype).min)


def BartLayerNorm(normalized_shape: torch.Size, eps: float = 1e-5, elementwise_affine: bool = True):
    if torch.cuda.is_available():
        try:
            from apex.normalization import FusedLayerNorm

            return FusedLayerNorm(normalized_shape, eps, elementwise_affine)
        except ImportError:
            pass
    return torch.nn.LayerNorm(normalized_shape, eps, elementwise_affine)


class BartLearnedPositionalEmbedding(nn.Embedding):
    """
    This module learns positional embeddings up to a fixed maximum size. Padding ids are ignored by either offsetting
    based on padding_idx or by setting padding_idx to None and ensuring that the appropriate position ids are passed to
    the forward function.
    """

    def __init__(self, num_embeddings: int, embedding_dim: int, padding_idx: int, offset: int):
        # Bart is set up so that if padding_idx is specified then offset the embedding ids by 2
        # and adjust num_embeddings appropriately. Other models dont have this hack
        self.offset = offset
        assert padding_idx is not None, "`padding_idx` should not be None, but of type int"
        num_embeddings += offset
        super().__init__(num_embeddings, embedding_dim, padding_idx=padding_idx)

    def forward(self, input_ids_shape: torch.Size, past_key_values_length: int = 0):
        """`input_ids_shape` is expected to be [bsz x seqlen]."""
        bsz, seq_len = input_ids_shape[:2]
        positions = torch.arange(
            past_key_values_length, past_key_values_length + seq_len, dtype=torch.long, device=self.weight.device
        )
        return super().forward(positions + self.offset)


class BartSinusoidalPositionalEmbedding(nn.Embedding):
    """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__(num_positions, embedding_dim)
        self.weight = self._init_weight(self.weight)

    @staticmethod
    def _init_weight(out: nn.Parameter):
        """
        Identical to the XLM create_sinusoidal_embeddings except features are not interleaved. The cos features are in
        the 2nd half of the vector. [dim // 2:]
        """
        n_pos, dim = out.shape
        position_enc = np.array(
            [[pos / np.power(10000, 2 * (j // 2) / dim) for j in range(dim)] for pos in range(n_pos)]
        )
        out.requires_grad = False  # set early to avoid an error in pytorch-1.8+
        sentinel = dim // 2 if dim % 2 == 0 else (dim // 2) + 1
        out[:, 0:sentinel] = torch.FloatTensor(np.sin(position_enc[:, 0::2]))
        out[:, sentinel:] = torch.FloatTensor(np.cos(position_enc[:, 1::2]))
        out.detach_()
        return out

    @torch.no_grad()
    def forward(self, input_ids_shape: torch.Size, past_key_values_length: int = 0):
        """`input_ids_shape` is expected to be [bsz x seqlen]."""
        bsz, seq_len = input_ids_shape[:2]
        positions = torch.arange(
            past_key_values_length, past_key_values_length + seq_len, dtype=torch.long, device=self.weight.device
        )
        return super().forward(positions)


class BartAttention(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,
    ):
        super().__init__()
        self.embed_dim = embed_dim
        self.num_heads = num_heads
        self.dropout = dropout
        self.head_dim = embed_dim // num_heads
        assert (
            self.head_dim * num_heads == self.embed_dim
        ), f"embed_dim must be divisible by num_heads (got `embed_dim`: {self.embed_dim} and `num_heads`: {num_heads})."
        self.scaling = self.head_dim ** -0.5
        self.is_decoder = is_decoder

        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,
        key_value_states: Optional[torch.Tensor] = None,
        past_key_value: Optional[Tuple[torch.Tensor]] = None,
        attn_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, embed_dim = hidden_states.size()

        # get query proj
        query_states = self.q_proj(hidden_states) * self.scaling
        # get key, value proj
        if is_cross_attention and past_key_value is not None:
            # reuse k,v, cross_attentions
            key_states = past_key_value[0]
            value_states = past_key_value[1]
        elif is_cross_attention:
            # cross_attentions
            key_states = self._shape(self.k_proj(key_value_states), -1, bsz)
            value_states = self._shape(self.v_proj(key_value_states), -1, bsz)
        elif past_key_value is not None:
            # reuse k, v, self_attention
            key_states = self._shape(self.k_proj(hidden_states), -1, bsz)
            value_states = self._shape(self.v_proj(hidden_states), -1, bsz)
            key_states = 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.view(*proj_shape)
        value_states = value_states.view(*proj_shape)

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

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

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

        attn_weights = F.softmax(attn_weights, dim=-1)

        if output_attentions:
            # this operation is a bit akward, but it's required to
            # make sure that attn_weights keeps its gradient.
            # In order to do so, attn_weights have to 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 = F.dropout(attn_weights, p=self.dropout, training=self.training)

        attn_output = torch.bmm(attn_probs, value_states)

        assert attn_output.size() == (
            bsz * self.num_heads,
            tgt_len,
            self.head_dim,
        ), f"`attn_output` should be of size {(bsz, self.num_heads, tgt_len, self.head_dim)}, but is {attn_output.size()}"

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

        attn_output = self.out_proj(attn_output)

        return attn_output, attn_weights_reshaped, past_key_value


class BartEncoderLayer(nn.Module):
    def __init__(self, config: BartConfig):
        super().__init__()
        self.embed_dim = config.d_model
        self.self_attn = BartAttention(
            embed_dim=self.embed_dim,
            num_heads=config.encoder_attention_heads,
            dropout=config.attention_dropout,
        )
        self.normalize_before = config.normalize_before
        self.self_attn_layer_norm = BartLayerNorm(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 = BartLayerNorm(self.embed_dim)

    def forward(
        self, hidden_states: torch.Tensor, encoder_padding_mask: torch.Tensor, output_attentions: bool = False
    ):
        """
        Args:
            hidden_states (:obj:`torch.FloatTensor`): input to the layer of shape `(seq_len, batch, embed_dim)`
            encoder_padding_mask (:obj:`torch.FloatTensor`): attention mask of size
                `(batch, 1, tgt_len, src_len)` where padding elements are indicated by very large negative values.

        Returns:
            encoded output of shape `(seq_len, batch, embed_dim)`
        """
        residual = hidden_states
        if self.normalize_before:
            hidden_states = self.self_attn_layer_norm(hidden_states)
        hidden_states, attn_weights, _ = self.self_attn(
            hidden_states=hidden_states, attn_mask=encoder_padding_mask, output_attentions=output_attentions
        )
        hidden_states = F.dropout(hidden_states, p=self.dropout, training=self.training)
        hidden_states = residual + hidden_states
        if not self.normalize_before:
            hidden_states = self.self_attn_layer_norm(hidden_states)

        residual = hidden_states
        if self.normalize_before:
            hidden_states = self.final_layer_norm(hidden_states)
        hidden_states = self.activation_fn(self.fc1(hidden_states))
        hidden_states = F.dropout(hidden_states, p=self.activation_dropout, training=self.training)
        hidden_states = self.fc2(hidden_states)
        hidden_states = F.dropout(hidden_states, p=self.dropout, training=self.training)
        hidden_states = residual + hidden_states
        if not self.normalize_before:
            hidden_states = self.final_layer_norm(hidden_states)
        if torch.isinf(hidden_states).any() or torch.isnan(hidden_states).any():
            clamp_value = torch.finfo(hidden_states.dtype).max - 1000
            hidden_states = torch.clamp(hidden_states, min=-clamp_value, max=clamp_value)
        return hidden_states, attn_weights


class BartDecoderLayer(nn.Module):
    def __init__(self, config: BartConfig):
        super().__init__()
        self.embed_dim = config.d_model

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

        self.self_attn_layer_norm = BartLayerNorm(self.embed_dim)
        self.encoder_attn = BartAttention(
            self.embed_dim,
            config.decoder_attention_heads,
            dropout=config.attention_dropout,
            is_decoder=True,
        )
        self.encoder_attn_layer_norm = BartLayerNorm(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 = BartLayerNorm(self.embed_dim)

    def forward(
        self,
        hidden_states: torch.Tensor,
        encoder_hidden_states: torch.Tensor,
        encoder_attn_mask: Optional[torch.Tensor] = None,
        past_key_value: Optional[Tuple[Tuple[torch.Tensor]]] = None,
        attn_mask: Optional[torch.Tensor] = None,
        output_attentions: Optional[torch.Tensor] = False,
    ):
        residual = hidden_states
        if self.normalize_before:
            hidden_states = self.self_attn_layer_norm(hidden_states)
        # Self Attention

        # decoder uni-directional self-attention cached key/values tuple is at first position
        self_attn_past_key_value = past_key_value[0] if past_key_value is not None else None
        hidden_states, self_attn_weights, self_attn_present_key_value = self.self_attn(
            hidden_states=hidden_states,
            past_key_value=self_attn_past_key_value,
            attn_mask=attn_mask,
            output_attentions=output_attentions,
        )
        hidden_states = F.dropout(hidden_states, p=self.dropout, training=self.training)
        hidden_states = residual + hidden_states
        if not self.normalize_before:
            hidden_states = self.self_attn_layer_norm(hidden_states)

        # Cross-Attention Block
        cross_attn_present_key_value = None
        cross_attn_weights = None
        if encoder_hidden_states is not None:
            residual = hidden_states
            if self.normalize_before:
                hidden_states = self.encoder_attn_layer_norm(hidden_states)

            # cross_attn cached key/values tuple is at second position
            cross_attn_past_key_value = past_key_value[1] if past_key_value is not None else None
            hidden_states, cross_attn_weights, cross_attn_present_key_value = self.encoder_attn(
                hidden_states=hidden_states,
                key_value_states=encoder_hidden_states,
                attn_mask=encoder_attn_mask,
                past_key_value=cross_attn_past_key_value,
                output_attentions=output_attentions,
            )
            hidden_states = F.dropout(hidden_states, p=self.dropout, training=self.training)
            hidden_states = residual + hidden_states
            if not self.normalize_before:
                hidden_states = self.encoder_attn_layer_norm(hidden_states)

        # Fully Connected
        residual = hidden_states
        if self.normalize_before:
            hidden_states = self.final_layer_norm(hidden_states)
        hidden_states = self.activation_fn(self.fc1(hidden_states))
        hidden_states = F.dropout(hidden_states, p=self.activation_dropout, training=self.training)
        hidden_states = self.fc2(hidden_states)
        hidden_states = F.dropout(hidden_states, p=self.dropout, training=self.training)
        hidden_states = residual + hidden_states
        if not self.normalize_before:
            hidden_states = self.final_layer_norm(hidden_states)

        # make sure decoder uni-directional self-attn at 1st position and cross-attn at 2nd position.
        present_key_value = (self_attn_present_key_value, cross_attn_present_key_value)

        return (
            hidden_states,
            self_attn_weights,
            present_key_value,
            cross_attn_weights,
        )


class BartClassificationHead(nn.Module):
    """Head for sentence-level classification tasks."""

    def __init__(
        self,
        input_dim: int,
        inner_dim: int,
        num_classes: int,
        pooler_dropout: float,
    ):
        super().__init__()
        self.dense = nn.Linear(input_dim, inner_dim)
        self.dropout = nn.Dropout(p=pooler_dropout)
        self.out_proj = nn.Linear(inner_dim, num_classes)

    def forward(self, hidden_states: torch.Tensor):
        hidden_states = self.dropout(hidden_states)
        hidden_states = self.dense(hidden_states)
        hidden_states = torch.tanh(hidden_states)
        hidden_states = self.dropout(hidden_states)
        hidden_states = self.out_proj(hidden_states)
        return hidden_states


class BartPretrainedModel(PreTrainedModel):
    config_class = BartConfig
    base_model_prefix = "model"

    def _init_weights(self, module):
        std = self.config.init_std
        if isinstance(module, nn.Linear):
            module.weight.data.normal_(mean=0.0, std=std)
            if module.bias is not None:
                module.bias.data.zero_()
        elif isinstance(module, BartSinusoidalPositionalEmbedding):
            pass
        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_()

    @property
    def dummy_inputs(self):
        pad_token = self.config.pad_token_id
        input_ids = torch.tensor([[0, 6, 10, 4, 2], [0, 8, 12, 2, pad_token]], device=self.device)
        dummy_inputs = {
            "attention_mask": input_ids.ne(pad_token),
            "input_ids": input_ids,
        }
        return dummy_inputs


class PretrainedBartModel(BartPretrainedModel):
    def __init_subclass__(self):
        warnings.warn(
            "The class `PretrainedBartModel` has been depreciated, please use `BartPretrainedModel` instead.",
            FutureWarning,
        )


BART_START_DOCSTRING = r"""
    This model inherits from :class:`~transformers.PreTrainedModel`. Check the superclass documentation for the generic
    methods the library implements for all its model (such as downloading or saving, resizing the input embeddings,
    pruning heads etc.)

    This model is also a PyTorch `torch.nn.Module <https://pytorch.org/docs/stable/nn.html#torch.nn.Module>`__
    subclass. Use it as a regular PyTorch Module and refer to the PyTorch documentation for all matter related to
    general usage and behavior.

    Parameters:
        config (:class:`~transformers.BartConfig`): 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 :meth:`~transformers.PreTrainedModel.from_pretrained` method to load the model
            weights.
"""

BART_GENERATION_EXAMPLE = r"""
    Summarization example::

        >>> from transformers import BartTokenizer, BartForConditionalGeneration, BartConfig

        >>> # see ``examples/summarization/bart/run_eval.py`` for a longer example
        >>> model = BartForConditionalGeneration.from_pretrained('facebook/bart-large-cnn')
        >>> tokenizer = BartTokenizer.from_pretrained('facebook/bart-large-cnn')

        >>> ARTICLE_TO_SUMMARIZE = "My friends are cool but they eat too many carbs."
        >>> inputs = tokenizer([ARTICLE_TO_SUMMARIZE], max_length=1024, return_tensors='pt')

        >>> # Generate Summary
        >>> summary_ids = model.generate(inputs['input_ids'], num_beams=4, max_length=5, early_stopping=True)
        >>> print([tokenizer.decode(g, skip_special_tokens=True, clean_up_tokenization_spaces=False) for g in summary_ids])
"""

BART_INPUTS_DOCSTRING = r"""
    Args:
        input_ids (:obj:`torch.LongTensor` of shape :obj:`(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 :class:`~transformers.BartTokenizer`. See
            :meth:`transformers.PreTrainedTokenizer.encode` and :meth:`transformers.PreTrainedTokenizer.__call__` for
            details.

            `What are input IDs? <../glossary.html#input-ids>`__
        attention_mask (:obj:`torch.Tensor` of shape :obj:`(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.html#attention-mask>`__
        decoder_input_ids (:obj:`torch.LongTensor` of shape :obj:`(batch_size, target_sequence_length)`, `optional`):
            Provide for translation and summarization training. By default, the model will create this tensor by
            shifting the :obj:`input_ids` to the right, following the paper.
        decoder_attention_mask (:obj:`torch.LongTensor` of shape :obj:`(batch_size, tgt_seq_len)`, `optional`):
            Default behavior: generate a tensor that ignores pad tokens in :obj:`decoder_input_ids`. Causal mask will
            also be used by default.

            If you want to change padding behavior, you should read :func:`modeling_bart._prepare_decoder_inputs` and
            modify to your needs. See diagram 1 in `the paper <https://arxiv.org/abs/1910.13461>`__ for more
            information on the default strategy.
        encoder_outputs (:obj:`tuple(tuple(torch.FloatTensor)`, `optional`):
            Tuple consists of (:obj:`last_hidden_state`, `optional`: :obj:`hidden_states`, `optional`:
            :obj:`attentions`) :obj:`last_hidden_state` of shape :obj:`(batch_size, sequence_length, hidden_size)`,
            `optional`) is a sequence of hidden-states at the output of the last layer of the encoder. Used in the
            cross-attention of the decoder.
        past_key_values (:obj:`Tuple[Tuple[Tuple[torch.Tensor]]]` of length :obj:`config.n_layers` with each tuple having 2 tuples each of which has 2 tensors of shape :obj:`(batch_size, num_heads, sequence_length - 1, embed_size_per_head)`):
            Contains precomputed key and value hidden-states of the attention blocks. Can be used to speed up decoding.

            If :obj:`past_key_values` are used, the user can optionally input only the last :obj:`decoder_input_ids`
            (those that don't have their past key value states given to this model) of shape :obj:`(batch_size, 1)`
            instead of all :obj:`decoder_input_ids`` of shape :obj:`(batch_size, sequence_length)`.
        inputs_embeds (:obj:`torch.FloatTensor` of shape :obj:`(batch_size, sequence_length, hidden_size)`, `optional`):
            Optionally, instead of passing :obj:`input_ids` you can choose to directly pass an embedded representation.
            This is useful if you want more control over how to convert :obj:`input_ids` indices into associated
            vectors than the model's internal embedding lookup matrix.
        decoder_inputs_embeds (:obj:`torch.FloatTensor` of shape :obj:`(batch_size, target_sequence_length, hidden_size)`, `optional`):
            Optionally, instead of passing :obj:`decoder_input_ids` you can choose to directly pass an embedded
            representation. If :obj:`past_key_values` is used, optionally only the last :obj:`decoder_inputs_embeds`
            have to be input (see :obj:`past_key_values`). This is useful if you want more control over how to convert
            :obj:`decoder_input_ids` indices into associated vectors than the model's internal embedding lookup matrix.

            If :obj:`decoder_input_ids` and :obj:`decoder_inputs_embeds` are both unset, :obj:`decoder_inputs_embeds`
            takes the value of :obj:`inputs_embeds`.
        use_cache (:obj:`bool`, `optional`):
            If set to :obj:`True`, :obj:`past_key_values` key value states are returned and can be used to speed up
            decoding (see :obj:`past_key_values`).
        output_attentions (:obj:`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 (:obj:`bool`, `optional`):
            Whether or not to return the hidden states of all layers. See ``hidden_states`` under returned tensors for
            more detail.
        return_dict (:obj:`bool`, `optional`):
            Whether or not to return a :class:`~transformers.file_utils.ModelOutput` instead of a plain tuple.
"""


class BartEncoder(BartPretrainedModel):
    """
    Transformer encoder consisting of *config.encoder_layers* self attention layers. Each layer is a
    :class:`BartEncoderLayer`.

    Args:
        config: BartConfig
        embed_tokens (torch.nn.Embedding): output embedding
    """

    def __init__(self, config: BartConfig, embed_tokens: Optional[nn.Embedding] = None):
        super().__init__(config)

        self.dropout = config.dropout
        self.layerdrop = config.encoder_layerdrop

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

        if embed_tokens is not None:
            self.embed_tokens = embed_tokens
        else:
            self.embed_tokens = nn.Embedding(config.vocab_size, embed_dim, self.padding_idx)

        if config.static_position_embeddings:
            self.embed_positions = BartSinusoidalPositionalEmbedding(
                config.max_position_embeddings, embed_dim, self.padding_idx
            )
        else:
            self.embed_positions = BartLearnedPositionalEmbedding(
                config.max_position_embeddings,
                embed_dim,
                self.padding_idx,
                config.extra_pos_embeddings,
            )
        self.layers = nn.ModuleList([BartEncoderLayer(config) for _ in range(config.encoder_layers)])
        self.layernorm_embedding = BartLayerNorm(embed_dim) if config.normalize_embedding else nn.Identity()
        # mbart has one extra layer_norm
        self.layer_norm = BartLayerNorm(config.d_model) if config.add_final_layer_norm else None

        self.init_weights()

    def forward(
        self,
        input_ids=None,
        attention_mask=None,
        inputs_embeds=None,
        output_attentions=None,
        output_hidden_states=None,
        return_dict=None,
    ):
        r"""
        Args:
            input_ids (:obj:`torch.LongTensor` of shape :obj:`(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 :class:`~transformers.BartTokenizer`. See
                :meth:`transformers.PreTrainedTokenizer.encode` and :meth:`transformers.PreTrainedTokenizer.__call__`
                for details.

                `What are input IDs? <../glossary.html#input-ids>`__
            attention_mask (:obj:`torch.Tensor` of shape :obj:`(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.html#attention-mask>`__
            inputs_embeds (:obj:`torch.FloatTensor` of shape :obj:`(batch_size, sequence_length, hidden_size)`, `optional`):
                Optionally, instead of passing :obj:`input_ids` you can choose to directly pass an embedded
                representation. This is useful if you want more control over how to convert :obj:`input_ids` indices
                into associated vectors than the model's internal embedding lookup matrix.
            output_attentions (:obj:`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 (:obj:`bool`, `optional`):
                Whether or not to return the hidden states of all layers. See ``hidden_states`` under returned tensors
                for more detail.
            return_dict (:obj:`bool`, `optional`):
                Whether or not to return a :class:`~transformers.file_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

        # retrieve input_ids and inputs_embeds
        if input_ids is not None and inputs_embeds is not None:
            raise ValueError("You cannot specify both input_ids and 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 input_ids or inputs_embeds")

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

        embed_pos = self.embed_positions(input_shape)

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

        # expand attention_mask
        if attention_mask is not None:
            # [bsz, seq_len] -> [bsz, 1, tgt_seq_len, src_seq_len]
            attention_mask = _expand_mask(attention_mask, inputs_embeds.dtype)

        encoder_states = () if output_hidden_states else None
        all_attentions = () if output_attentions else None
        for encoder_layer in self.layers:
            if output_hidden_states:
                encoder_states = encoder_states + (hidden_states,)
            # add LayerDrop (see https://arxiv.org/abs/1909.11556 for description)
            dropout_probability = random.uniform(0, 1)
            if self.training and (dropout_probability < self.layerdrop):  # skip the layer
                attn = None
            else:
                hidden_states, attn = encoder_layer(hidden_states, attention_mask, output_attentions=output_attentions)

            if output_attentions:
                all_attentions = all_attentions + (attn,)

        if self.layer_norm:
            hidden_states = self.layer_norm(hidden_states)

        if output_hidden_states:
            encoder_states = encoder_states + (hidden_states,)

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


class BartDecoder(BartPretrainedModel):
    """
    Transformer decoder consisting of *config.decoder_layers* layers. Each layer is a :class:`BartDecoderLayer`

    Args:
        config: BartConfig
        embed_tokens (torch.nn.Embedding): output embedding
    """

    def __init__(self, config: BartConfig, embed_tokens: Optional[nn.Embedding] = None):
        super().__init__(config)
        self.dropout = config.dropout
        self.layerdrop = config.decoder_layerdrop
        self.do_blenderbot_90_layernorm = config.do_blenderbot_90_layernorm  # layernorm variant
        self.padding_idx = config.pad_token_id
        self.max_target_positions = config.max_position_embeddings
        self.embed_scale = math.sqrt(config.d_model) if config.scale_embedding else 1.0

        if embed_tokens is not None:
            self.embed_tokens = embed_tokens
        else:
            self.embed_tokens = nn.Embedding(config.vocab_size, config.d_model, self.padding_idx)

        if config.static_position_embeddings:
            self.embed_positions = BartSinusoidalPositionalEmbedding(
                config.max_position_embeddings, config.d_model, config.pad_token_id
            )
        else:
            self.embed_positions = BartLearnedPositionalEmbedding(
                config.max_position_embeddings,
                config.d_model,
                self.padding_idx,
                config.extra_pos_embeddings,
            )
        self.layers = nn.ModuleList([BartDecoderLayer(config) for _ in range(config.decoder_layers)])
        self.layernorm_embedding = BartLayerNorm(config.d_model) if config.normalize_embedding else nn.Identity()
        self.layer_norm = BartLayerNorm(config.d_model) if config.add_final_layer_norm else None

        self.init_weights()

    def forward(
        self,
        input_ids=None,
        attention_mask=None,
        encoder_hidden_states=None,
        encoder_attention_mask=None,
        past_key_values=None,
        inputs_embeds=None,
        use_cache=None,
        output_attentions=None,
        output_hidden_states=None,
        return_dict=None,
    ):
        r"""
        Args:
            input_ids (:obj:`torch.LongTensor` of shape :obj:`(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 :class:`~transformers.BartTokenizer`. See
                :meth:`transformers.PreTrainedTokenizer.encode` and :meth:`transformers.PreTrainedTokenizer.__call__`
                for details.

                `What are input IDs? <../glossary.html#input-ids>`__
            attention_mask (:obj:`torch.Tensor` of shape :obj:`(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.html#attention-mask>`__
            encoder_hidden_states (:obj:`torch.FloatTensor` of shape :obj:`(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 (:obj:`torch.LongTensor` of shape :obj:`(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.html#attention-mask>`__
            past_key_values (:obj:`Tuple[Tuple[Tuple[torch.Tensor]]]` of length :obj:`config.n_layers` with each tuple having 2 tuples each of which has 2 tensors of shape :obj:`(batch_size, num_heads, sequence_length - 1, embed_size_per_head)`):
                Contains precomputed key and value hidden-states of the attention blocks. Can be used to speed up
                decoding.

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

        # past_key_values_length
        past_key_values_length = past_key_values[0][0][0].shape[2] if past_key_values is not None else 0

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

        # create causal mask
        # [bsz, seq_len] -> [bsz, 1, tgt_seq_len, src_seq_len]
        attn_mask = None
        if input_shape[-1] > 1:
            attn_mask = _make_causal_mask(
                input_shape, inputs_embeds.dtype, past_key_values_length=past_key_values_length
            ).to(self.device)

        # create decoder_padding_mask if not provided and needed
        # 4.12.20 (PVP): Not a fan of this "magical" function that
        # automatically creates attention_mask for padded tokens
        # => this is inconsistent with other models
        # => Pegasus uses the pad_token as decoder_start_token_id, so that this could
        # pose some problems.
        if (
            attention_mask is None
            and input_ids is not None
            and input_shape[-1] > 1
            and self.config.pad_token_id in input_ids
        ):
            # should be kept for backwards compatibility
            attention_mask = input_ids.ne(self.config.pad_token_id).to(torch.long)
            # never mask leading token, even if it is pad
            attention_mask[:, 0] = attention_mask[:, 1]

        if attention_mask is not None and attn_mask is not None:
            # [bsz, seq_len] -> [bsz, 1, tgt_seq_len, src_seq_len]
            attn_mask = attn_mask + _expand_mask(
                attention_mask, inputs_embeds.dtype, past_key_values_length=past_key_values_length
            )

        # expand encoder attention mask
        if encoder_hidden_states is not None and encoder_attention_mask is not None:
            # [bsz, seq_len] -> [bsz, 1, tgt_seq_len, src_seq_len]
            encoder_attention_mask = _expand_mask(encoder_attention_mask, inputs_embeds.dtype, tgt_len=input_shape[-1])

        # embed positions
        positions = self.embed_positions(input_shape, past_key_values_length)

        if self.do_blenderbot_90_layernorm:
            hidden_states = self.layernorm_embedding(inputs_embeds)
            hidden_states += positions
        else:
            hidden_states = inputs_embeds + positions
            hidden_states = self.layernorm_embedding(hidden_states)

        hidden_states = F.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 else None
        next_decoder_cache = () if use_cache else None
        for idx, decoder_layer in enumerate(self.layers):
            # add LayerDrop (see https://arxiv.org/abs/1909.11556 for description)
            if output_hidden_states:
                all_hidden_states += (hidden_states,)
            dropout_probability = random.uniform(0, 1)
            if self.training and (dropout_probability < self.layerdrop):
                continue

            past_key_value = past_key_values[idx] if past_key_values is not None else None

            hidden_states, layer_self_attn, present_key_value, layer_cross_attn = decoder_layer(
                hidden_states,
                encoder_hidden_states,
                encoder_attn_mask=encoder_attention_mask,
                attn_mask=attn_mask,
                past_key_value=past_key_value,
                output_attentions=output_attentions,
            )

            if use_cache:
                next_decoder_cache += (present_key_value,)

            if output_attentions:
                all_self_attns += (layer_self_attn,)
                all_cross_attentions += (layer_cross_attn,)

        # add hidden states from the last decoder layer
        if output_hidden_states:
            all_hidden_states += (hidden_states,)

        # if config.add_final_layer_norm (mBART)
        if self.layer_norm:
            hidden_states = self.layer_norm(hidden_states)

        next_cache = next_decoder_cache if use_cache else None
        if not return_dict:
            return tuple(
                v
                for v in [hidden_states, next_cache, all_hidden_states, all_self_attns, all_cross_attentions]
                if v is not None
            )
        return BaseModelOutputWithPastAndCrossAttentions(
            last_hidden_state=hidden_states,
            past_key_values=next_cache,
            hidden_states=all_hidden_states,
            attentions=all_self_attns,
            cross_attentions=all_cross_attentions,
        )


@add_start_docstrings(
    "The bare BART Model outputting raw hidden-states without any specific head on top.",
    BART_START_DOCSTRING,
)
class BartModel(BartPretrainedModel):
    def __init__(self, config: BartConfig):
        super().__init__(config)

        padding_idx, vocab_size = config.pad_token_id, config.vocab_size
        self.shared = nn.Embedding(vocab_size, config.d_model, padding_idx)

        self.encoder = BartEncoder(config, self.shared)
        self.decoder = BartDecoder(config, self.shared)

        self.init_weights()

    def get_input_embeddings(self):
        return self.shared

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

    def get_encoder(self):
        return self.encoder

    def get_decoder(self):
        return self.decoder

    @add_start_docstrings_to_model_forward(BART_INPUTS_DOCSTRING)
    @add_code_sample_docstrings(
        tokenizer_class=_TOKENIZER_FOR_DOC,
        checkpoint="facebook/bart-large",
        output_type=Seq2SeqModelOutput,
        config_class=_CONFIG_FOR_DOC,
    )
    def forward(
        self,
        input_ids=None,
        attention_mask=None,
        decoder_input_ids=None,
        decoder_attention_mask=None,
        encoder_outputs=None,
        past_key_values=None,
        inputs_embeds=None,
        decoder_inputs_embeds=None,
        use_cache=None,
        output_attentions=None,
        output_hidden_states=None,
        return_dict=None,
    ):

        # 4.12.20 (PVP): Not a fan of this "magical" function and
        # also wonder how often it's actually used ... keep now
        # for backward compatibility
        # -> is this used for backward compatibility
        if decoder_input_ids is None and decoder_inputs_embeds is None:
            decoder_input_ids = shift_tokens_right(input_ids, self.config.pad_token_id)

        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_ids=input_ids,
                attention_mask=attention_mask,
                inputs_embeds=inputs_embeds,
                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=False
        elif return_dict and not isinstance(encoder_outputs, BaseModelOutput):
            encoder_outputs = BaseModelOutput(
                last_hidden_state=encoder_outputs[0],
                hidden_states=encoder_outputs[1] if len(encoder_outputs) > 1 else None,
                attentions=encoder_outputs[2] if len(encoder_outputs) > 2 else None,
            )

        # decoder outputs consists of (dec_features, past_key_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=attention_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,
        )

        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,
        )


@add_start_docstrings(
    "The BART Model with a language modeling head. Can be used for summarization.", BART_START_DOCSTRING
)
class BartForConditionalGeneration(BartPretrainedModel):
    base_model_prefix = "model"
    _keys_to_ignore_on_load_missing = [
        r"final_logits_bias",
        r"encoder\.version",
        r"decoder\.version",
        r"lm_head\.weight",
    ]

    def __init__(self, config: BartConfig):
        super().__init__(config)
        self.model = BartModel(config)
        self.register_buffer("final_logits_bias", torch.zeros((1, self.model.shared.num_embeddings)))
        self.lm_head = nn.Linear(config.d_model, self.model.shared.num_embeddings, bias=False)

        self.init_weights()

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

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

    def resize_token_embeddings(self, new_num_tokens: int) -> nn.Embedding:
        new_embeddings = super().resize_token_embeddings(new_num_tokens)
        self._resize_final_logits_bias(new_num_tokens)
        return new_embeddings

    def _resize_final_logits_bias(self, new_num_tokens: int) -> None:
        old_num_tokens = self.final_logits_bias.shape[-1]
        if new_num_tokens <= old_num_tokens:
            new_bias = self.final_logits_bias[:, :new_num_tokens]
        else:
            extra_bias = torch.zeros((1, new_num_tokens - old_num_tokens), device=self.final_logits_bias.device)
            new_bias = torch.cat([self.final_logits_bias, extra_bias], dim=1)
        self.register_buffer("final_logits_bias", new_bias)

    def get_output_embeddings(self):
        return self.lm_head

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

    @add_start_docstrings_to_model_forward(BART_INPUTS_DOCSTRING)
    @replace_return_docstrings(output_type=Seq2SeqLMOutput, config_class=_CONFIG_FOR_DOC)
    @add_end_docstrings(BART_GENERATION_EXAMPLE)
    def forward(
        self,
        input_ids=None,
        attention_mask=None,
        decoder_input_ids=None,
        decoder_attention_mask=None,
        encoder_outputs=None,
        past_key_values=None,
        inputs_embeds=None,
        decoder_inputs_embeds=None,
        labels=None,
        use_cache=None,
        output_attentions=None,
        output_hidden_states=None,
        return_dict=None,
    ):
        r"""
        labels (:obj:`torch.LongTensor` of shape :obj:`(batch_size, sequence_length)`, `optional`):
            Labels for computing the masked language modeling loss. Indices should either be in ``[0, ...,
            config.vocab_size]`` or -100 (see ``input_ids`` docstring). Tokens with indices set to ``-100`` are ignored
            (masked), the loss is only computed for the tokens with labels in ``[0, ..., config.vocab_size]``.

        Returns:

        Conditional generation example::

            >>> # Mask filling only works for bart-large
            >>> from transformers import BartTokenizer, BartForConditionalGeneration
            >>> tokenizer = BartTokenizer.from_pretrained('facebook/bart-large')
            >>> TXT = "My friends are <mask> but they eat too many carbs."

            >>> model = BartForConditionalGeneration.from_pretrained('facebook/bart-large')
            >>> input_ids = tokenizer([TXT], return_tensors='pt')['input_ids']
            >>> logits = model(input_ids).logits

            >>> masked_index = (input_ids[0] == tokenizer.mask_token_id).nonzero().item()
            >>> probs = logits[0, masked_index].softmax(dim=0)
            >>> values, predictions = probs.topk(5)

            >>> tokenizer.decode(predictions).split()
            >>> # ['good', 'great', 'all', 'really', 'very']
        """
        return_dict = return_dict if return_dict is not None else self.config.use_return_dict

        if labels is not None:
            use_cache = False
            if decoder_input_ids is None:
                decoder_input_ids = shift_tokens_right(labels, self.config.pad_token_id)

        outputs = self.model(
            input_ids,
            attention_mask=attention_mask,
            decoder_input_ids=decoder_input_ids,
            encoder_outputs=encoder_outputs,
            decoder_attention_mask=decoder_attention_mask,
            past_key_values=past_key_values,
            inputs_embeds=inputs_embeds,
            decoder_inputs_embeds=decoder_inputs_embeds,
            use_cache=use_cache,
            output_attentions=output_attentions,
            output_hidden_states=output_hidden_states,
            return_dict=return_dict,
        )
        lm_logits = self.lm_head(outputs[0]) + self.final_logits_bias

        masked_lm_loss = None
        if labels is not None:
            loss_fct = CrossEntropyLoss()
            # TODO(SS): do we need to ignore pad tokens in labels?
            masked_lm_loss = loss_fct(lm_logits.view(-1, self.config.vocab_size), labels.view(-1))

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

        return Seq2SeqLMOutput(
            loss=masked_lm_loss,
            logits=lm_logits,
            past_key_values=outputs.past_key_values,
            decoder_hidden_states=outputs.decoder_hidden_states,
            decoder_attentions=outputs.decoder_attentions,
            cross_attentions=outputs.cross_attentions,
            encoder_last_hidden_state=outputs.encoder_last_hidden_state,
            encoder_hidden_states=outputs.encoder_hidden_states,
            encoder_attentions=outputs.encoder_attentions,
        )

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

        return {
            "input_ids": None,  # encoder_outputs is defined. input_ids not needed
            "encoder_outputs": encoder_outputs,
            "past_key_values": past,
            "decoder_input_ids": decoder_input_ids,
            "attention_mask": attention_mask,
            "use_cache": use_cache,  # change this to avoid caching (presumably for debugging)
        }

    def adjust_logits_during_generation(self, logits, cur_len, max_length):
        if cur_len == 1 and self.config.force_bos_token_to_be_generated:
            self._force_token_id_to_be_generated(logits, self.config.bos_token_id)
        elif cur_len == max_length - 1 and self.config.eos_token_id is not None:
            self._force_token_id_to_be_generated(logits, self.config.eos_token_id)
        return logits

    @staticmethod
    def _force_token_id_to_be_generated(scores, token_id) -> None:
        """force one of token_ids to be generated by setting prob of all other tokens to 0 (logprob=-float("inf"))"""
        scores[:, [x for x in range(scores.shape[1]) if x != token_id]] = -float("inf")

    @staticmethod
    def _reorder_cache(past, beam_idx):
        def _reorder_buffer(cache: Tuple[torch.Tensor], new_order) -> Dict:
            return tuple(past_state.index_select(0, new_order) for past_state in cache)

        reordered_past = ()
        for layer_past in past:
            reordered_past += (tuple(_reorder_buffer(cache, beam_idx) for cache in layer_past),)
        return reordered_past


@add_start_docstrings(
    """
    Bart model with a sequence classification/head on top (a linear layer on top of the pooled output) e.g. for GLUE
    tasks.
    """,
    BART_START_DOCSTRING,
)
class BartForSequenceClassification(BartPretrainedModel):
    def __init__(self, config: BartConfig, **kwargs):
        super().__init__(config, **kwargs)
        self.model = BartModel(config)
        self.classification_head = BartClassificationHead(
            config.d_model,
            config.d_model,
            config.num_labels,
            config.classifier_dropout,
        )
        self.model._init_weights(self.classification_head.dense)
        self.model._init_weights(self.classification_head.out_proj)

    @add_start_docstrings_to_model_forward(BART_INPUTS_DOCSTRING)
    @add_code_sample_docstrings(
        tokenizer_class=_TOKENIZER_FOR_DOC,
        checkpoint="facebook/bart-large",
        output_type=Seq2SeqSequenceClassifierOutput,
        config_class=_CONFIG_FOR_DOC,
    )
    def forward(
        self,
        input_ids=None,
        attention_mask=None,
        decoder_input_ids=None,
        decoder_attention_mask=None,
        encoder_outputs=None,
        inputs_embeds=None,
        decoder_inputs_embeds=None,
        labels=None,
        use_cache=None,
        output_attentions=None,
        output_hidden_states=None,
        return_dict=None,
    ):
        r"""
        labels (:obj:`torch.LongTensor` of shape :obj:`(batch_size,)`, `optional`):
            Labels for computing the sequence classification/regression loss. Indices should be in :obj:`[0, ...,
            config.num_labels - 1]`. If :obj:`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
        if labels is not None:
            use_cache = False

        if input_ids is None and inputs_embeds is not None:
            raise NotImplementedError(
                f"Passing input embeddings is currently not supported for {self.__class__.__name__}"
            )

        outputs = self.model(
            input_ids,
            attention_mask=attention_mask,
            decoder_input_ids=decoder_input_ids,
            decoder_attention_mask=decoder_attention_mask,
            encoder_outputs=encoder_outputs,
            inputs_embeds=inputs_embeds,
            decoder_inputs_embeds=decoder_inputs_embeds,
            use_cache=use_cache,
            output_attentions=output_attentions,
            output_hidden_states=output_hidden_states,
            return_dict=return_dict,
        )
        hidden_states = outputs[0]  # last hidden state

        eos_mask = input_ids.eq(self.config.eos_token_id)

        if len(torch.unique(eos_mask.sum(1))) > 1:
            raise ValueError("All examples must have the same number of <eos> tokens.")
        sentence_representation = hidden_states[eos_mask, :].view(hidden_states.size(0), -1, hidden_states.size(-1))[
            :, -1, :
        ]
        logits = self.classification_head(sentence_representation)

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

        return Seq2SeqSequenceClassifierOutput(
            loss=loss,
            logits=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,
        )


@add_start_docstrings(
    """
    BART Model with a span classification head on top for extractive question-answering tasks like SQuAD (a linear
    layer on top of the hidden-states output to compute `span start logits` and `span end logits`).
    """,
    BART_START_DOCSTRING,
)
class BartForQuestionAnswering(BartPretrainedModel):
    def __init__(self, config):
        super().__init__(config)

        config.num_labels = 2
        self.num_labels = config.num_labels

        self.model = BartModel(config)
        self.qa_outputs = nn.Linear(config.hidden_size, config.num_labels)

        self.model._init_weights(self.qa_outputs)

    @add_start_docstrings_to_model_forward(BART_INPUTS_DOCSTRING)
    @add_code_sample_docstrings(
        tokenizer_class=_TOKENIZER_FOR_DOC,
        checkpoint="facebook/bart-large",
        output_type=Seq2SeqQuestionAnsweringModelOutput,
        config_class=_CONFIG_FOR_DOC,
    )
    def forward(
        self,
        input_ids=None,
        attention_mask=None,
        decoder_input_ids=None,
        decoder_attention_mask=None,
        encoder_outputs=None,
        start_positions=None,
        end_positions=None,
        inputs_embeds=None,
        decoder_inputs_embeds=None,
        use_cache=None,
        output_attentions=None,
        output_hidden_states=None,
        return_dict=None,
    ):
        r"""
        start_positions (:obj:`torch.LongTensor` of shape :obj:`(batch_size,)`, `optional`):
            Labels for position (index) of the start of the labelled span for computing the token classification loss.
            Positions are clamped to the length of the sequence (`sequence_length`). Position outside of the sequence
            are not taken into account for computing the loss.
        end_positions (:obj:`torch.LongTensor` of shape :obj:`(batch_size,)`, `optional`):
            Labels for position (index) of the end of the labelled span for computing the token classification loss.
            Positions are clamped to the length of the sequence (`sequence_length`). Position outside of the sequence
            are not taken into account for computing the loss.
        """
        return_dict = return_dict if return_dict is not None else self.config.use_return_dict
        if start_positions is not None and end_positions is not None:
            use_cache = False

        outputs = self.model(
            input_ids,
            attention_mask=attention_mask,
            decoder_input_ids=decoder_input_ids,
            decoder_attention_mask=decoder_attention_mask,
            encoder_outputs=encoder_outputs,
            inputs_embeds=inputs_embeds,
            decoder_inputs_embeds=decoder_inputs_embeds,
            use_cache=use_cache,
            output_attentions=output_attentions,
            output_hidden_states=output_hidden_states,
            return_dict=return_dict,
        )

        sequence_output = outputs[0]

        logits = self.qa_outputs(sequence_output)
        start_logits, end_logits = logits.split(1, dim=-1)
        start_logits = start_logits.squeeze(-1)
        end_logits = end_logits.squeeze(-1)

        total_loss = None
        if start_positions is not None and end_positions is not None:
            # If we are on multi-GPU, split add a dimension
            if len(start_positions.size()) > 1:
                start_positions = start_positions.squeeze(-1)
            if len(end_positions.size()) > 1:
                end_positions = end_positions.squeeze(-1)
            # sometimes the start/end positions are outside our model inputs, we ignore these terms
            ignored_index = start_logits.size(1)
            start_positions.clamp_(0, ignored_index)
            end_positions.clamp_(0, ignored_index)

            loss_fct = CrossEntropyLoss(ignore_index=ignored_index)
            start_loss = loss_fct(start_logits, start_positions)
            end_loss = loss_fct(end_logits, end_positions)
            total_loss = (start_loss + end_loss) / 2

        if not return_dict:
            output = (
                start_logits,
                end_logits,
            ) + outputs[1:]
            return ((total_loss,) + output) if total_loss is not None else output

        return Seq2SeqQuestionAnsweringModelOutput(
            loss=total_loss,
            start_logits=start_logits,
            end_logits=end_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,
        )