modeling_bart.py

# 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.
""" BART model configuration """
import warnings
import logging
import pdb
import torch.nn.functional as F
import numpy as np
import wandb
from random import random
from transformers import PretrainedConfig

from modeling_static_spattn import build_static_sparsity_mask, apply_static_sparsity_mask
from modeling_sanger_spattn import gen_sparsity_mask, quant_qk_matmul, prune_attn_scores
from quant_utils import build_quant_matmul


logger = logging.getLogger(__name__)


BART_PRETRAINED_CONFIG_ARCHIVE_MAP = {
    "facebook/bart-large": "https://huggingface.co/facebook/bart-large/resolve/main/config.json",
    # See all BART models at https://huggingface.co/models?filter=bart
}

class BartConfig(PretrainedConfig):
    r"""
    This is the configuration class to store the configuration of a :class:`~transformers.BartModel`. It is used to
    instantiate a BART model according to the specified arguments, defining the model architecture. Instantiating a
    configuration with the defaults will yield a similar configuration to that of the BART `facebook/bart-large
    <https://huggingface.co/facebook/bart-large>`__ architecture.

    Configuration objects inherit from :class:`~transformers.PretrainedConfig` and can be used to control the model
    outputs. Read the documentation from :class:`~transformers.PretrainedConfig` for more information.


    Args:
        vocab_size (:obj:`int`, `optional`, defaults to 50265):
            Vocabulary size of the BART model. Defines the number of different tokens that can be represented by the
            :obj:`inputs_ids` passed when calling :class:`~transformers.BartModel` or
            :class:`~transformers.TFBartModel`.
        d_model (:obj:`int`, `optional`, defaults to 1024):
            Dimensionality of the layers and the pooler layer.
        encoder_layers (:obj:`int`, `optional`, defaults to 12):
            Number of encoder layers.
        decoder_layers (:obj:`int`, `optional`, defaults to 12):
            Number of decoder layers.
        encoder_attention_heads (:obj:`int`, `optional`, defaults to 16):
            Number of attention heads for each attention layer in the Transformer encoder.
        decoder_attention_heads (:obj:`int`, `optional`, defaults to 16):
            Number of attention heads for each attention layer in the Transformer decoder.
        decoder_ffn_dim (:obj:`int`, `optional`, defaults to 4096):
            Dimensionality of the "intermediate" (often named feed-forward) layer in decoder.
        encoder_ffn_dim (:obj:`int`, `optional`, defaults to 4096):
            Dimensionality of the "intermediate" (often named feed-forward) layer in decoder.
        activation_function (:obj:`str` or :obj:`function`, `optional`, defaults to :obj:`"gelu"`):
            The non-linear activation function (function or string) in the encoder and pooler. If string,
            :obj:`"gelu"`, :obj:`"relu"`, :obj:`"silu"` and :obj:`"gelu_new"` are supported.
        dropout (:obj:`float`, `optional`, defaults to 0.1):
            The dropout probability for all fully connected layers in the embeddings, encoder, and pooler.
        attention_dropout (:obj:`float`, `optional`, defaults to 0.0):
            The dropout ratio for the attention probabilities.
        activation_dropout (:obj:`float`, `optional`, defaults to 0.0):
            The dropout ratio for activations inside the fully connected layer.
        classifier_dropout (:obj:`float`, `optional`, defaults to 0.0):
            The dropout ratio for classifier.
        max_position_embeddings (:obj:`int`, `optional`, defaults to 1024):
            The maximum sequence length that this model might ever be used with. Typically set this to something large
            just in case (e.g., 512 or 1024 or 2048).
        init_std (:obj:`float`, `optional`, defaults to 0.02):
            The standard deviation of the truncated_normal_initializer for initializing all weight matrices.
        encoder_layerdrop: (:obj:`float`, `optional`, defaults to 0.0):
            The LayerDrop probability for the encoder. See the `LayerDrop paper <see
            https://arxiv.org/abs/1909.11556>`__ for more details.
        decoder_layerdrop: (:obj:`float`, `optional`, defaults to 0.0):
            The LayerDrop probability for the decoder. See the `LayerDrop paper <see
            https://arxiv.org/abs/1909.11556>`__ for more details.
        gradient_checkpointing (:obj:`bool`, `optional`, defaults to :obj:`False`):
            If True, use gradient checkpointing to save memory at the expense of slower backward pass.
        scale_embedding (:obj:`bool`, `optional`, defaults to :obj:`False`):
            Scale embeddings by diving by sqrt(d_model).
        use_cache (:obj:`bool`, `optional`, defaults to :obj:`True`):
            Whether or not the model should return the last key/values attentions (not used by all models).
        num_labels: (:obj:`int`, `optional`, defaults to 3):
            The number of labels to use in :class:`~transformers.BartForSequenceClassification`.
        forced_eos_token_id (:obj:`int`, `optional`, defaults to 2):
            The id of the token to force as the last generated token when :obj:`max_length` is reached. Usually set to
            :obj:`eos_token_id`.

    Example::

        >>> from transformers import BartModel, BartConfig

        >>> # Initializing a BART facebook/bart-large style configuration
        >>> configuration = BartConfig()

        >>> # Initializing a model from the facebook/bart-large style configuration
        >>> model = BartModel(configuration)

        >>> # Accessing the model configuration
        >>> configuration = model.config
    """
    model_type = "bart"
    keys_to_ignore_at_inference = ["past_key_values"]

    def __init__(
        self,
        vocab_size=50265,
        max_position_embeddings=1024,
        encoder_layers=12,
        encoder_ffn_dim=4096,
        encoder_attention_heads=16,
        decoder_layers=12,
        decoder_ffn_dim=4096,
        decoder_attention_heads=16,
        encoder_layerdrop=0.0,
        decoder_layerdrop=0.0,
        activation_function="gelu",
        d_model=1024,
        dropout=0.1,
        attention_dropout=0.0,
        activation_dropout=0.0,
        init_std=0.02,
        classifier_dropout=0.0,
        scale_embedding=False,
        gradient_checkpointing=False,
        use_cache=True,
        num_labels=3,
        pad_token_id=1,
        bos_token_id=0,
        eos_token_id=2,
        is_encoder_decoder=True,
        decoder_start_token_id=2,
        forced_eos_token_id=2,
        **kwargs
    ):

        super().__init__(
            num_labels=num_labels,
            pad_token_id=pad_token_id,
            bos_token_id=bos_token_id,
            eos_token_id=eos_token_id,
            is_encoder_decoder=is_encoder_decoder,
            decoder_start_token_id=decoder_start_token_id,
            forced_eos_token_id=forced_eos_token_id,
            **kwargs,
        )


        self.vocab_size = vocab_size
        self.max_position_embeddings = max_position_embeddings
        self.d_model = d_model
        self.encoder_ffn_dim = encoder_ffn_dim
        self.encoder_layers = encoder_layers
        self.encoder_attention_heads = encoder_attention_heads
        self.decoder_ffn_dim = decoder_ffn_dim
        self.decoder_layers = decoder_layers
        self.decoder_attention_heads = decoder_attention_heads
        self.dropout = dropout
        self.attention_dropout = attention_dropout
        self.activation_dropout = activation_dropout
        self.activation_function = activation_function
        self.init_std = init_std
        self.encoder_layerdrop = encoder_layerdrop
        self.decoder_layerdrop = decoder_layerdrop
        self.classifier_dropout = classifier_dropout
        self.use_cache = use_cache
        self.num_hidden_layers = encoder_layers
        self.gradient_checkpointing = gradient_checkpointing
        self.scale_embedding = scale_embedding  # scale factor will be sqrt(d_model) if True

        # ensure backward compatibility for BART CNN models
        if self.forced_bos_token_id is None and kwargs.get("force_bos_token_to_be_generated", False):
            self.forced_bos_token_id = self.bos_token_id
            warnings.warn(
                f"Please make sure the config includes `forced_bos_token_id={self.bos_token_id}` in future versions."
                "The config can simply be saved and uploaded again to be fixed."
            )

    @property
    def num_attention_heads(self) -> int:
        return self.encoder_attention_heads

    @property
    def hidden_size(self) -> int:
        return self.d_model

# 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.

from transformers.models.roberta import RobertaTokenizer

VOCAB_FILES_NAMES = {"vocab_file": "vocab.json", "merges_file": "merges.txt"}

# See all BART models at https://huggingface.co/models?filter=bart
PRETRAINED_VOCAB_FILES_MAP = {
    "vocab_file": {
        "facebook/bart-base": "https://huggingface.co/facebook/bart-base/resolve/main/vocab.json",
        "facebook/bart-large": "https://huggingface.co/facebook/bart-large/resolve/main/vocab.json",
        "facebook/bart-large-mnli": "https://huggingface.co/facebook/bart-large-mnli/resolve/main/vocab.json",
        "facebook/bart-large-cnn": "https://huggingface.co/facebook/bart-large-cnn/resolve/main/vocab.json",
        "facebook/bart-large-xsum": "https://huggingface.co/facebook/bart-large-xsum/resolve/main/vocab.json",
        "yjernite/bart_eli5": "https://huggingface.co/yjernite/bart_eli5/resolve/main/vocab.json",
    },
    "merges_file": {
        "facebook/bart-base": "https://huggingface.co/facebook/bart-base/resolve/main/merges.txt",
        "facebook/bart-large": "https://huggingface.co/facebook/bart-large/resolve/main/merges.txt",
        "facebook/bart-large-mnli": "https://huggingface.co/facebook/bart-large-mnli/resolve/main/merges.txt",
        "facebook/bart-large-cnn": "https://huggingface.co/facebook/bart-large-cnn/resolve/main/merges.txt",
        "facebook/bart-large-xsum": "https://huggingface.co/facebook/bart-large-xsum/resolve/main/merges.txt",
        "yjernite/bart_eli5": "https://huggingface.co/yjernite/bart_eli5/resolve/main/merges.txt",
    },
}

PRETRAINED_POSITIONAL_EMBEDDINGS_SIZES = {
    "facebook/bart-base": 1024,
    "facebook/bart-large": 1024,
    "facebook/bart-large-mnli": 1024,
    "facebook/bart-large-cnn": 1024,
    "facebook/bart-large-xsum": 1024,
    "yjernite/bart_eli5": 1024,
}


class BartTokenizer(RobertaTokenizer):
    r"""
    Construct a BART tokenizer.

    :class:`~transformers.BartTokenizer` is identical to :class:`~transformers.RobertaTokenizer`. Refer to superclass
    :class:`~transformers.RobertaTokenizer` for usage examples and documentation concerning the initialization
    parameters and other methods.
    """
    vocab_files_names = VOCAB_FILES_NAMES
    pretrained_vocab_files_map = PRETRAINED_VOCAB_FILES_MAP
    max_model_input_sizes = PRETRAINED_POSITIONAL_EMBEDDINGS_SIZES

# 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 BART model. """
import copy
import math
import random
import warnings
from typing import Optional, Tuple

import torch
import torch.utils.checkpoint
from torch import nn
from torch.nn import CrossEntropyLoss, MSELoss

from transformers.activations import ACT2FN
from transformers.file_utils import (
    add_code_sample_docstrings,
    add_end_docstrings,
    add_start_docstrings,
    add_start_docstrings_to_model_forward,
    replace_return_docstrings,
)
from transformers.modeling_outputs import (
    BaseModelOutput,
    BaseModelOutputWithPastAndCrossAttentions,
    CausalLMOutputWithCrossAttentions,
    Seq2SeqLMOutput,
    Seq2SeqModelOutput,
    Seq2SeqQuestionAnsweringModelOutput,
    Seq2SeqSequenceClassifierOutput,
)
from transformers.modeling_utils import PreTrainedModel


_CHECKPOINT_FOR_DOC = "facebook/bart-large"
_CONFIG_FOR_DOC = "BartConfig"
_TOKENIZER_FOR_DOC = "BartTokenizer"


BART_PRETRAINED_MODEL_ARCHIVE_LIST = [
    "facebook/bart-large",
    # See all BART models at https://huggingface.co/models?filter=bart
]


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

    assert pad_token_id is not None, "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


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):
    """
    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)

    inverted_mask = 1.0 - expanded_mask

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


class BartLearnedPositionalEmbedding(nn.Embedding):
    """
    This module learns positional embeddings up to a fixed maximum size.
    """

    def __init__(self, num_embeddings: int, embedding_dim: 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 don't have this hack
        self.offset = 2
        super().__init__(num_embeddings + self.offset, embedding_dim)

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

    def __init__(
        self,
        config,
        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)
        
        do_quant = getattr(config, 'quant_qk', False)
        if do_quant:
            self.quant_matmul = build_quant_matmul(config)

        assert config.encoder_attention_heads == config.decoder_attention_heads
        num_attention_heads = config.encoder_attention_heads
        if getattr(config, 'static_sparsity', False):
            sparsity_mask = build_static_sparsity_mask(
                json_file=config.static_sparsity, 
                num_heads=num_attention_heads, 
                max_position=config.max_position_embeddings,
            )
            self.register_buffer('static_sparsity_mask', sparsity_mask)
        
        self.config = config
            
    def _shape(self, tensor: torch.Tensor, seq_len: int, bsz: int):
        return tensor.view(bsz, seq_len, self.num_heads, self.head_dim).transpose(1, 2).contiguous()

    def forward(
        self,
        hidden_states: torch.Tensor,
        key_value_states: Optional[torch.Tensor] = None,
        past_key_value: Optional[Tuple[torch.Tensor]] = None,
        attention_mask: Optional[torch.Tensor] = None,
        layer_head_mask: Optional[torch.Tensor] = None,
        output_attentions: bool = False,
    ) -> Tuple[torch.Tensor, Optional[torch.Tensor], Optional[Tuple[torch.Tensor]]]:
        """Input shape: Batch x Time x Channel"""

        # if key_value_states are provided this layer is used as a cross-attention layer
        # for the decoder
        is_cross_attention = key_value_states is not None
        bsz, tgt_len, embed_dim = hidden_states.size()
        src_len = tgt_len

        # 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)
        # both [b*n_head, l, d]

        """ sparse attention """
        key_states = key_states.transpose(1, 2)

        if attention_mask is None:
            attention_mask = key_states.new_zeros(bsz, 1, tgt_len, src_len)

        do_quant = getattr(self.config, 'quant_qk', False)
        do_prune = getattr(self.config, 'prune_score', False)

        attn_weights = torch.matmul(query_states, key_states)
        attention_mask = attention_mask.expand(-1, self.num_heads, -1, -1)
        attention_mask = attention_mask.reshape_as(attn_weights)

        if do_quant:
            quant_attn_weights = quant_qk_matmul(query_states, key_states, self.config, self.quant_matmul)
        else:
            quant_attn_weights = None

        if do_prune:
            attn_scores = quant_attn_weights if do_quant else attn_weights
            sparsity_mask = prune_attn_scores(attn_scores, attention_mask, self.config)
            attn_weights += sparsity_mask

        # if hasattr(self, 'quant_matmul'):
        #     with torch.no_grad():
        #         if attention_mask is None:
        #             attention_mask = 0
        #         attention_mask = attention_mask + torch.zeros((bsz, self.num_heads, tgt_len, src_len), device = key_states.device)

        #         with torch.no_grad():
        #             query_norm = query_states.norm(dim = 2, keepdim=True)
        #             key_norm = key_states.norm(dim = 1, keepdim=True)
        #             normalized_query_states = query_states/query_norm
        #             normalized_key_states = key_states/key_norm

        #         quant_attention_scores = self.quant_matmul(normalized_query_states, normalized_key_states)
        #         quant_attention_scores *= query_norm*key_norm

        #         quant_attention_scores = quant_attention_scores.view(*attention_mask.shape)
        #         sparsity_mask = prune_attn_scores(quant_attention_scores, attention_mask, self.config)
                
        # attention_mask =  attention_mask + sparsity_mask
        
        # src_len = key_states.size(2)
        # attn_weights = torch.bmm(query_states, key_states)

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

        do_static_sparsity = getattr(self.config, 'static_sparsity', False)
        if do_static_sparsity:
            attn_weights = attn_weights.view(bsz, self.num_heads, tgt_len, src_len)
            apply_static_sparsity_mask(self.static_sparsity_mask, attn_weights)
        
        attn_weights += attention_mask

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

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

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

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

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

        attn_output = attn_output.view(bsz, self.num_heads, tgt_len, self.head_dim)
        attn_output = attn_output.transpose(1, 2)
        attn_output = attn_output.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(
            config,
            embed_dim=self.embed_dim,
            num_heads=config.encoder_attention_heads,
            dropout=config.attention_dropout,
        )
        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,
    ):
        """
        Args:
            hidden_states (:obj:`torch.FloatTensor`): input to the layer of shape `(seq_len, batch, embed_dim)`
            attention_mask (:obj:`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 (:obj:`torch.FloatTensor`): mask for attention heads in a given layer of size
                `(encoder_attention_heads,)`.
            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.
        """
        residual = 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
        hidden_states = self.self_attn_layer_norm(hidden_states)

        residual = 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
        hidden_states = self.final_layer_norm(hidden_states)

        if hidden_states.dtype == torch.float16 and (
            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)

        outputs = (hidden_states,)

        if output_attentions:
            outputs += (attn_weights,)

        return outputs


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

        self.self_attn = BartAttention(
            config,
            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.self_attn_layer_norm = nn.LayerNorm(self.embed_dim)
        self.encoder_attn = BartAttention(
            config,
            self.embed_dim,
            config.decoder_attention_heads,
            dropout=config.attention_dropout,
            is_decoder=True,
        )
        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)

    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[Tuple[torch.Tensor]] = None,
        output_attentions: Optional[bool] = False,
        use_cache: Optional[bool] = True,
    ):
        """
        Args:
            hidden_states (:obj:`torch.FloatTensor`): input to the layer of shape `(batch, seq_len, embed_dim)`
            attention_mask (:obj:`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 (:obj:`torch.FloatTensor`): cross attention input to the layer of shape `(batch, seq_len, embed_dim)`
            encoder_attention_mask (:obj:`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 (:obj:`torch.FloatTensor`): mask for attention heads in a given layer of size
                `(encoder_attention_heads,)`.
            cross_attn_layer_head_mask (:obj:`torch.FloatTensor`): mask for cross-attention heads in a given layer of
                size `(decoder_attention_heads,)`.
            past_key_value (:obj:`Tuple(torch.FloatTensor)`): cached past key and value projection states
            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.
        """
        residual = hidden_states

        # Self Attention
        # decoder uni-directional self-attention cached key/values tuple is at positions 1,2
        self_attn_past_key_value = past_key_value[:2] if past_key_value is not None else None
        # add present self-attn cache to positions 1,2 of present_key_value tuple
        hidden_states, self_attn_weights, present_key_value = self.self_attn(
            hidden_states=hidden_states,
            past_key_value=self_attn_past_key_value,
            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
        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

            # cross_attn cached key/values tuple is at positions 3,4 of present_key_value tuple
            cross_attn_past_key_value = past_key_value[-2:] 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,
                attention_mask=encoder_attention_mask,
                layer_head_mask=cross_attn_layer_head_mask,
                past_key_value=cross_attn_past_key_value,
                output_attentions=output_attentions,
            )
            hidden_states = nn.functional.dropout(hidden_states, p=self.dropout, training=self.training)
            hidden_states = residual + hidden_states
            hidden_states = self.encoder_attn_layer_norm(hidden_states)

            # add cross-attn to positions 3,4 of present_key_value tuple
            present_key_value = present_key_value + cross_attn_present_key_value

        # Fully Connected
        residual = 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
        hidden_states = self.final_layer_norm(hidden_states)

        outputs = (hidden_states,)

        if output_attentions:
            outputs += (self_attn_weights, cross_attn_weights)

        if use_cache:
            outputs += (present_key_value,)

        return outputs


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"
    _keys_to_ignore_on_load_unexpected = [r"encoder\.version", r"decoder\.version"]

    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, 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

        >>> 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])

    Mask filling example::

        >>> 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()
"""

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`):
            Indices of decoder input sequence tokens in the vocabulary.

            Indices can be obtained using :class:`~transformers.BartTokenizer`. See
            :meth:`transformers.PreTrainedTokenizer.encode` and :meth:`transformers.PreTrainedTokenizer.__call__` for
            details.

            `What are decoder input IDs? <../glossary.html#decoder-input-ids>`__

            Bart uses the :obj:`eos_token_id` as the starting token for :obj:`decoder_input_ids` generation. If
            :obj:`past_key_values` is used, optionally only the last :obj:`decoder_input_ids` have to be input (see
            :obj:`past_key_values`).

            For translation and summarization training, :obj:`decoder_input_ids` should be provided. If no
            :obj:`decoder_input_ids` is provided, the model will create this tensor by shifting the :obj:`input_ids` to
            the right for denoising pre-training following the paper.
        decoder_attention_mask (:obj:`torch.LongTensor` of shape :obj:`(batch_size, target_sequence_length)`, `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.
        head_mask (:obj:`torch.Tensor` of shape :obj:`(encoder_layers, encoder_attention_heads)`, `optional`):
            Mask to nullify selected heads of the attention modules in the encoder. Mask values selected in ``[0, 1]``:

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

        decoder_head_mask (:obj:`torch.Tensor` of shape :obj:`(decoder_layers, decoder_attention_heads)`, `optional`):
            Mask to nullify selected heads of the attention modules in the decoder. Mask values selected in ``[0, 1]``:

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

        cross_attn_head_mask (:obj:`torch.Tensor` of shape :obj:`(decoder_layers, decoder_attention_heads)`, `optional`):
            Mask to nullify selected heads of the cross-attention modules in the decoder. Mask values selected in ``[0,
            1]``:

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

        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(torch.FloatTensor))`, `optional`, returned when ``use_cache=True`` is passed or when ``config.use_cache=True``):
            Tuple of :obj:`tuple(torch.FloatTensor)` of length :obj:`config.n_layers`, with each tuple having 2 tensors
            of shape :obj:`(batch_size, num_heads, sequence_length, embed_size_per_head)`) and 2 additional tensors of
            shape :obj:`(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 :obj:`past_key_values` input) to speed up sequential 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 (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.padding_idx = config.pad_token_id
        self.max_source_positions = config.max_position_embeddings
        self.embed_scale = math.sqrt(embed_dim) 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, embed_dim, self.padding_idx)

        self.embed_positions = BartLearnedPositionalEmbedding(
            config.max_position_embeddings,
            embed_dim,
        )
        self.layers = nn.ModuleList([BartEncoderLayer(config) for _ in range(config.encoder_layers)])
        self.layernorm_embedding = nn.LayerNorm(embed_dim)

        self.init_weights()

    def forward(
        self,
        input_ids=None,
        attention_mask=None,
        head_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>`__
            head_mask (:obj:`torch.Tensor` of shape :obj:`(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**.

            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 = nn.functional.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

        # 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://arxiv.org/abs/1909.11556 for description)
            dropout_probability = random.uniform(0, 1)
            if self.training and (dropout_probability < self.layerdrop):  # skip the layer
                layer_outputs = (None, None)
            else:
                if getattr(self.config, "gradient_checkpointing", False) and self.training:

                    def create_custom_forward(module):
                        def custom_forward(*inputs):
                            return module(*inputs, output_attentions)

                        return custom_forward

                    layer_outputs = torch.utils.checkpoint.checkpoint(
                        create_custom_forward(encoder_layer),
                        hidden_states,
                        attention_mask,
                        (head_mask[idx] if head_mask is not None else 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],)

        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 (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.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)

        self.embed_positions = BartLearnedPositionalEmbedding(
            config.max_position_embeddings,
            config.d_model,
        )
        self.layers = nn.ModuleList([BartDecoderLayer(config) for _ in range(config.decoder_layers)])
        self.layernorm_embedding = nn.LayerNorm(config.d_model)

        self.init_weights()

    def get_input_embeddings(self):
        return self.embed_tokens

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

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

        if attention_mask is not None:
            # [bsz, seq_len] -> [bsz, 1, tgt_seq_len, src_seq_len]
            expanded_attn_mask = _expand_mask(attention_mask, inputs_embeds.dtype, tgt_len=input_shape[-1])
            combined_attention_mask = (
                expanded_attn_mask if combined_attention_mask is None else expanded_attn_mask + combined_attention_mask
            )

        return combined_attention_mask

    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,
    ):
        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>`__
            head_mask (:obj:`torch.Tensor` of shape :obj:`(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 (:obj:`torch.Tensor` of shape :obj:`(decoder_layers, decoder_attention_heads)`, `optional`):
                Mask to nullify selected heads of the cross-attention modules in the decoder 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 (:obj:`tuple(tuple(torch.FloatTensor))`, `optional`, returned when ``use_cache=True`` is passed or when ``config.use_cache=True``):
                Tuple of :obj:`tuple(torch.FloatTensor)` of length :obj:`config.n_layers`, with each tuple having 2
                tensors of shape :obj:`(batch_size, num_heads, sequence_length, embed_size_per_head)`) and 2 additional
                tensors of shape :obj:`(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 :obj:`past_key_values` input) to speed up sequential
                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].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

        attention_mask = self._prepare_decoder_attention_mask(
            attention_mask, input_shape, inputs_embeds, 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)

        hidden_states = inputs_embeds + positions
        hidden_states = self.layernorm_embedding(hidden_states)

        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
        next_decoder_cache = () if use_cache 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 {head_mask.size()[0]}."
        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

            if getattr(self.config, "gradient_checkpointing", False) and self.training:

                if use_cache:
                    logger.warning(
                        "`use_cache=True` is incompatible with `config.gradient_checkpointing=True`. Setting "
                        "`use_cache=False`..."
                    )
                    use_cache = False

                def create_custom_forward(module):
                    def custom_forward(*inputs):
                        # None for past_key_value
                        return module(*inputs, output_attentions, use_cache)

                    return custom_forward

                layer_outputs = torch.utils.checkpoint.checkpoint(
                    create_custom_forward(decoder_layer),
                    hidden_states,
                    attention_mask,
                    encoder_hidden_states,
                    encoder_attention_mask,
                    head_mask[idx] if head_mask is not None else None,
                    cross_attn_head_mask[idx] if cross_attn_head_mask is not None else None,
                    None,
                )
            else:

                layer_outputs = decoder_layer(
                    hidden_states,
                    attention_mask=attention_mask,
                    encoder_hidden_states=encoder_hidden_states,
                    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_value,
                    output_attentions=output_attentions,
                    use_cache=use_cache,
                )
            hidden_states = layer_outputs[0]

            if use_cache:
                next_decoder_cache += (layer_outputs[3 if output_attentions else 1],)

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

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

        # add hidden states from the last decoder layer
        if output_hidden_states:
            all_hidden_states += (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=_CHECKPOINT_FOR_DOC,
        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,
        head_mask=None,
        decoder_head_mask=None,
        cross_attn_head_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,
    ):

        # different to other models, Bart automatically creates decoder_input_ids from
        # input_ids if no decoder_input_ids are provided
        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, self.config.decoder_start_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,
                head_mask=head_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=True
        elif return_dict and not isinstance(encoder_outputs, BaseModelOutput):
            encoder_outputs = BaseModelOutput(
                last_hidden_state=encoder_outputs[0],
                hidden_states=encoder_outputs[1] if len(encoder_outputs) > 1 else None,
                attentions=encoder_outputs[2] if len(encoder_outputs) > 2 else None,
            )

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

        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"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,
        head_mask=None,
        decoder_head_mask=None,
        cross_attn_head_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:
        """
        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:
                decoder_input_ids = shift_tokens_right(
                    labels, self.config.pad_token_id, self.config.decoder_start_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,
            head_mask=head_mask,
            decoder_head_mask=decoder_head_mask,
            cross_attn_head_mask=cross_attn_head_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()
            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,
        head_mask=None,
        decoder_head_mask=None,
        cross_attn_head_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,
            "head_mask": head_mask,
            "decoder_head_mask": decoder_head_mask,
            "cross_attn_head_mask": cross_attn_head_mask,
            "use_cache": use_cache,  # change this to avoid caching (presumably for debugging)
        }

    def prepare_decoder_input_ids_from_labels(self, labels: torch.Tensor):
        return shift_tokens_right(labels, self.config.pad_token_id, self.config.decoder_start_token_id)

    @staticmethod
    def _reorder_cache(past, beam_idx):
        reordered_past = ()
        for layer_past in past:
            # cached cross_attention states don't have to be reordered -> they are always the same
            reordered_past += (
                tuple(past_state.index_select(0, beam_idx) for past_state in layer_past[:2]) + layer_past[2:],
            )
        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=_CHECKPOINT_FOR_DOC,
        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,
        head_mask=None,
        decoder_head_mask=None,
        cross_attn_head_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,
            head_mask=head_mask,
            decoder_head_mask=decoder_head_mask,
            cross_attn_head_mask=cross_attn_head_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:
            if self.config.num_labels == 1:
                # regression
                loss_fct = MSELoss()
                loss = loss_fct(logits.view(-1), labels.view(-1))
            else:
                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=_CHECKPOINT_FOR_DOC,
        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,
        head_mask=None,
        decoder_head_mask=None,
        cross_attn_head_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,
            head_mask=head_mask,
            decoder_head_mask=decoder_head_mask,
            cross_attn_head_mask=cross_attn_head_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).contiguous()
        end_logits = end_logits.squeeze(-1).contiguous()

        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 = start_positions.clamp(0, ignored_index)
            end_positions = 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,
        )



class BartDecoderWrapper(BartPretrainedModel):
    """
    This wrapper class is a helper class to correctly load pretrained checkpoints when the causal language model is
    used in combination with the :class:`~transformers.EncoderDecoderModel` framework.
    """

    def __init__(self, config):
        super().__init__(config)
        self.decoder = BartDecoder(config)

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




class BartForCausalLM(BartPretrainedModel):
    def __init__(self, config):
        super().__init__(config)
        config = copy.deepcopy(config)
        config.is_decoder = True
        config.is_encoder_decoder = False
        self.model = BartDecoderWrapper(config)

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

        self.init_weights()

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

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

    def get_output_embeddings(self):
        return self.lm_head

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

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

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



    @replace_return_docstrings(output_type=CausalLMOutputWithCrossAttentions, config_class=_CONFIG_FOR_DOC)
    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,
        labels=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, sequence_length, hidden_size)`, `optional`):
                Sequence of hidden-states at the output of the last layer of the encoder. Used in the cross-attention
                if the model is configured as a decoder.
            encoder_attention_mask (:obj:`torch.FloatTensor` of shape :obj:`(batch_size, sequence_length)`, `optional`):
                Mask to avoid performing attention on the padding token indices of the encoder input. This mask is used
                in the cross-attention if the model is configured as a decoder. Mask values selected in ``[0, 1]``:
            head_mask (:obj:`torch.Tensor` of shape :obj:`(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 (:obj:`torch.Tensor` of shape :obj:`(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**.

            past_key_values (:obj:`tuple(tuple(torch.FloatTensor))`, `optional`, returned when ``use_cache=True`` is passed or when ``config.use_cache=True``):
                Tuple of :obj:`tuple(torch.FloatTensor)` of length :obj:`config.n_layers`, with each tuple having 2
                tensors of shape :obj:`(batch_size, num_heads, sequence_length, embed_size_per_head)`) and 2 additional
                tensors of shape :obj:`(batch_size, num_heads, encoder_sequence_length, embed_size_per_head)`. The two
                additional tensors are only required when the model is used as a decoder in a Sequence to Sequence
                model.

                Contains pre-computed hidden-states (key and values in the self-attention blocks and in the
                cross-attention blocks) that can be used (see :obj:`past_key_values` input) to speed up sequential
                decoding.

                If :obj:`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 :obj:`(batch_size, 1)`
                instead of all ``decoder_input_ids`` of shape :obj:`(batch_size, sequence_length)`.
            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]``.
            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`).

                - 1 for tokens that are **not masked**,
                - 0 for tokens that are **masked**.
            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.

        Returns:

        Example::

            >>> from transformers import BartTokenizer, BartForCausalLM

            >>> tokenizer = BartTokenizer.from_pretrained('facebook/bart-large')
            >>> model = BartForCausalLM.from_pretrained('facebook/bart-large', add_cross_attention=False)
            >>> assert model.config.is_decoder, f"{model.__class__} has to be configured as a decoder."
            >>> inputs = tokenizer("Hello, my dog is cute", return_tensors="pt")
            >>> outputs = model(**inputs)

            >>> last_hidden_states = outputs.last_hidden_state
        """

        output_attentions = output_attentions if output_attentions is not None else self.config.output_attentions
        output_hidden_states = (
            output_hidden_states if output_hidden_states is not None else self.config.output_hidden_states
        )
        return_dict = return_dict if return_dict is not None else self.config.use_return_dict

        # decoder outputs consists of (dec_features, layer_state, dec_hidden, dec_attn)
        outputs = self.model.decoder(
            input_ids=input_ids,
            attention_mask=attention_mask,
            encoder_hidden_states=encoder_hidden_states,
            encoder_attention_mask=encoder_attention_mask,
            head_mask=head_mask,
            cross_attn_head_mask=cross_attn_head_mask,
            past_key_values=past_key_values,
            inputs_embeds=inputs_embeds,
            use_cache=use_cache,
            output_attentions=output_attentions,
            output_hidden_states=output_hidden_states,
            return_dict=return_dict,
        )

        logits = self.lm_head(outputs[0])

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

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

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


    def prepare_inputs_for_generation(self, input_ids, past=None, attention_mask=None, use_cache=None, **kwargs):
        # if model is used as a decoder in encoder-decoder model, the decoder attention mask is created on the fly
        if attention_mask is None:
            attention_mask = input_ids.new_ones(input_ids.shape)

        if past:
            input_ids = input_ids[:, -1:]
        # first step, decoder_cached_states are empty
        return {
            "input_ids": input_ids,  # encoder_outputs is defined. input_ids not needed
            "attention_mask": attention_mask,
            "past_key_values": past,
            "use_cache": use_cache,
        }

    @staticmethod
    def _reorder_cache(past, beam_idx):
        reordered_past = ()
        for layer_past in past:
            reordered_past += (tuple(past_state.index_select(0, beam_idx) for past_state in layer_past),)
        return reordered_past