modeling_bart.py

# coding=utf-8
# Copyright 2020 The Facebook AI Research Team Authors and The HuggingFace Inc. team.
#
# Licensed under the Apache License, Version 2.0 (the "License");
# you may not use this file except in compliance with the License.
# You may obtain a copy of the License at
#
#     http://www.apache.org/licenses/LICENSE-2.0
#
# Unless required by applicable law or agreed to in writing, software
# distributed under the License is distributed on an "AS IS" BASIS,
# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
# See the License for the specific language governing permissions and
# limitations under the License.
"""PyTorch BART model, ported from the fairseq repo."""
import logging
import math
import random
from typing import Dict, List, Optional, Tuple

import torch
import torch.nn.functional as F
from torch import Tensor, nn

from .configuration_bart import BartConfig
from .file_utils import add_start_docstrings, add_start_docstrings_to_callable
from .modeling_utils import BeamHypotheses, PreTrainedModel, create_position_ids_from_input_ids


logger = logging.getLogger(__name__)


BART_PRETRAINED_MODEL_ARCHIVE_MAP = {
    "bart-large": "https://s3.amazonaws.com/models.huggingface.co/bert/facebook/bart-large/pytorch_model.bin",
    "bart-large-mnli": "https://s3.amazonaws.com/models.huggingface.co/bert/facebook/bart-large-mnli/pytorch_model.bin",
    "bart-large-cnn": "https://s3.amazonaws.com/models.huggingface.co/bert/facebook/bart-large-cnn/pytorch_model.bin",
}

BART_START_DOCSTRING = r"""

    This model is a PyTorch `torch.nn.Module <https://pytorch.org/docs/stable/nn.html#torch.nn.Module>`_ sub-class. Use it as a regular PyTorch Module and
    refer to the PyTorch documentation for all matters 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_INPUTS_DOCSTRING = r"""
    Args:
        input_ids (:obj:`torch.LongTensor` of shape :obj:`(batch_size, sequence_length)`):
               Indices of input sequence tokens in the vocabulary. Use BartTokenizer.encode to produce them.
            Padding will be ignored by default should you provide it.
            Indices can be obtained using :class:`transformers.BartTokenizer.encode(text)`.
        attention_mask (:obj:`torch.Tensor` of shape :obj:`(batch_size, sequence_length)`, `optional`, defaults to :obj:`None`):
            Mask to avoid performing attention on padding token indices in input_ids.
            Mask values selected in ``[0, 1]``:
            ``1`` for tokens that are NOT MASKED, ``0`` for MASKED tokens.
        decoder_input_ids (:obj:`torch.LongTensor` of shape :obj:`(batch_size, target_sequence_length)`, `optional`, defaults to :obj:`None`):
            Provide for translation and summarization training. By default, the model will create this tensor by shifting the input_ids right, following the paper.
        decoder_attention_mask (:obj:`torch.Tensor` of shape :obj:`(batch_size, 1, tgt_seq_len, tgt_seq_len)`, `optional`, defaults to :obj:`None`):
            Default behavior: generate a tensor that ignores pad tokens and future tokens, as in the paper.
            If you want to change padding behavior, you should read :func:`~transformers.modeling_bart._prepare_decoder_inputs` and modify.
            See diagram 1 in the paper for more info on the default strategy
"""
LARGE_NEGATIVE = -1e4


def _prepare_bart_decoder_inputs(
    config, input_ids, decoder_input_ids=None, decoder_attn_mask=None,
):
    """Prepare masks that ignore padding tokens  decoder and a causal lm mask for the decoder if
    none are provided. This mimics the default behavior in fairseq. To override it pass in masks.
    """
    pad_token_id = config.pad_token_id
    need_causal_mask = not config.output_past
    if decoder_input_ids is None:
        decoder_input_ids = shift_tokens_right(input_ids, pad_token_id)
    bsz, tgt_len = decoder_input_ids.size()[:2]
    if decoder_attn_mask is None:
        decoder_padding_mask = make_padding_mask(decoder_input_ids, pad_token_id)
        if need_causal_mask:
            causal_lm_mask = torch.triu(fill_with_neg_inf(torch.zeros(tgt_len, tgt_len)), 1)
        else:
            causal_lm_mask = None
        new_shape = (bsz, tgt_len, tgt_len)
        # make it broadcastable so can just be added to the attention coefficients
        decoder_attn_mask = _combine_masks(decoder_padding_mask, causal_lm_mask, new_shape).to(device=input_ids.device)
    assert decoder_attn_mask is None or decoder_attn_mask.shape == (bsz, 1, tgt_len, tgt_len)
    return decoder_input_ids, decoder_attn_mask


class PretrainedBartModel(PreTrainedModel):
    config_class = BartConfig
    base_model_prefix = "model"
    pretrained_model_archive_map = BART_PRETRAINED_MODEL_ARCHIVE_MAP

    def _init_weights(self, module):
        std = self.config.init_std

        # called init_bert_params in fairseq
        if isinstance(module, nn.Linear):
            module.weight.data.normal_(mean=0.0, std=std)
            if module.bias is not None:
                module.bias.data.zero_()
        if 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 = 1
        input_ids = torch.Tensor(
            [
                [0, 31414, 232, 328, 740, 1140, 12695, 69, 46078, 1588, 2],
                [0, 31414, 232, 328, 740, 1140, 12695, 69, 46078, 2, pad_token],
            ]
        ).long()
        decoder_input_ids, decoder_attn_mask = _prepare_bart_decoder_inputs(
            self.config, input_ids, attention_mask=None, decoder_input_ids=None, decoder_attn_mask=None
        )
        dummy_inputs = {
            "decoder_input_ids": decoder_input_ids,
            "attention_mask": input_ids.ne(pad_token),
            "input_ids": input_ids,
            "decoder_attention_mask": decoder_attn_mask,
        }
        return dummy_inputs


def _make_linear_from_emb(emb):
    vocab_size, emb_size = emb.weight.shape
    lin_layer = nn.Linear(vocab_size, emb_size, bias=False)
    lin_layer.weight.data = emb.weight.data  # .T
    return lin_layer


# Helper Functions, mostly for making masks
def _check_shapes(shape_1, shape2):
    if shape_1 != shape2:
        raise AssertionError("shape mismatch: {} != {}".format(shape_1, shape2))


def _combine_masks(key_padding_mask, attn_mask, targ_size):
    # targ_size = (bsz, tgt_len, src_len)
    a = torch.zeros(targ_size)
    b = torch.zeros(targ_size)
    if key_padding_mask is not None:  # (bsz, tgt_len) -> targ_size
        _check_shapes(key_padding_mask.shape, targ_size[:2])
        reshaped = key_padding_mask.unsqueeze(2).expand(*targ_size)
        a[reshaped] = 1e-8

    if attn_mask is not None:  # (tgt_len, src_len) -> targ_size
        _check_shapes(attn_mask.shape, targ_size[-2:])
        b = attn_mask.unsqueeze(0).expand(*targ_size)
    return (a + b).unsqueeze(1).clamp(LARGE_NEGATIVE,)


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


def make_padding_mask(input_ids, padding_idx=1):
    """True for pad tokens"""
    padding_mask = input_ids.eq(padding_idx)
    if not padding_mask.any():
        padding_mask = None
    return padding_mask


# Helper Modules


class EncoderLayer(nn.Module):
    def __init__(self, config: BartConfig):
        super().__init__()
        self.embed_dim = config.d_model
        self.output_attentions = config.output_attentions
        self.self_attn = SelfAttention(
            self.embed_dim, config.encoder_attention_heads, dropout=config.attention_dropout,
        )
        self.self_attn_layer_norm = LayerNorm(self.embed_dim)
        self.dropout = config.dropout
        self.activation_fn = F.gelu
        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 = LayerNorm(self.embed_dim)

    def forward(self, x, encoder_padding_mask):
        """
        Args:
            x (Tensor): input to the layer of shape `(seq_len, batch, embed_dim)`
            encoder_padding_mask (ByteTensor): binary ByteTensor of shape
                `(batch, src_len)` where padding elements are indicated by ``1``.
            for t_tgt, t_src is excluded (or masked out), =0 means it is
            included in attention

        Returns:
            encoded output of shape `(seq_len, batch, embed_dim)`
        """
        residual = x
        x, attn_weights = self.self_attn(
            query=x, key=x, value=x, key_padding_mask=encoder_padding_mask, need_weights=self.output_attentions,
        )
        x = F.dropout(x, p=self.dropout, training=self.training)
        x = residual + x
        x = self.self_attn_layer_norm(x)

        residual = x
        x = self.activation_fn(self.fc1(x))
        x = F.dropout(x, p=self.activation_dropout, training=self.training)
        x = self.fc2(x)
        x = F.dropout(x, p=self.dropout, training=self.training)
        x = residual + x
        x = self.final_layer_norm(x)
        return x, attn_weights


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

    Args:
        config: BartConfig
    """

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

        self.dropout = config.dropout
        self.layerdrop = config.encoder_layerdrop
        self.output_attentions = config.output_attentions
        self.output_hidden_states = config.output_hidden_states

        embed_dim = embed_tokens.embedding_dim
        self.padding_idx = embed_tokens.padding_idx
        self.max_source_positions = config.max_position_embeddings

        self.embed_tokens = embed_tokens

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

    def forward(
        self, input_ids=None, attention_mask=None,
    ):
        """
        Args:
            input_ids (LongTensor): tokens in the source language of shape
                `(batch, src_len)`
            attention_mask (torch.LongTensor): indicating which indices are padding tokens.
        Returns:
            namedtuple:
                - **x** (Tensor): the last encoder layer's output of
                  shape `(src_len, batch, embed_dim)`

                - **encoder_states** (List[Tensor]): all intermediate
                  hidden states of shape `(src_len, batch, embed_dim)`.
                  Only populated if *return_all_hiddens* is True.
                - **all_attentions** (List[Tensor]): Attention weights for each layer.
                During training might not be of length n_layers because of layer dropout.
        """
        inputs_embeds = self.embed_tokens(input_ids)
        embed_pos = self.embed_positions(input_ids)
        x = inputs_embeds + embed_pos
        x = self.layernorm_embedding(x)
        x = F.dropout(x, p=self.dropout, training=self.training)

        # B x T x C -> T x B x C
        x = x.transpose(0, 1)

        encoder_states, all_attentions = [], []

        # encoder layers
        for encoder_layer in self.layers:

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

            if self.output_attentions:
                all_attentions.append(attn)

        if self.output_hidden_states:
            encoder_states.append(x)

        encoder_states = [hidden_state.transpose(0, 1) for hidden_state in encoder_states]

        return x, encoder_states, all_attentions


class DecoderLayer(nn.Module):
    def __init__(self, config: BartConfig):
        super().__init__()
        self.embed_dim = config.d_model
        self.self_attn = SelfAttention(
            embed_dim=self.embed_dim, num_heads=config.decoder_attention_heads, dropout=config.attention_dropout,
        )
        self.dropout = config.dropout
        self.activation_fn = F.gelu
        self.activation_dropout = config.activation_dropout

        self.self_attn_layer_norm = LayerNorm(self.embed_dim)
        self.encoder_attn = SelfAttention(
            self.embed_dim,
            config.decoder_attention_heads,
            dropout=config.attention_dropout,
            encoder_decoder_attention=True,
        )
        self.encoder_attn_layer_norm = 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 = LayerNorm(self.embed_dim)

    def forward(
        self,
        x,
        encoder_hidden_states,
        encoder_attn_mask=None,
        layer_state=None,
        attention_mask=None,
        need_attn_weights=False,
    ):
        """
        Args:
            x (Tensor): input to the layer of shape `(seq_len, batch, embed_dim)`
            encoder_attn_mask (ByteTensor, optional): binary
                ByteTensor of shape `(batch, src_len)` where padding
                elements are indicated by ``1``.
            need_attn_weights (bool, optional): return attention weights
                for each head (default: return average over heads).

        Returns:
            encoded output of shape `(seq_len, batch, embed_dim)`
        """

        residual = x
        y = x  # TODO(SS): figure out why fairseq did this, then hopefully delete it

        if layer_state is None:
            layer_state = {}
        # next line mutates layer state
        x, self_attn_weights = self.self_attn(
            query=x, key=y, value=y, layer_state=layer_state, need_weights=need_attn_weights, attn_mask=attention_mask,
        )
        x = F.dropout(x, p=self.dropout, training=self.training)
        x = residual + x
        x = self.self_attn_layer_norm(x)
        residual = x
        assert self.encoder_attn.cache_key != self.self_attn.cache_key

        x, encoder_attn_weights = self.encoder_attn(
            query=x,
            key=encoder_hidden_states,  # could be None
            value=encoder_hidden_states,
            key_padding_mask=encoder_attn_mask,
            layer_state=layer_state,  # mutates layer state
            static_kv=True,
            need_weights=False,  # not returning it so why compute it
        )
        x = F.dropout(x, p=self.dropout, training=self.training)
        x = residual + x

        x = self.encoder_attn_layer_norm(x)

        residual = x
        x = self.activation_fn(self.fc1(x))
        x = F.dropout(x, p=self.activation_dropout, training=self.training)
        x = self.fc2(x)
        x = F.dropout(x, p=self.dropout, training=self.training)
        x = residual + x
        x = self.final_layer_norm(x)
        return (
            x,
            self_attn_weights,
            layer_state,
        )  # just self_attn weights for now, following t5, layer_state = cache for decoding


class BartDecoder(nn.Module):
    """
    Transformer decoder consisting of *config.decoder_layers* layers. Each layer
    is a :class:`DecoderLayer`.
    Args:
        config: BartConfig
        embed_tokens (torch.nn.Embedding): output embedding
    """

    def __init__(self, config: BartConfig, embed_tokens: nn.Embedding):
        super().__init__()
        self.output_past = config.output_past
        self.output_attentions = config.output_attentions
        self.output_hidden_states = config.output_hidden_states
        self.dropout = config.dropout
        self.layerdrop = config.decoder_layerdrop
        self.padding_idx = embed_tokens.padding_idx
        self.max_target_positions = config.max_position_embeddings
        self.embed_tokens = embed_tokens
        self.embed_positions = LearnedPositionalEmbedding(
            config.max_position_embeddings, config.d_model, self.padding_idx,
        )
        self.layers = nn.ModuleList(
            [DecoderLayer(config) for _ in range(config.decoder_layers)]
        )  # type: List[DecoderLayer]
        self.layernorm_embedding = LayerNorm(config.d_model)
        self.generation_mode = False

    def forward(
        self,
        input_ids,
        encoder_hidden_states,
        encoder_padding_mask,
        combined_mask,
        decoder_cached_states=None,
        **unused
    ):
        """
        Includes several features from "Jointly Learning to Align and
        Translate with Transformer Models" (Garg et al., EMNLP 2019).

        Args:
            input_ids (LongTensor): previous decoder outputs of shape
                `(batch, tgt_len)`, for teacher forcing
            encoder_hidden_states: output from the encoder, used for
                encoder-side attention
            encoder_padding_mask: for ignoring pad tokens
            decoder_cached_states (dict or None): dictionary used for storing state during generation

        Returns:
            tuple:
                - the decoder's features of shape `(batch, tgt_len, embed_dim)`
                - hidden states
                - attentions
        """
        # embed positions
        positions = self.embed_positions(input_ids, generation_mode=self.generation_mode)

        if self.generation_mode:
            input_ids = input_ids[:, -1:]
            positions = positions[:, -1:]  # happens after we embed them
            assert input_ids.ne(self.padding_idx).any()

        x = self.embed_tokens(input_ids)
        x += positions

        x = self.layernorm_embedding(x)
        x = F.dropout(x, p=self.dropout, training=self.training)
        x = x.transpose(0, 1)  # (seq_len, BS, model_dim)
        # decoder layers
        all_hidden_states = ()
        all_self_attns = ()
        next_decoder_cache = []

        for i, decoder_layer in enumerate(self.layers):
            decoder_layer  # type: DecoderLayer
            # 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):
                continue

            layer_state = decoder_cached_states[i] if decoder_cached_states is not None else None
            x, layer_self_attn, layer_past = decoder_layer(
                x,
                encoder_hidden_states,
                encoder_padding_mask,
                layer_state=layer_state,
                attention_mask=combined_mask,
                need_attn_weights=self.output_attentions,
            )

            if self.output_past:
                next_decoder_cache.append(layer_past.copy())
            if self.output_hidden_states:
                all_hidden_states += (x,)
            if self.output_attentions:
                all_self_attns += (layer_self_attn,)

        # Convert shapes from (seq_len, BS, model_dim) to (BS, seq_len, model_dim)
        all_hidden_states = [hidden_state.transpose(0, 1) for hidden_state in all_hidden_states]
        x = x.transpose(0, 1)

        if self.output_past:
            next_cache = ((encoder_hidden_states, encoder_padding_mask), next_decoder_cache)
        else:
            next_cache = None
        return x, next_cache, all_hidden_states, list(all_self_attns)


def reorder_attn_buffer(input_buffer, new_order):
    """Reorder buffered internal state (for incremental generation)."""
    # input_buffer = self._get_input_buffer(incremental_state)
    for k in input_buffer.keys():
        input_buffer_k = input_buffer[k]
        if input_buffer_k is not None:
            input_buffer[k] = input_buffer_k.index_select(0, new_order)
        # incremental_state = self._set_input_buffer(incremental_state, input_buffer)
    return input_buffer


class SelfAttention(nn.Module):
    """Multi-headed attention from "Attention Is All You Need"""

    def __init__(
        self,
        embed_dim,
        num_heads,
        kdim=None,
        vdim=None,
        dropout=0.0,
        bias=True,
        encoder_decoder_attention=False,  # otherwise self_attention
    ):
        super().__init__()
        self.embed_dim = embed_dim
        self.kdim = kdim if kdim is not None else embed_dim
        self.vdim = vdim if vdim is not None else 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, "embed_dim must be divisible by num_heads"
        self.scaling = self.head_dim ** -0.5

        self.encoder_decoder_attention = encoder_decoder_attention
        qkv_same_dim = self.kdim == embed_dim and self.vdim == embed_dim  # True for all BART

        assert self.encoder_decoder_attention or qkv_same_dim, (
            "Self-attention requires query, key and " "value to be of the same size"
        )
        self.k_proj = nn.Linear(self.kdim, embed_dim, bias=bias)
        self.v_proj = nn.Linear(self.vdim, 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)
        self.cache_key = "encoder_decoder" if self.encoder_decoder_attention else "self"

    def _shape(self, tensor, dim_0, bsz):
        return tensor.contiguous().view(dim_0, bsz * self.num_heads, self.head_dim).transpose(0, 1)

    def forward(
        self,
        query,
        key: Optional[Tensor],
        value: Optional[Tensor],
        key_padding_mask: Optional[Tensor] = None,
        layer_state: Optional[Dict[str, Dict[str, Optional[Tensor]]]] = None,
        need_weights: bool = False,
        static_kv: bool = False,
        attn_mask: Optional[Tensor] = None,
    ) -> Tuple[Tensor, Optional[Tensor]]:
        """Input shape: Time(SeqLen) x Batch x Channel

        Args:

            key_padding_mask (ByteTensor, optional): mask to exclude
                keys that are pads, of shape `(batch, src_len)`, where
                padding elements are indicated by 1s.
            need_weights (bool, optional): return the attention weights,
                averaged over heads (default: False).
            attn_mask (ByteTensor, optional): typically used to
                implement causal attention, where the mask prevents the
                attention from looking forward in time (default: None).
        """
        tgt_len, bsz, embed_dim = query.size()
        assert embed_dim == self.embed_dim
        assert list(query.size()) == [tgt_len, bsz, embed_dim]
        # get here for encoder decoder cause of static_kv
        if layer_state is not None:  # get the last k,v and mask for reuse
            saved_state = layer_state.get(self.cache_key, {})
            if "prev_key" in saved_state:
                # previous time steps are cached - no need to recompute key and value if they are static
                if static_kv:
                    assert self.encoder_decoder_attention
                    key = value = None
        else:
            saved_state = None
            layer_state = {}

        q = self.q_proj(query) * self.scaling
        if self.encoder_decoder_attention:
            if key is None:
                assert value is None
                k = v = None
            else:
                k = self.k_proj(key)
                v = self.v_proj(key)
        else:
            k = self.k_proj(query)
            v = self.v_proj(query)

        q = self._shape(q, tgt_len, bsz)
        if k is not None:
            k = self._shape(k, -1, bsz)
        if v is not None:
            v = self._shape(v, -1, bsz)

        if saved_state is not None:
            k, v, key_padding_mask = self._use_saved_state(k, v, saved_state, key_padding_mask, static_kv, bsz)
        # assert self.cache_key != 'encoder_decoder' or key_padding_mask is None

        # Update cache
        layer_state[self.cache_key] = {
            "prev_key": k.view(bsz, self.num_heads, -1, self.head_dim),
            "prev_value": v.view(bsz, self.num_heads, -1, self.head_dim),
            "prev_key_padding_mask": key_padding_mask if not static_kv else None,
        }

        assert k is not None
        src_len = k.size(1)
        attn_weights = torch.bmm(q, k.transpose(1, 2))

        assert attn_weights.size() == (bsz * self.num_heads, tgt_len, src_len)

        if attn_mask is not None:
            attn_weights = attn_weights.view(bsz, self.num_heads, tgt_len, src_len) + attn_mask
            attn_weights = attn_weights.view(bsz * self.num_heads, tgt_len, src_len)

        # This is part of a workaround to get around fork/join parallelism not supporting Optional types.
        if key_padding_mask is not None and key_padding_mask.dim() == 0:
            key_padding_mask = None
        assert key_padding_mask is None or key_padding_mask.size()[:2] == (bsz, src_len,)

        if key_padding_mask is not None:  # don't attend to padding symbols
            attn_weights = attn_weights.view(bsz, self.num_heads, tgt_len, src_len)
            reshaped = key_padding_mask.unsqueeze(1).unsqueeze(2).to(torch.bool)
            attn_weights = attn_weights.masked_fill(reshaped, float("-inf"))
            attn_weights = attn_weights.view(bsz * self.num_heads, tgt_len, src_len)
        attn_weights = F.softmax(attn_weights, dim=-1)
        attn_probs = F.dropout(attn_weights, p=self.dropout, training=self.training,)

        assert v is not None
        attn_output = torch.bmm(attn_probs, v)
        assert attn_output.size() == (bsz * self.num_heads, tgt_len, self.head_dim)
        attn_output = attn_output.transpose(0, 1).contiguous().view(tgt_len, bsz, embed_dim)
        attn_output = self.out_proj(attn_output)
        attn_weights = attn_weights.view(bsz, self.num_heads, tgt_len, src_len)
        return attn_output, attn_weights

    def _use_saved_state(self, k, v, saved_state, key_padding_mask, static_kv, bsz):
        # saved states are stored with shape (bsz, num_heads, seq_len, head_dim)
        if "prev_key" in saved_state:
            _prev_key = saved_state["prev_key"]
            assert _prev_key is not None
            prev_key = _prev_key.view(bsz * self.num_heads, -1, self.head_dim)
            if static_kv:
                k = prev_key
            else:
                assert k is not None
                k = torch.cat([prev_key, k], dim=1)
        if "prev_value" in saved_state:
            _prev_value = saved_state["prev_value"]
            assert _prev_value is not None
            prev_value = _prev_value.view(bsz * self.num_heads, -1, self.head_dim)
            if static_kv:
                v = prev_value
            else:
                assert v is not None
                v = torch.cat([prev_value, v], dim=1)
        assert k is not None and v is not None
        prev_key_padding_mask = saved_state.get("prev_key_padding_mask", None)  # type: Optional[Tensor]
        key_padding_mask = self._cat_prev_key_padding_mask(
            key_padding_mask, prev_key_padding_mask, bsz, k.size(1), static_kv
        )
        return k, v, key_padding_mask

    @staticmethod
    def _cat_prev_key_padding_mask(
        key_padding_mask: Optional[Tensor],
        prev_key_padding_mask: Optional[Tensor],
        batch_size: int,
        src_len: int,
        static_kv: bool,
    ) -> Optional[Tensor]:
        # saved key padding masks have shape (bsz, seq_len)
        if prev_key_padding_mask is not None and static_kv:
            new_key_padding_mask = prev_key_padding_mask
        elif prev_key_padding_mask is not None and key_padding_mask is not None:
            new_key_padding_mask = torch.cat([prev_key_padding_mask.float(), key_padding_mask.float()], dim=1)
        # During incremental decoding, as the padding token enters and
        # leaves the frame, there will be a time when prev or current is None
        elif prev_key_padding_mask is not None:
            filler = torch.zeros(batch_size, src_len - prev_key_padding_mask.size(1))
            if prev_key_padding_mask.is_cuda:
                filler = filler.to(prev_key_padding_mask.device)
            new_key_padding_mask = torch.cat([prev_key_padding_mask.float(), filler.float()], dim=1)
        elif key_padding_mask is not None:
            filler = torch.zeros(batch_size, src_len - key_padding_mask.size(1))
            if key_padding_mask.is_cuda:
                filler = filler.cuda()
            new_key_padding_mask = torch.cat([filler.float(), key_padding_mask.float()], dim=1)
        else:
            new_key_padding_mask = prev_key_padding_mask
        return new_key_padding_mask


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

    # This can trivially be shared with RobertaClassificationHead

    def __init__(
        self, input_dim, inner_dim, num_classes, pooler_dropout,
    ):
        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, x):
        x = self.dropout(x)
        x = self.dense(x)
        x = torch.tanh(x)
        x = self.dropout(x)
        x = self.out_proj(x)
        return x


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

    def __init__(
        self, num_embeddings: int, embedding_dim: int, padding_idx: int,
    ):
        # if padding_idx is specified then offset the embedding ids by
        # this index and adjust num_embeddings appropriately
        assert padding_idx is not None
        num_embeddings += padding_idx + 1  # WHY?
        super().__init__(num_embeddings, embedding_dim, padding_idx=padding_idx)

    def forward(self, input, generation_mode=False):
        """Input is expected to be of size [bsz x seqlen]."""
        if generation_mode:  # the position is our current step in the decoded sequence
            pos = int(self.padding_idx + input.size(1))
            positions = input.data.new(1, 1).fill_(pos)
        else:
            positions = create_position_ids_from_input_ids(input, self.padding_idx)
        return super().forward(positions)


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

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


def fill_with_neg_inf(t):
    """FP16-compatible function that fills a input_ids with -inf."""
    return t.float().fill_(float("-inf")).type_as(t)


def _filter_out_falsey_values(tup) -> Tuple:
    """Remove entries that are None or [] from an iterable."""
    return tuple(x for x in tup if isinstance(x, torch.Tensor) or x)


RET_DOCSTRING = r"""
    Return:
        :obj:`tuple(torch.FloatTensor)` comprising various elements depending on the configuration (:class:`~transformers.BertConfig`) and inputs:
        last_hidden_state (:obj:`torch.FloatTensor` of shape :obj:`(batch_size, sequence_length, hidden_size)`):
            Sequence of hidden-states at the output of the last layer of the model.
        hidden_states (:obj:`tuple(torch.FloatTensor)`, `optional`, returned when ``config.output_hidden_states=True``):
            Tuple of :obj:`torch.FloatTensor` (one for the output of the embeddings + one for the output of each layer)
            of shape :obj:`(batch_size, sequence_length, hidden_size)`.
            Hidden-states of the model at the output of each layer plus the initial embedding outputs.
        attentions (:obj:`tuple(torch.FloatTensor)`, `optional`, returned when ``config.output_attentions=True``):
            Tuple of :obj:`torch.FloatTensor` (one for each layer) of shape
            :obj:`(batch_size, num_heads, sequence_length, sequence_length)`.
            Attentions weights after the attention softmax, used to compute the weighted average in the self-attention
            heads.
"""
# Public API


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

        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()

    @add_start_docstrings_to_callable(BART_INPUTS_DOCSTRING)
    def forward(
        self,
        input_ids,
        attention_mask=None,
        decoder_input_ids=None,
        encoder_outputs=None,  # type: Tuple
        decoder_attention_mask=None,
        decoder_cached_states=None,
    ):
        if attention_mask is not None:
            assert attention_mask.dim() == 2

            attention_mask = (1.0 - attention_mask.long()) * -10000.0
            assert attention_mask.max() <= 0

        # make masks if user doesn't supply
        if not self.decoder.generation_mode:
            decoder_input_ids, decoder_attention_mask = _prepare_bart_decoder_inputs(
                self.config, input_ids, decoder_input_ids=decoder_input_ids, decoder_attn_mask=decoder_attention_mask,
            )
        assert decoder_input_ids is not None
        if encoder_outputs is None:
            encoder_outputs = self.encoder(input_ids=input_ids, attention_mask=attention_mask)
        assert isinstance(encoder_outputs, tuple)
        # decoder outputs consists of (dec_features, layer_state, dec_hidden, dec_attn)
        decoder_outputs = self.decoder(
            decoder_input_ids,
            encoder_outputs[0],
            attention_mask,
            decoder_attention_mask,
            decoder_cached_states=decoder_cached_states,
        )
        # Attention and hidden_states will be [] or None if they aren't needed
        decoder_outputs = _filter_out_falsey_values(decoder_outputs)  # type: tuple
        assert isinstance(decoder_outputs[0], torch.Tensor)
        encoder_outputs = _filter_out_falsey_values(encoder_outputs)  # type: tuple
        return decoder_outputs + encoder_outputs

    def get_input_embeddings(self):
        return self.shared

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

    def get_output_embeddings(self):
        return _make_linear_from_emb(self.shared)  # make it on the fly


@add_start_docstrings(
    "The bare BART Model with a language modeling head. This is the model used for summarization.",
    BART_START_DOCSTRING,
)
class BartForMaskedLM(PretrainedBartModel):
    base_model_prefix = "model"

    def __init__(self, config: BartConfig):
        super().__init__(config)
        # if base_model is None:
        base_model = BartModel(config)
        self.model = base_model
        self.lm_head = _make_linear_from_emb(self.model.shared)

    def tie_weights(self):
        pass  # hack to prevent changing lm_head.out_features. The input and output embeddings are still the same.

    @add_start_docstrings_to_callable(BART_INPUTS_DOCSTRING)
    def forward(
        self,
        input_ids,
        attention_mask=None,
        encoder_outputs=None,
        decoder_input_ids=None,
        decoder_attention_mask=None,
        decoder_cached_states=None,
        lm_labels=None,
        **unused
    ):
        r"""
        masked_lm_labels (:obj:`torch.LongTensor` of shape :obj:`(batch_size, sequence_length)`, `optional`, defaults to :obj:`None`):
            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:
        :obj:`tuple(torch.FloatTensor)` comprising various elements depending on the configuration (:class:`~transformers.RobertaConfig`) and inputs:
        masked_lm_loss (`optional`, returned when ``masked_lm_labels`` is provided) ``torch.FloatTensor`` of shape ``(1,)``:
            Masked language modeling loss.
        prediction_scores (:obj:`torch.FloatTensor` of shape :obj:`(batch_size, sequence_length, config.vocab_size)`)
            Prediction scores of the language modeling head (scores for each vocabulary token before SoftMax).
        hidden_states (:obj:`tuple(torch.FloatTensor)`, `optional`, returned when ``config.output_hidden_states=True``):
            Tuple of :obj:`torch.FloatTensor` (one for the output of the embeddings + one for the output of each layer)
            of shape :obj:`(batch_size, sequence_length, hidden_size)`.

            Hidden-states of the model at the output of each layer plus the initial embedding outputs.
        attentions (:obj:`tuple(torch.FloatTensor)`, `optional`, returned when ``config.output_attentions=True``):
            Tuple of :obj:`torch.FloatTensor` (one for each layer) of shape
            :obj:`(batch_size, num_heads, sequence_length, sequence_length)`.

            Attentions weights after the attention softmax, used to compute the weighted average in the self-attention
            heads.

    Examples::

            tokenizer = BartTokenizer.from_pretrained('bart-large')
            model = BartForMaskedLM.from_pretrained('bart-large')
            input_ids = torch.tensor(tokenizer.encode("Hello, my dog is cute")).unsqueeze(0)  # Batch size 1
            outputs = model(input_ids=input_ids, lm_labels=input_ids)
            loss, prediction_scores = outputs[:2]
        """
        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,
            decoder_cached_states=decoder_cached_states,
        )
        lm_logits = self.lm_head(outputs[0])
        outputs = (lm_logits,) + outputs[1:]  # Add hidden states and attention if they are here
        if lm_labels is not None:
            loss_fct = nn.CrossEntropyLoss()
            # TODO(SS): do we need to ignore pad tokens in lm_labels?
            masked_lm_loss = loss_fct(lm_logits.view(-1, self.config.vocab_size), lm_labels.view(-1))
            outputs = (masked_lm_loss,) + outputs

        return outputs

    @staticmethod
    def prepare_inputs_for_generation(input_ids, past, decoder_input_ids, attention_mask):
        if past is None:  # first step
            encoder_outputs, decoder_cached_states = None, None
        else:
            encoder_outputs, decoder_cached_states = past
        return {
            "input_ids": input_ids,  # ignored after first pass
            "decoder_cached_states": decoder_cached_states,
            "decoder_input_ids": decoder_input_ids,
            "encoder_outputs": encoder_outputs,
            "attention_mask": attention_mask,
            # "decoder_attention_mask": decoder_attention_mask,
        }

    @staticmethod
    def _reorder_cache(past, beam_idx):
        ((enc_out, enc_mask), decoder_cached_states) = past
        reordered_past = []
        for layer_past in decoder_cached_states:
            # get the correct batch idx from decoder layer's batch dim for cross and self-attn
            layer_past_new = {
                attn_key: reorder_attn_buffer(attn_cache, beam_idx) for attn_key, attn_cache in layer_past.items()
            }
            # reordered_layer_past = [layer_past[:, i].unsqueeze(1).clone().detach() for i in beam_idx]
            # reordered_layer_past = torch.cat(reordered_layer_past, dim=1)
            reordered_past.append(layer_past_new)
        new_enc_out = enc_out if enc_out is None else enc_out.index_select(1, beam_idx)
        new_enc_mask = enc_mask if enc_mask is None else enc_mask.index_select(0, beam_idx)

        past = ((new_enc_out, new_enc_mask), reordered_past)
        return past

    def get_output_embeddings(self):
        return self.lm_head

    @torch.no_grad()
    def generate(
        self,
        input_ids,
        attention_mask=None,
        max_length=20,
        num_beams=1,
        repetition_penalty=1.0,
        length_penalty=1.0,
        num_return_sequences=1,
        min_len=0,
        no_repeat_ngram_size=0,
    ):
        r""" Generates sequences for models with a LM head. The method currently supports greedy or penalized greedy decoding, sampling with top-k or nucleus sampling
        and beam-search.

        Adapted in part from Facebook's `XLM beam search code`_ and `Fairseq beam search code`_.

        .. _`XLM beam search code`:
           https://github.com/facebookresearch/XLM/blob/9e6f6814d17be4fe5b15f2e6c43eb2b2d76daeb4/src/model/transformer.py#L529
        .. _`Fairseq beam search code`:
           https://github.com/pytorch/fairseq/blob/master/fairseq/sequence_generator.py


        Parameters:

            input_ids: (`optional`) `torch.LongTensor` of shape `(batch_size, sequence_length)`
                The sequence used as a prompt for the generation. If `None` the method initializes
                it as an empty `torch.LongTensor` of shape `(1,)`.

            max_length: (`optional`) int
                The max length of the sequence to be generated. Does not include tokens in input_ids.

            num_beams: (`optional`) int
                Number of beams for beam search. Must be between 1 and infinity. 1 means no beam search. Default to 1.

            repetition_penalty: (`optional`) float
                The parameter for repetition penalty. Between 1.0 and infinity. 1.0 means no penalty. Default to 1.0.

            length_penalty: (`optional`) float
                Exponential penalty to the length. Default to 1.

            num_return_sequences: (`optional`) int
                The number of independently computed returned sequences for each element in the batch. Default to 1.

            min_len: (`optional`) int

        Returns:
            `torch.LongTensor` of shape `(batch_size * num_return_sequences, sequence_length)`
                sequence_length is <= max_length (examples can finish early)

        Examples::

            config = BartConfig(vocab_size=50264, output_past=True)
            model = AutoModelWithLMHead.from_pretrained('bart-large-cnn', config=config)
            tokenizer = AutoTokenizer.from_pretrained('bart-large-cnn')
            ARTICLE_TO_SUMMARIZE = "My friends are cool but they eat too many carbs."
            inputs = tokenizer.batch_encode_plus([ARTICLE_TO_SUMMARIZE], max_length=1024, return_tensors='pt')
            # Generate Summary
            generated_ids = model.generate(inputs['input_ids'], attention_mask=inputs['attention_mask'], num_beams=4, max_length=5)
            print([tokenizer.decode(g, skip_special_tokens=True, clean_up_tokenization_spaces=False) for g in generated_ids])

        """
        bos_token_id = self.config.bos_token_id
        pad_token_id = self.config.pad_token_id
        eos_token_id = self.config.eos_token_id
        batch_size, cur_len = input_ids.shape
        assert input_ids is not None
        assert self.config.output_past, "Generating with bart requires instantiating a config with output_past=True"
        assert isinstance(max_length, int) and max_length > 0, "`max_length` should be a strictly positive integer."
        assert isinstance(num_beams, int) and num_beams > 0, "`num_beams` should be a strictly positive integer."
        assert repetition_penalty >= 1.0, "`repetition_penalty` should be >= 1."
        assert isinstance(pad_token_id, int)
        assert bos_token_id == 0, "configurable bos_token_id not yet supported"
        assert length_penalty > 0, "`length_penalty` should be strictly positive."
        assert (
            isinstance(num_return_sequences, int) and num_return_sequences > 0
        ), "`num_return_sequences` should be a positive integer."

        # current position and vocab size
        cur_len = input_ids.shape[1]
        vocab_size = self.config.vocab_size

        if num_return_sequences != 1:
            # Expand input to num return sequences
            input_ids = input_ids.unsqueeze(1).expand(batch_size, num_return_sequences, cur_len)
            input_ids = input_ids.contiguous().view(
                batch_size * num_return_sequences, cur_len
            )  # shape: (batch_size * num_return_sequences, cur_len)
            batch_size *= num_return_sequences

        # Below here somewhat similar to PretrainedModel._generate_beam_search
        # Expand input to num beams
        input_ids = input_ids.unsqueeze(1).expand(batch_size, num_beams, cur_len)

        input_ids = input_ids.contiguous().view(batch_size * num_beams, cur_len)  # (batch_size * num_beams, cur_len)
        if attention_mask is not None:
            attention_mask = (
                attention_mask.unsqueeze(1)
                .expand(batch_size, num_beams, cur_len)
                .contiguous()
                .view(batch_size * num_beams, cur_len)
            )  # RESHAPE

        # generated hypotheses
        finalized_hyps = [  # they end in EOS and we wont work on them more!
            BeamHypotheses(num_beams, max_length, length_penalty, early_stopping=True) for _ in range(batch_size)
        ]

        # scores for each sentence in the beam
        beam_scores = torch.zeros((batch_size, num_beams), dtype=torch.float, device=input_ids.device)
        beam_scores[:, 1:] = -1e9  # avoid ties in first step
        beam_scores = beam_scores.view(-1)  # shape (batch_size * num_beams,)

        # decoder tokens
        prev_output_tokens = input_ids.new(batch_size * num_beams, 1).long().fill_(-1)
        prev_output_tokens[:, 0] = 2  # HARDCODED EOS, which will be removed at the end.
        decoder_cache = None
        done = [False for _ in range(batch_size)]  # done sentences

        self.model.decoder.generation_mode = True  # tells decoder not to use causal mask
        for step in range(max_length + 1):
            decoder_input_ids = prev_output_tokens.clone()
            model_inputs = self.prepare_inputs_for_generation(
                input_ids, decoder_cache, decoder_input_ids, attention_mask,
            )
            outputs = self(**model_inputs)
            lprobs = F.log_softmax(outputs[0][:, -1, :], dim=-1)

            lprobs[lprobs != lprobs] = -math.inf  # block nans
            lprobs[:, pad_token_id] = -math.inf
            # TODO(SS): fairseq also takes out <unk> every step, and has unk at slot 3

            if step == 0:  # Force BOS to be chosen
                lprobs[:, bos_token_id + 1 :] = -math.inf
            elif step < min_len:  # Prevent EOS from being chosen
                lprobs[:, eos_token_id] = -math.inf
            elif step == max_length:  # FORCE EOS to be chosen
                lprobs[:, :eos_token_id] = -math.inf
                lprobs[:, eos_token_id + 1 :] = -math.inf
            assert self._do_output_past(outputs)
            decoder_cache = outputs[1]
            if repetition_penalty != 1.0:
                self.enforce_repetition_penalty_(lprobs, batch_size, num_beams, prev_output_tokens, repetition_penalty)
            num_hypos = batch_size * num_beams
            if no_repeat_ngram_size > 0:  # copied from fairseq
                # for each sentence, calculate a list of banned tokens to prevent repetitively generating the same ngrams
                banned_tokens = self.calc_banned_tokens(prev_output_tokens, num_hypos, no_repeat_ngram_size, step)
                # then set their probabilities tof -inf
                for idx in range(num_hypos):
                    lprobs[idx, banned_tokens[idx]] = -math.inf
            assert lprobs.size() == (batch_size * num_beams, vocab_size)
            _scores = lprobs + beam_scores[:, None].expand_as(lprobs)  # (batch_size * num_beams, vocab_size)

            # re-organize to group the beam together (we are keeping top hypothesis across beams)
            _scores = _scores.view(batch_size, num_beams * vocab_size)  # (batch_size, num_beams * vocab_size)
            # Take the best 2 x beam_size predictions for each example, we'll choose the first beam_size of these which don't predict eos to continue with.
            next_scores, next_words = torch.topk(_scores, 2 * num_beams)
            assert next_scores.size() == next_words.size() == (batch_size, 2 * num_beams)

            # list of (batch_size * num_beams)
            next_batch_beam = []  # Tuple(next score, next word, current position in the batch)
            for batch_idx in range(batch_size):
                # if we are done with this sentence (because we can't improve)
                if done[batch_idx]:  # then pad all associated hypotheses
                    assert (
                        len(finalized_hyps[batch_idx]) >= num_beams
                    ), "Example can only be done if at least {} beams have been generated".format(num_beams)
                    next_batch_beam.extend([(0, pad_token_id, 0)] * num_beams)  # pad the batch
                    continue

                # Otherwise generate some next word choices
                next_sent_beam = []
                # add next words for this sentence
                for i, (idx, score) in enumerate(zip(next_words[batch_idx], next_scores[batch_idx])):
                    beam_id = idx // vocab_size
                    word_id = idx % vocab_size
                    assert prev_output_tokens.shape[1] == (step + 1)
                    if word_id.item() == eos_token_id:
                        if i >= num_beams:
                            continue
                        finalized_hyps[batch_idx].add(
                            prev_output_tokens[batch_idx * num_beams + beam_id].clone(), score.item(),
                        )
                    else:
                        next_sent_beam.append((score, word_id, batch_idx * num_beams + beam_id))

                    if len(next_sent_beam) == num_beams:  # TODO(SS): can we delete this?
                        break
                # Check if were done so that we can save a pad step if all(done)
                done[batch_idx] = done[batch_idx] or finalized_hyps[batch_idx].is_done(
                    next_scores[batch_idx].max().item(), cur_len=step + 1,
                )
                assert len(next_sent_beam) == num_beams, "Beam should always be full"
                next_batch_beam.extend(next_sent_beam)
                assert len(next_batch_beam) == num_beams * (batch_idx + 1)

            if all(done):
                break

            # sanity check / prepare next batch
            assert len(next_batch_beam) == batch_size * num_beams
            beam_scores = beam_scores.new([x[0] for x in next_batch_beam])
            beam_words = input_ids.new([x[1] for x in next_batch_beam])
            beam_idx = input_ids.new([x[2] for x in next_batch_beam])
            # re-order decoder inputs to [beam_idx]
            prev_output_tokens = prev_output_tokens[beam_idx]
            prev_output_tokens = torch.cat([prev_output_tokens, beam_words.unsqueeze(1)], dim=-1)

            # re-order internal states
            decoder_cache = self._reorder_cache(decoder_cache, beam_idx)

        for batch_idx in range(batch_size):
            # Add all open beam hypothesis to generated_hyps
            if done[batch_idx]:
                continue
            offset = batch_idx * num_beams
            for i in range(num_beams):
                score = beam_scores[offset + i]
                final_tokens = prev_output_tokens[offset + i]
                finalized_hyps[batch_idx].add(final_tokens, score.item())

        # select the best hypotheses
        sent_lengths = input_ids.new(batch_size)
        best = []
        for i, hypotheses in enumerate(finalized_hyps):
            best_hyp = max(hypotheses.beams, key=lambda x: x[0])[1]
            sent_lengths[i] = len(best_hyp)
            best.append(best_hyp)

        # shorter batches are filled with pad_token
        if sent_lengths.min().item() != sent_lengths.max().item():
            # TODO(SS): decoded = torch.rnn.utils.pad_sequence(best, batch_first=True, padding_value=pad_token_id)
            sent_max_len = min(sent_lengths.max().item() + 1, max_length + 1)  # TODO(SS): same as step?
            decoded = input_ids.new(batch_size, sent_max_len).fill_(pad_token_id)
            # fill with hypothesis and eos_token_id if necessary
            for i, hypo in enumerate(best):
                decoded[i, : sent_lengths[i]] = hypo
                if sent_lengths[i] < max_length:
                    decoded[i, sent_lengths[i]] = eos_token_id
        else:
            assert (len(hypo) == max_length for hypo in best)
            decoded = torch.stack(best).type(torch.long).to(next(self.parameters()).device)
        return decoded[:, 1:]  # get rid of starting EOS

    @staticmethod
    def calc_banned_tokens(prev_output_tokens, num_hypos, no_repeat_ngram_size, step):
        """Copied from fairseq for no_repeat_ngram in beam_search"""
        # TODO(SS): this can go on parent if there is demand
        if step + 2 < no_repeat_ngram_size:
            return [
                [] for _ in range(num_hypos)
            ]  # no banned tokens if we haven't generated no_repeat_ngram_size tokens yet
        gen_ngrams = [{} for _ in range(num_hypos)]
        for idx in range(num_hypos):
            gen_tokens = prev_output_tokens[idx].tolist()
            for ngram in zip(*[gen_tokens[i:] for i in range(no_repeat_ngram_size)]):
                k = tuple(ngram[:-1])
                gen_ngrams[idx][k] = gen_ngrams[idx].get(k, []) + [ngram[-1]]

        def _get_generated_ngrams(hypo_idx):
            """Before decoding the next token, prevent decoding of ngrams that have already appeared"""
            ngram_index = tuple(prev_output_tokens[hypo_idx, step + 2 - no_repeat_ngram_size : step + 1].tolist())
            return gen_ngrams[hypo_idx].get(ngram_index, [])

        banned_tokens = [_get_generated_ngrams(hypo_idx) for hypo_idx in range(num_hypos)]
        return banned_tokens


@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(PretrainedBartModel):
    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.classif_dropout,
        )
        self.model._init_weights(self.classification_head.dense)
        self.model._init_weights(self.classification_head.out_proj)

    @add_start_docstrings_to_callable(BART_INPUTS_DOCSTRING)
    def forward(
        self,
        input_ids,
        attention_mask=None,
        encoder_outputs=None,
        decoder_input_ids=None,
        decoder_attention_mask=None,
        labels=None,
    ):
        r"""
        labels (:obj:`torch.LongTensor` of shape :obj:`(batch_size,)`, `optional`, defaults to :obj:`None`):
            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).

    Returns:
        :obj:`tuple(torch.FloatTensor)` comprising various elements depending on the configuration (:class:`~transformers.BartConfig`) and inputs:
            loss (:obj:`torch.FloatTensor` of shape :obj:`(1,)`, `optional`, returned when :obj:`label` is provided):
                Classification  loss (cross entropy)
            logits (:obj:`torch.FloatTensor` of shape :obj:`(batch_size, config.num_labels)`):
                Classification (or regression if config.num_labels==1) scores (before SoftMax).
            hidden_states (:obj:`tuple(torch.FloatTensor)`, `optional`, returned when ``config.output_hidden_states=True``):
                Tuple of :obj:`torch.FloatTensor` (one for the output of the embeddings + one for the output of each layer)
                of shape :obj:`(batch_size, sequence_length, hidden_size)`.
                Hidden-states of the model at the output of each layer plus the initial embedding outputs.
            attentions (:obj:`tuple(torch.FloatTensor)`, `optional`, returned when ``config.output_attentions=True``):
                Tuple of :obj:`torch.FloatTensor` (one for each layer) of shape :obj:`(batch_size, num_heads, sequence_length, sequence_length)`.
                Attentions weights after the attention softmax, used to compute the weighted average in the
                self-attention
                heads.

    Examples::

        from transformers import BartTokenizer, BartForSequenceClassification
        import torch

        tokenizer = BartTokenizer.from_pretrained('bart-large')
        model = BartForSequenceClassification.from_pretrained('bart-large')
        input_ids = torch.tensor(tokenizer.encode("Hello, my dog is cute",
        add_special_tokens=True)).unsqueeze(0)  # Batch size 1
        labels = torch.tensor([1]).unsqueeze(0)  # Batch size 1
        outputs = model(input_ids, labels=labels)
        loss, logits = outputs[:2]

        """
        outputs = self.model(
            input_ids,
            attention_mask=attention_mask,
            decoder_input_ids=decoder_input_ids,
            decoder_attention_mask=decoder_attention_mask,
            encoder_outputs=encoder_outputs,
        )
        x = 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 = x[eos_mask, :].view(x.size(0), -1, x.size(-1))[:, -1, :]
        logits = self.classification_head(sentence_representation)
        # Prepend logits
        outputs = (logits,) + outputs[1:]  # Add hidden states and attention if they are here
        if labels is not None:  # prepend loss to output,
            loss = F.cross_entropy(logits.view(-1, self.config.num_labels), labels.view(-1))
            outputs = (loss,) + outputs

        return outputs