transformers_4_35_0/models/bigbird_pegasus/modeling_bigbird_pegasus.py

# coding=utf-8
# Copyright 2021 Google Research 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 BigBirdPegasus model."""


import copy
import math
from typing import List, Optional, Tuple, Union

import numpy as np
import torch
from torch import nn
from torch.nn import BCEWithLogitsLoss, CrossEntropyLoss, MSELoss

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


logger = logging.get_logger(__name__)

_CHECKPOINT_FOR_DOC = "google/bigbird-pegasus-large-arxiv"
_CONFIG_FOR_DOC = "BigBirdPegasusConfig"
_EXPECTED_OUTPUT_SHAPE = [1, 7, 1024]


BIGBIRD_PEGASUS_PRETRAINED_MODEL_ARCHIVE_LIST = [
    "google/bigbird-pegasus-large-arxiv",
    "google/bigbird-pegasus-large-pubmed",
    "google/bigbird-pegasus-large-bigpatent",
    # See all BigBirdPegasus models at https://huggingface.co/models?filter=bigbird_pegasus
]


def shift_tokens_right(input_ids: torch.Tensor, pad_token_id: int, decoder_start_token_id: int):
    """
    Shift input ids one token to the right.
    """
    shifted_input_ids = input_ids.new_zeros(input_ids.shape)
    shifted_input_ids[:, 1:] = input_ids[:, :-1].clone()
    shifted_input_ids[:, 0] = decoder_start_token_id

    if pad_token_id is None:
        raise ValueError("self.model.config.pad_token_id has to be defined.")
    # replace possible -100 values in labels by `pad_token_id`
    shifted_input_ids.masked_fill_(shifted_input_ids == -100, pad_token_id)

    return shifted_input_ids


# Copied from transformers.models.bart.modeling_bart._make_causal_mask
def _make_causal_mask(
    input_ids_shape: torch.Size, dtype: torch.dtype, device: torch.device, 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), torch.finfo(dtype).min, device=device)
    mask_cond = torch.arange(mask.size(-1), device=device)
    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, device=device), 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 BigBirdPegasusLearnedPositionalEmbedding(nn.Embedding):
    """
    This module learns positional embeddings up to a fixed maximum size.
    """

    def __init__(self, num_embeddings: int, embedding_dim: int):
        super().__init__(num_embeddings, 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)


# Copied from transformers.models.big_bird.modeling_big_bird.BigBirdSelfAttention with BigBird->BigBirdPegasus
class BigBirdPegasusSelfAttention(nn.Module):
    def __init__(self, config):
        super().__init__()
        if config.hidden_size % config.num_attention_heads != 0 and not hasattr(config, "embedding_size"):
            raise ValueError(
                f"The hidden size ({config.hidden_size}) is not a multiple of the number of attention "
                f"heads ({config.num_attention_heads})"
            )

        self.num_attention_heads = config.num_attention_heads
        self.attention_head_size = int(config.hidden_size / config.num_attention_heads)
        self.all_head_size = self.num_attention_heads * self.attention_head_size

        self.query = nn.Linear(config.hidden_size, self.all_head_size, bias=config.use_bias)
        self.key = nn.Linear(config.hidden_size, self.all_head_size, bias=config.use_bias)
        self.value = nn.Linear(config.hidden_size, self.all_head_size, bias=config.use_bias)

        self.dropout = nn.Dropout(config.attention_probs_dropout_prob)
        self.is_decoder = config.is_decoder

    def transpose_for_scores(self, x):
        new_x_shape = x.size()[:-1] + (self.num_attention_heads, self.attention_head_size)
        x = x.view(*new_x_shape)
        return x.permute(0, 2, 1, 3)

    def forward(
        self,
        hidden_states,
        attention_mask=None,
        head_mask=None,
        encoder_hidden_states=None,
        encoder_attention_mask=None,
        past_key_value=None,
        output_attentions=False,
    ):
        mixed_query_layer = self.query(hidden_states)

        # If this is instantiated as a cross-attention module, the keys
        # and values come from an encoder; the attention mask needs to be
        # such that the encoder's padding tokens are not attended to.
        is_cross_attention = encoder_hidden_states is not None

        if is_cross_attention and past_key_value is not None:
            # reuse k,v, cross_attentions
            key_layer = past_key_value[0]
            value_layer = past_key_value[1]
            attention_mask = encoder_attention_mask
        elif is_cross_attention:
            key_layer = self.transpose_for_scores(self.key(encoder_hidden_states))
            value_layer = self.transpose_for_scores(self.value(encoder_hidden_states))
            attention_mask = encoder_attention_mask
        elif past_key_value is not None:
            key_layer = self.transpose_for_scores(self.key(hidden_states))
            value_layer = self.transpose_for_scores(self.value(hidden_states))
            key_layer = torch.cat([past_key_value[0], key_layer], dim=2)
            value_layer = torch.cat([past_key_value[1], value_layer], dim=2)
        else:
            key_layer = self.transpose_for_scores(self.key(hidden_states))
            value_layer = self.transpose_for_scores(self.value(hidden_states))

        query_layer = self.transpose_for_scores(mixed_query_layer)

        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_layer, value_layer)

        # Take the dot product between "query" and "key" to get the raw attention scores.
        attention_scores = torch.matmul(query_layer, key_layer.transpose(-1, -2))

        attention_scores = attention_scores / math.sqrt(self.attention_head_size)
        if attention_mask is not None:
            # Apply the attention mask is (precomputed for all layers in BigBirdPegasusModel forward() function)
            attention_scores = attention_scores + attention_mask

        # Normalize the attention scores to probabilities.
        attention_probs = nn.functional.softmax(attention_scores, dim=-1)

        # This is actually dropping out entire tokens to attend to, which might
        # seem a bit unusual, but is taken from the original Transformer paper.
        attention_probs = self.dropout(attention_probs)

        # Mask heads if we want to
        if head_mask is not None:
            attention_probs = attention_probs * head_mask

        context_layer = torch.matmul(attention_probs, value_layer)

        context_layer = context_layer.permute(0, 2, 1, 3).contiguous()
        new_context_layer_shape = context_layer.size()[:-2] + (self.all_head_size,)
        context_layer = context_layer.view(*new_context_layer_shape)

        outputs = (context_layer, attention_probs) if output_attentions else (context_layer,)

        if self.is_decoder:
            outputs = outputs + (past_key_value,)
        return outputs


# Copied from transformers.models.big_bird.modeling_big_bird.BigBirdBlockSparseAttention with BigBird->BigBirdPegasus
class BigBirdPegasusBlockSparseAttention(nn.Module):
    def __init__(self, config, seed=None):
        super().__init__()

        self.max_seqlen = config.max_position_embeddings
        self.seed = seed

        if config.hidden_size % config.num_attention_heads != 0:
            raise ValueError(
                f"The hidden size {config.hidden_size} is not a multiple of the number of attention "
                f"heads {config.num_attention_heads}."
            )

        self.num_attention_heads = config.num_attention_heads
        self.num_random_blocks = config.num_random_blocks
        self.block_size = config.block_size

        self.attention_head_size = int(config.hidden_size / config.num_attention_heads)
        self.all_head_size = self.num_attention_heads * self.attention_head_size

        self.query = nn.Linear(config.hidden_size, self.all_head_size, bias=config.use_bias)
        self.key = nn.Linear(config.hidden_size, self.all_head_size, bias=config.use_bias)
        self.value = nn.Linear(config.hidden_size, self.all_head_size, bias=config.use_bias)

    def transpose_for_scores(self, x):
        new_x_shape = x.size()[:-1] + (self.num_attention_heads, self.attention_head_size)
        x = x.view(*new_x_shape)
        return x.permute(0, 2, 1, 3)

    def forward(
        self,
        hidden_states,
        band_mask=None,
        from_mask=None,
        to_mask=None,
        from_blocked_mask=None,
        to_blocked_mask=None,
        output_attentions=None,
    ):
        # Currently this `class` can't be used in decoder.

        batch_size, seqlen, _ = hidden_states.size()
        to_seq_length = from_seq_length = seqlen
        from_block_size = to_block_size = self.block_size

        if from_seq_length % from_block_size != 0:
            raise ValueError("Query sided sequence length must be multiple of block size")

        if to_seq_length % to_block_size != 0:
            raise ValueError("Key/Value sided sequence length must be multiple of block size")

        query_layer = self.transpose_for_scores(self.query(hidden_states))
        key_layer = self.transpose_for_scores(self.key(hidden_states))
        value_layer = self.transpose_for_scores(self.value(hidden_states))

        context_layer, attention_probs = self.bigbird_block_sparse_attention(
            query_layer,
            key_layer,
            value_layer,
            band_mask,
            from_mask,
            to_mask,
            from_blocked_mask,
            to_blocked_mask,
            self.num_attention_heads,
            self.num_random_blocks,
            self.attention_head_size,
            from_block_size,
            to_block_size,
            batch_size,
            from_seq_length,
            to_seq_length,
            seed=self.seed,
            plan_from_length=None,
            plan_num_rand_blocks=None,
            output_attentions=output_attentions,
        )

        context_layer = context_layer.contiguous().view(batch_size, from_seq_length, -1)

        outputs = (context_layer, attention_probs) if output_attentions else (context_layer,)
        return outputs

    @staticmethod
    def torch_bmm_nd(inp_1, inp_2, ndim=None):
        """Fast nd matrix multiplication"""
        # faster replacement of torch.einsum ("bhqk,bhkd->bhqd")
        return torch.bmm(inp_1.reshape((-1,) + inp_1.shape[-2:]), inp_2.reshape((-1,) + inp_2.shape[-2:])).view(
            inp_1.shape[: ndim - 2] + (inp_1.shape[ndim - 2], inp_2.shape[ndim - 1])
        )

    @staticmethod
    def torch_bmm_nd_transpose(inp_1, inp_2, ndim=None):
        """Fast nd matrix multiplication with transpose"""
        # faster replacement of torch.einsum (bhqd,bhkd->bhqk)
        return torch.bmm(
            inp_1.reshape((-1,) + inp_1.shape[-2:]), inp_2.reshape((-1,) + inp_2.shape[-2:]).transpose(1, 2)
        ).view(inp_1.shape[: ndim - 2] + (inp_1.shape[ndim - 2], inp_2.shape[ndim - 2]))

    def bigbird_block_sparse_attention(
        self,
        query_layer,
        key_layer,
        value_layer,
        band_mask,
        from_mask,
        to_mask,
        from_blocked_mask,
        to_blocked_mask,
        n_heads,
        n_rand_blocks,
        attention_head_size,
        from_block_size,
        to_block_size,
        batch_size,
        from_seq_len,
        to_seq_len,
        seed,
        plan_from_length,
        plan_num_rand_blocks,
        output_attentions,
    ):
        # BigBirdPegasus block-sparse attention as suggested in paper

        # ITC:
        #     global tokens: 2 x block_size
        #     window tokens: 3 x block_size
        #     random tokens: num_rand_tokens x block_size

        # ETC:
        #     global tokens: extra_globals_tokens + 2 x block_size
        #     window tokens: 3 x block_size
        #     random tokens: num_rand_tokens x block_size

        # Note:
        #     1) Currently, ETC is not supported.
        #     2) Window size is fixed to 3 blocks & it can be changed only by
        #     changing `block_size`.
        #     3) Number of global blocks are fixed (2 blocks here) & global tokens can be
        #     controlled only by `block_size`.

        # attention is calculated separately for q[0], q[1], q[2:-2], q[-2], q[-1] in order to use special trick of shifting tokens (for calculating sliding attention)
        # hence following code can be divided into 5 parts.

        if from_seq_len // from_block_size != to_seq_len // to_block_size:
            raise ValueError("Error the number of blocks needs to be same!")

        rsqrt_d = 1 / math.sqrt(attention_head_size)
        bsz = batch_size
        attn_mask_penalty = -10000.0

        # generate random attention and corresponding masks
        np.random.seed(seed)
        if from_seq_len in [1024, 3072, 4096]:  # old plans used in paper
            rand_attn = [
                self._bigbird_block_rand_mask(
                    self.max_seqlen, self.max_seqlen, from_block_size, to_block_size, n_rand_blocks, last_idx=1024
                )[: (from_seq_len // from_block_size - 2)]
                for _ in range(n_heads)
            ]
        else:
            if plan_from_length is None:
                plan_from_length, plan_num_rand_blocks = self._get_rand_attn_plan(
                    from_seq_len, from_block_size, n_rand_blocks
                )

            rand_attn = self._bigbird_block_rand_mask_with_head(
                from_seq_length=from_seq_len,
                to_seq_length=to_seq_len,
                from_block_size=from_block_size,
                to_block_size=to_block_size,
                num_heads=n_heads,
                plan_from_length=plan_from_length,
                plan_num_rand_blocks=plan_num_rand_blocks,
            )

        rand_attn = np.stack(rand_attn, axis=0)
        rand_attn = torch.tensor(rand_attn, device=query_layer.device, dtype=torch.long)
        rand_attn.unsqueeze_(0)
        rand_attn = torch.cat([rand_attn for _ in range(batch_size)], dim=0)

        rand_mask = self._create_rand_mask_from_inputs(
            from_blocked_mask, to_blocked_mask, rand_attn, n_heads, n_rand_blocks, bsz, from_seq_len, from_block_size
        )

        blocked_query_matrix = query_layer.view(bsz, n_heads, from_seq_len // from_block_size, from_block_size, -1)
        blocked_key_matrix = key_layer.view(bsz, n_heads, to_seq_len // to_block_size, to_block_size, -1)
        blocked_value_matrix = value_layer.view(bsz, n_heads, to_seq_len // to_block_size, to_block_size, -1)

        # preparing block for randn attn
        gathered_key = self.torch_gather_b2(blocked_key_matrix, rand_attn)
        gathered_key = gathered_key.view(
            bsz, n_heads, to_seq_len // to_block_size - 2, n_rand_blocks * to_block_size, -1
        )  # [bsz, n_heads, to_seq_len//to_block_size-2, n_rand_blocks, to_block_size, -1]
        gathered_value = self.torch_gather_b2(blocked_value_matrix, rand_attn)
        gathered_value = gathered_value.view(
            bsz, n_heads, to_seq_len // to_block_size - 2, n_rand_blocks * to_block_size, -1
        )  # [bsz, n_heads, to_seq_len//to_block_size-2, n_rand_blocks, to_block_size, -1]

        # 1st PART
        # 1st block (global block) attention scores
        # q[0] x (k[0], k[1], k[2], k[3], k[4] .... )

        # [bsz, n_heads, from_block_size, -1] x [bsz, n_heads, to_seq_len, -1] ==> [bsz, n_heads, from_block_size, to_seq_len]
        first_product = self.torch_bmm_nd_transpose(blocked_query_matrix[:, :, 0], key_layer, ndim=4)

        first_product = first_product * rsqrt_d
        first_product += (1.0 - to_mask) * attn_mask_penalty
        first_attn_weights = nn.functional.softmax(
            first_product, dim=-1
        )  # [bsz, n_heads, from_block_size, to_seq_len]

        # [bsz, n_heads, from_block_size, to_seq_len] x [bsz, n_heads, to_seq_len, -1] ==> [bsz, n_heads, from_block_size, -1]
        first_context_layer = self.torch_bmm_nd(first_attn_weights, value_layer, ndim=4)
        first_context_layer.unsqueeze_(2)

        # 2nd PART
        # 2nd block attention scores
        # q[1] x (sliding_keys, random_keys, global_keys)
        # sliding key blocks -> 2nd, 3rd blocks
        # global key blocks -> 1st block

        second_key_mat = torch.cat(
            [
                blocked_key_matrix[:, :, 0],
                blocked_key_matrix[:, :, 1],
                blocked_key_matrix[:, :, 2],
                blocked_key_matrix[:, :, -1],
                gathered_key[:, :, 0],
            ],
            dim=2,
        )  # [bsz, n_heads, (4+n_rand_blocks)*to_block_size, -1]
        second_value_mat = torch.cat(
            [
                blocked_value_matrix[:, :, 0],
                blocked_value_matrix[:, :, 1],
                blocked_value_matrix[:, :, 2],
                blocked_value_matrix[:, :, -1],
                gathered_value[:, :, 0],
            ],
            dim=2,
        )  # [bsz, n_heads, (4+n_rand_blocks)*to_block_size, -1]

        # [bsz, n_heads, from_block_size, -1] x [bsz, n_heads, (4+n_rand_blocks)*to_block_size, -1] ==> [bsz, n_heads, from_block_size, (4+n_rand_blocks)*to_block_size]
        second_product = self.torch_bmm_nd_transpose(blocked_query_matrix[:, :, 1], second_key_mat, ndim=4)
        second_seq_pad = torch.cat(
            [
                to_mask[:, :, :, : 3 * to_block_size],
                to_mask[:, :, :, -to_block_size:],
                to_mask.new_ones([bsz, 1, 1, n_rand_blocks * to_block_size]),
            ],
            dim=3,
        )
        second_rand_pad = torch.cat(
            [
                rand_mask.new_ones([bsz, n_heads, from_block_size, 4 * to_block_size]),
                rand_mask[:, :, 0],
            ],
            dim=3,
        )
        second_product = second_product * rsqrt_d
        second_product += (1.0 - torch.minimum(second_seq_pad, second_rand_pad)) * attn_mask_penalty
        second_attn_weights = nn.functional.softmax(
            second_product, dim=-1
        )  # [bsz, n_heads, from_block_size, (4+n_rand_blocks)*to_block_size]

        # [bsz, n_heads, from_block_size, (4+n_rand_blocks)*to_block_size] x [bsz, n_heads, (4+n_rand_blocks)*to_block_size, -1] ==> [bsz, n_heads, from_block_size, -1]
        second_context_layer = self.torch_bmm_nd(second_attn_weights, second_value_mat, ndim=4)

        second_context_layer.unsqueeze_(2)

        # 3rd PART
        # Middle blocks attention scores
        # q[-2:2] x (sliding_keys, random_keys, global_keys)
        # sliding attn is calculated using special trick of shifting tokens as discussed in paper
        # random keys are generated by taking random indices as per `rand_attn`
        # global keys -> 1st & last block

        exp_blocked_key_matrix = torch.cat(
            [blocked_key_matrix[:, :, 1:-3], blocked_key_matrix[:, :, 2:-2], blocked_key_matrix[:, :, 3:-1]], dim=3
        )  # [bsz, n_heads, from_seq_len//from_block_size-4, 3*to_block_size, -1]
        exp_blocked_value_matrix = torch.cat(
            [blocked_value_matrix[:, :, 1:-3], blocked_value_matrix[:, :, 2:-2], blocked_value_matrix[:, :, 3:-1]],
            dim=3,
        )  # [bsz, n_heads, from_seq_len//from_block_size-4, 3*to_block_size, -1]
        middle_query_matrix = blocked_query_matrix[:, :, 2:-2]

        # sliding attention scores for q[-2:2]
        # [bsz, n_heads, from_seq_len//from_block_size-4, from_block_size, -1] x [b, n_heads, from_seq_len//from_block_size-4, 3*to_block_size, -1]
        inner_band_product = self.torch_bmm_nd_transpose(middle_query_matrix, exp_blocked_key_matrix, ndim=5)
        #     ==> [bsz, n_heads, from_seq_len//from_block_size-4, from_block_size, 3*to_block_size]
        inner_band_product = inner_band_product * rsqrt_d

        # randn attention scores for q[-2:2]
        # [bsz, n_heads, from_seq_len//from_block_size-4, from_block_size, -1] x [bsz, n_heads, from_seq_len//from_block_size-4, n_rand_blocks*to_block_size, -1]
        rand_band_product = self.torch_bmm_nd_transpose(middle_query_matrix, gathered_key[:, :, 1:-1], ndim=5)
        #     ==> [bsz, n_heads, from_seq_len//from_block_size-4, from_block_size, n_rand_blocks*to_block_size]
        rand_band_product = rand_band_product * rsqrt_d

        # Including 1st block (since it's global)
        first_band_product = torch.einsum(
            "bhlqd,bhkd->bhlqk", middle_query_matrix, blocked_key_matrix[:, :, 0]
        )  # [bsz, n_heads, from_seq_len//from_block_size-4, from_block_size, -1] x [bsz, n_heads, to_block_size, -1] ==> [bsz, n_heads, from_seq_len//from_block_size-4, from_block_size, to_block_size]
        first_band_product = first_band_product * rsqrt_d

        # Including last block (since it's global)
        last_band_product = torch.einsum(
            "bhlqd,bhkd->bhlqk", middle_query_matrix, blocked_key_matrix[:, :, -1]
        )  # [bsz, n_heads, from_seq_len//from_block_size-4, from_block_size, -1] x [bsz, n_heads, to_block_size, -1] ==> [bsz, n_heads, from_seq_len//from_block_size-4, from_block_size, to_block_size]
        last_band_product = last_band_product * rsqrt_d

        # masking padded tokens
        inner_band_product += (1.0 - band_mask) * attn_mask_penalty
        first_band_product += (1.0 - to_mask[:, :, :, :to_block_size].unsqueeze(3)) * attn_mask_penalty
        last_band_product += (1.0 - to_mask[:, :, :, -to_block_size:].unsqueeze(3)) * attn_mask_penalty
        rand_band_product += (1.0 - rand_mask[:, :, 1:-1]) * attn_mask_penalty

        # completing attention scores matrix for all q[-2:2]
        band_product = torch.cat(
            [first_band_product, inner_band_product, rand_band_product, last_band_product], dim=-1
        )  # [bsz, n_heads, from_seq_len//from_block_size-4, from_block_size, (5+n_rand_blocks)*to_block_size]

        # safely doing softmax since attention matrix is completed
        attn_weights = nn.functional.softmax(
            band_product, dim=-1
        )  # [bsz, n_heads, from_seq_len//from_block_size-4, from_block_size, (5+n_rand_blocks)*to_block_size]

        # contribution of sliding keys
        # [bsz, n_heads, m//from_block_size-4, from_block_size, 3*to_block_size] x [bsz, n_heads, from_seq_len//from_block_size-4, 3*to_block_size, -1]
        context_layer = self.torch_bmm_nd(
            attn_weights[:, :, :, :, to_block_size : 4 * to_block_size], exp_blocked_value_matrix, ndim=5
        )
        #     ==> [bsz, n_heads, from_seq_len//from_block_size-4, from_block_size, -1]

        # adding contribution of random keys
        # [bsz, n_heads, from_seq_len//from_block_size-4, from_block_size, n_rand_blocks*to_block_size] x [bsz, n_heads, from_seq_len//from_block_size-4, n_rand_blocks*to_block_size, -1]
        context_layer += self.torch_bmm_nd(
            attn_weights[:, :, :, :, 4 * to_block_size : -to_block_size], gathered_value[:, :, 1:-1], ndim=5
        )
        #     ==> [bsz, n_heads, from_seq_len//from_block_size-4, from_block_size, -1]

        # adding contribution of global keys
        context_layer += torch.einsum(
            "bhlqk,bhkd->bhlqd", attn_weights[:, :, :, :, :to_block_size], blocked_value_matrix[:, :, 0]
        )  # [bsz, n_heads, from_seq_len//from_block_size-4, from_block_size, to_block_size] x [bsz, n_heads, to_block_size, -1] ==> [bsz, n_heads, from_seq_len//from_block_size-4, from_block_size, -1]
        context_layer += torch.einsum(
            "bhlqk,bhkd->bhlqd", attn_weights[:, :, :, :, -to_block_size:], blocked_value_matrix[:, :, -1]
        )  # [bsz, n_heads, from_seq_len//from_block_size-4, from_block_size, to_block_size] x [bsz, n_heads, to_block_size, -1] ==> [bsz, n_heads, from_seq_len//from_block_size-4, from_block_size, -1]

        # 4th PART
        # last 2nd token attention scores
        # q[-2] x (sliding_keys, random_keys, global_keys)
        # sliding key blocks -> last 3 blocks
        # global key block -> 1st block
        # random key block -> based on indices stored in `randn_attn`

        second_last_key_mat = torch.cat(
            [
                blocked_key_matrix[:, :, 0],
                blocked_key_matrix[:, :, -3],
                blocked_key_matrix[:, :, -2],
                blocked_key_matrix[:, :, -1],
                gathered_key[:, :, -1],
            ],
            dim=2,
        )  # [bsz, n_heads, (4+n_random_blocks)*to_block_size, -1]
        second_last_value_mat = torch.cat(
            [
                blocked_value_matrix[:, :, 0],
                blocked_value_matrix[:, :, -3],
                blocked_value_matrix[:, :, -2],
                blocked_value_matrix[:, :, -1],
                gathered_value[:, :, -1],
            ],
            dim=2,
        )  # [bsz, n_heads, (4+r)*to_block_size, -1]

        # [bsz, n_heads, from_block_size, -1] x [bsz, n_heads, (4+n_rand_blocks)*to_block_size, -1] ==> [bsz, n_heads, from_block_size, (4+n_rand_blocks)*to_block_size]
        second_last_product = self.torch_bmm_nd_transpose(blocked_query_matrix[:, :, -2], second_last_key_mat, ndim=4)
        second_last_seq_pad = torch.cat(
            [
                to_mask[:, :, :, :to_block_size],
                to_mask[:, :, :, -3 * to_block_size :],
                to_mask.new_ones([bsz, 1, 1, n_rand_blocks * to_block_size]),
            ],
            dim=3,
        )
        second_last_rand_pad = torch.cat(
            [
                rand_mask.new_ones([bsz, n_heads, from_block_size, 4 * to_block_size]),
                rand_mask[:, :, -1],
            ],
            dim=3,
        )
        second_last_product = second_last_product * rsqrt_d
        second_last_product += (1.0 - torch.minimum(second_last_seq_pad, second_last_rand_pad)) * attn_mask_penalty
        second_last_attn_weights = nn.functional.softmax(
            second_last_product, dim=-1
        )  # [bsz, n_heads, from_block_size, (4+n_rand_blocks)*to_block_size]

        # [bsz, n_heads, from_block_size, (4+n_rand_blocks)*to_block_size] x [bsz, n_heads, (4+n_rand_blocks)*to_block_size, -1] ==> [bsz, n_heads, from_block_size, -1]
        second_last_context_layer = self.torch_bmm_nd(second_last_attn_weights, second_last_value_mat, ndim=4)
        second_last_context_layer.unsqueeze_(2)

        # 5th PART
        # last block (global) attention scores
        # q[-1] x (k[0], k[1], k[2], k[3], .... )

        # [bsz, n_heads, from_block_size, -1] x [bsz, n_heads, to_seq_len, -1] ==> [bsz, n_heads, from_block_size, to_seq_len]
        last_product = self.torch_bmm_nd_transpose(blocked_query_matrix[:, :, -1], key_layer, ndim=4)
        last_product = last_product * rsqrt_d
        last_product += (1.0 - to_mask) * attn_mask_penalty
        last_attn_weights = nn.functional.softmax(last_product, dim=-1)  # [bsz, n_heads, from_block_size, n]

        # [bsz, n_heads, from_block_size, to_seq_len] x [bsz, n_heads, to_seq_len, -1] ==> [bsz, n_heads, from_block_size, -1]
        last_context_layer = self.torch_bmm_nd(last_attn_weights, value_layer, ndim=4)
        last_context_layer.unsqueeze_(2)

        # combining representations of all tokens
        context_layer = torch.cat(
            [first_context_layer, second_context_layer, context_layer, second_last_context_layer, last_context_layer],
            dim=2,
        )
        context_layer = context_layer.view((bsz, n_heads, from_seq_len, -1)) * from_mask
        context_layer = torch.transpose(context_layer, 1, 2)

        # this is just for visualizing; forward pass doesn't depend on following code
        if output_attentions:
            # TODO(PVP): need to verify if below code is correct
            attention_probs = torch.zeros(
                bsz, n_heads, from_seq_len, to_seq_len, dtype=torch.float, device=context_layer.device
            )

            # 1st query block
            # corresponding to `first_context_layer`
            attention_probs[:, :, :from_block_size, :] = first_attn_weights  # all keys global

            # 2nd query block
            # corresponding to `second_context_layer`
            attention_probs[:, :, from_block_size : 2 * from_block_size, : 3 * to_block_size] = second_attn_weights[
                :, :, :, : 3 * to_block_size
            ]  # 1st three key blocks (global + sliding)
            attention_probs[:, :, from_block_size : 2 * from_block_size, -to_block_size:] = second_attn_weights[
                :, :, :, 3 * to_block_size : 4 * to_block_size
            ]  # last key block (global)
            # random keys
            for p1, i1, w1 in zip(range(bsz), rand_attn, second_attn_weights):
                # p1, i1, w1 corresponds to batch_dim i.e. following operation is done for each sequence in batch
                for p2, i2, w2 in zip(range(n_heads), i1, w1):
                    # p2, i2, w2 corresponds to head_dim i.e. following operation is done for each heads
                    attn_probs_view = attention_probs.view(
                        bsz,
                        n_heads,
                        from_seq_len // from_block_size,
                        from_block_size,
                        to_seq_len // to_block_size,
                        to_block_size,
                    )
                    right_slice = w2[:, 4 * to_block_size :]
                    attn_probs_view[p1, p2, 1, :, i2[0]] = right_slice.view(
                        from_block_size, n_rand_blocks, to_block_size
                    )

            # Middle query blocks
            # corresponding to `context_layer`
            # sliding keys
            for q_idx in range(from_seq_len // from_block_size - 4):
                attn_probs_view = attention_probs.view(
                    bsz,
                    n_heads,
                    from_seq_len // from_block_size,
                    from_block_size,
                    to_seq_len // to_block_size,
                    to_block_size,
                )[:, :, 2:-2, :, 1:-1, :]
                right_slice = attn_weights[:, :, q_idx, :, to_block_size : 4 * to_block_size]
                attn_probs_view[:, :, q_idx, :, q_idx : q_idx + 3, :] = right_slice.view(
                    bsz, n_heads, from_block_size, 3, to_block_size
                )  # inner_band_product
            # global keys (corresponding to 1st key block)
            attention_probs[:, :, 2 * from_block_size : -2 * from_block_size, :to_block_size] = attn_weights[
                :, :, :, :, :to_block_size
            ].view(
                bsz, n_heads, -1, to_block_size
            )  # first_band_product
            # global keys (corresponding to last key block)
            attention_probs[:, :, 2 * from_block_size : -2 * from_block_size, -to_block_size:] = attn_weights[
                :, :, :, :, -to_block_size:
            ].view(
                bsz, n_heads, -1, to_block_size
            )  # last_band_product
            # random keys
            for p1, i1, w1 in zip(range(bsz), rand_attn, attn_weights):
                # p1, i1, w1 corresponds to batch_dim i.e. following operation is done for each sequence in batch
                for p2, i2, w2 in zip(range(n_heads), i1, w1):
                    # p2, i2, w2 corresponds to head_dim i.e. following operation is done for each heads
                    for q_idx in range(1, len(i2) - 1):
                        attn_probs_view = attention_probs.view(
                            bsz,
                            n_heads,
                            from_seq_len // from_block_size,
                            from_block_size,
                            to_seq_len // to_block_size,
                            to_block_size,
                        )
                        right_slice = w2[q_idx - 1, :, 4 * to_block_size : -to_block_size]
                        attn_probs_view[p1, p2, q_idx + 1, :, i2[q_idx]] = right_slice.view(
                            from_block_size, n_rand_blocks, to_block_size
                        )

            # Second-last query block
            # corresponding to `second_last_context_layer`
            attention_probs[:, :, -2 * from_block_size : -from_block_size, :to_block_size] = second_last_attn_weights[
                :, :, :, :to_block_size
            ]  # 1st key block (global)
            attention_probs[
                :, :, -2 * from_block_size : -from_block_size, -3 * to_block_size :
            ] = second_last_attn_weights[
                :, :, :, to_block_size : 4 * to_block_size
            ]  # last three blocks (global + sliding)
            # random keys
            for p1, i1, w1 in zip(range(bsz), rand_attn, second_last_attn_weights):
                # p1, i1, w1 corresponds to batch_dim i.e. following operation is done for each sequence in batch
                for p2, i2, w2 in zip(range(n_heads), i1, w1):
                    # p2, i2, w2 corresponds to head_dim i.e. following operation is done for each heads
                    attn_probs_view = attention_probs.view(
                        bsz,
                        n_heads,
                        from_seq_len // from_block_size,
                        from_block_size,
                        to_seq_len // to_block_size,
                        to_block_size,
                    )
                    right_slice = w2[:, 4 * to_block_size :]
                    attn_probs_view[p1, p2, -2, :, i2[-1]] = right_slice.view(
                        from_block_size, n_rand_blocks, to_block_size
                    )

            # last query block
            # corresponding to `last_context_layer`
            attention_probs[:, :, -from_block_size:, :] = last_attn_weights  # all keys global

        else:
            attention_probs = None

        return context_layer, attention_probs

    @staticmethod
    def torch_gather_b2(params, indices):
        # this operation is equivalent to tf.gather when batch_dims=2

        if params.shape[:2] != indices.shape[:2]:
            raise ValueError(
                "Make sure that the first two dimensions of params and indices are identical,                 but"
                f" they are params: {params.shape[:2]} vs. indices: {indices.shape[:2]}"
            )
        num_indices_to_gather = indices.shape[-2] * indices.shape[-1]
        num_indices_to_pick_from = params.shape[2]

        shift = torch.arange(indices.shape[0] * indices.shape[1] * num_indices_to_gather, device=indices.device)
        indices_shift = torch.div(shift, num_indices_to_gather, rounding_mode="floor") * num_indices_to_pick_from

        flattened_indices = indices.view(-1) + indices_shift
        flattened_params = params.reshape(-1, params.shape[-2], params.shape[-1])

        out_flattened = flattened_params.index_select(0, flattened_indices)

        out = out_flattened.reshape(params.shape[:2] + (num_indices_to_gather,) + params.shape[3:])
        return out

    @staticmethod
    def _create_rand_mask_from_inputs(
        from_blocked_mask,
        to_blocked_mask,
        rand_attn,
        num_attention_heads,
        num_rand_blocks,
        batch_size,
        from_seq_length,
        from_block_size,
    ):
        """
        Create 3D attention mask from a 2D tensor mask.

        Args:
            from_blocked_mask: 2D Tensor of shape [batch_size,
            from_seq_length//from_block_size, from_block_size].
            to_blocked_mask: int32 Tensor of shape [batch_size,
            to_seq_length//to_block_size, to_block_size].
            rand_attn: [batch_size, num_attention_heads,
            from_seq_length//from_block_size-2, num_rand_blocks]
            num_attention_heads: int. Number of attention heads.
            num_rand_blocks: int. Number of random chunks per row.
            batch_size: int. Batch size for computation.
            from_seq_length: int. length of from sequence.
            from_block_size: int. size of block in from sequence.

        Returns:
            float Tensor of shape [batch_size, num_attention_heads, from_seq_length//from_block_size-2,
            from_block_size, num_rand_blocks*to_block_size].
        """
        num_windows = from_seq_length // from_block_size - 2
        rand_mask = torch.stack([p1[i1.flatten()] for p1, i1 in zip(to_blocked_mask, rand_attn)])
        rand_mask = rand_mask.view(batch_size, num_attention_heads, num_windows, num_rand_blocks * from_block_size)
        rand_mask = torch.einsum("blq,bhlk->bhlqk", from_blocked_mask[:, 1:-1], rand_mask)
        return rand_mask

    @staticmethod
    def _get_rand_attn_plan(from_seq_length, from_block_size, num_rand_blocks):
        """
        Gives the plan of where to put random attention.

        Args:
            from_seq_length: int. length of from sequence.
            from_block_size: int. size of block in from sequence.
            num_rand_blocks: int. Number of random chunks per row.

        Returns:
            plan_from_length: ending location of from block plan_num_rand_blocks: number of random ending location for
            each block
        """

        plan_from_length = []
        plan_num_rand_blocks = []
        if (2 * num_rand_blocks + 5) < (from_seq_length // from_block_size):
            plan_from_length.append(int((2 * num_rand_blocks + 5) * from_block_size))
            plan_num_rand_blocks.append(num_rand_blocks)
            plan_from_length.append(from_seq_length)
            plan_num_rand_blocks.append(0)
        elif (num_rand_blocks + 5) < (from_seq_length // from_block_size):
            plan_from_length.append(int((num_rand_blocks + 5) * from_block_size))
            plan_num_rand_blocks.append(num_rand_blocks // 2)
            plan_from_length.append(from_seq_length)
            plan_num_rand_blocks.append(num_rand_blocks - (num_rand_blocks // 2))
        else:
            plan_from_length.append(from_seq_length)
            plan_num_rand_blocks.append(num_rand_blocks)

        return plan_from_length, plan_num_rand_blocks

    def _bigbird_block_rand_mask(
        self, from_seq_length, to_seq_length, from_block_size, to_block_size, num_rand_blocks, last_idx=-1
    ):
        """
        Create adjacency list of random attention.

        Args:
            from_seq_length: int. length of from sequence.
            to_seq_length: int. length of to sequence.
            from_block_size: int. size of block in from sequence.
            to_block_size: int. size of block in to sequence.
            num_rand_blocks: int. Number of random chunks per row.
            last_idx: if -1 then num_rand_blocks blocks chosen anywhere in to sequence,
            if positive then num_rand_blocks blocks chosen only up to last_idx.

        Returns:
            adjacency list of size from_seq_length//from_block_size-2 by num_rand_blocks
        """
        # using this method when from_seq_length in [1024, 3072, 4096]

        if from_seq_length // from_block_size != to_seq_length // to_block_size:
            raise ValueError("Error the number of blocks needs to be same!")

        rand_attn = np.zeros((from_seq_length // from_block_size - 2, num_rand_blocks), dtype=np.int32)
        # During inference (eval) no randomness
        if not self.training:
            return rand_attn
        middle_seq = np.arange(1, to_seq_length // to_block_size - 1, dtype=np.int32)
        last = to_seq_length // to_block_size - 1
        if last_idx > (2 * to_block_size):
            last = (last_idx // to_block_size) - 1

        r = num_rand_blocks  # shorthand
        for i in range(1, from_seq_length // from_block_size - 1):
            start = i - 2
            end = i
            if i == 1:
                rand_attn[i - 1, :] = np.random.permutation(middle_seq[2:last])[:r]
            elif i == 2:
                rand_attn[i - 1, :] = np.random.permutation(middle_seq[3:last])[:r]
            elif i == from_seq_length // from_block_size - 3:
                rand_attn[i - 1, :] = np.random.permutation(middle_seq[:last])[:r]
            # Missing -3: should have been sliced till last-3
            elif i == from_seq_length // from_block_size - 2:
                rand_attn[i - 1, :] = np.random.permutation(middle_seq[:last])[:r]
            # Missing -4: should have been sliced till last-4
            else:
                if start > last:
                    start = last
                    rand_attn[i - 1, :] = np.random.permutation(middle_seq[:start])[:r]
                elif (end + 1) == last:
                    rand_attn[i - 1, :] = np.random.permutation(middle_seq[:start])[:r]
                else:
                    rand_attn[i - 1, :] = np.random.permutation(
                        np.concatenate((middle_seq[:start], middle_seq[end + 1 : last]))
                    )[:r]
        return rand_attn

    def _bigbird_block_rand_mask_with_head(
        self,
        from_seq_length,
        to_seq_length,
        from_block_size,
        to_block_size,
        num_heads,
        plan_from_length,
        plan_num_rand_blocks,
        window_block_left=1,
        window_block_right=1,
        global_block_top=1,
        global_block_bottom=1,
        global_block_left=1,
        global_block_right=1,
    ):
        """
        Create adjacency list of random attention.

        Args:
            from_seq_length: int. length of from sequence.
            to_seq_length: int. length of to sequence.
            from_block_size: int. size of block in from sequence.
            to_block_size: int. size of block in to sequence.
            num_heads: int. total number of heads.
            plan_from_length: list. plan from length where num_random_blocks are chosen from.
            plan_num_rand_blocks: list. number of rand blocks within the plan.
            window_block_left: int. number of blocks of window to left of a block.
            window_block_right: int. number of blocks of window to right of a block.
            global_block_top: int. number of blocks at the top.
            global_block_bottom: int. number of blocks at the bottom.
            global_block_left: int. Number of blocks globally used to the left.
            global_block_right: int. Number of blocks globally used to the right.

        Returns:
            adjacency list of size num_head where each element is of size from_seq_length//from_block_size-2 by
            num_rand_blocks
        """
        # using this method when from_seq_length not in [1024, 3072, 4096]

        if from_seq_length // from_block_size != to_seq_length // to_block_size:
            raise ValueError("Error the number of blocks needs to be same!")

        if from_seq_length not in plan_from_length:
            raise ValueError("Error from sequence length not in plan!")

        # Total number of blocks in the mmask
        num_blocks = from_seq_length // from_block_size
        # Number of blocks per plan
        plan_block_length = np.array(plan_from_length) // from_block_size
        # till when to follow plan
        max_plan_idx = plan_from_length.index(from_seq_length)

        # Random Attention adjacency list
        rand_attn = [
            np.zeros((num_blocks, np.sum(plan_num_rand_blocks[: max_plan_idx + 1])), dtype=np.int32)
            for i in range(num_heads)
        ]
        # During inference (eval) no randomness
        if not self.training:
            for nh in range(num_heads):
                rand_attn[nh] = rand_attn[nh][global_block_top : num_blocks - global_block_bottom, :]
            return rand_attn

        # We will go iteratively over the plan blocks and pick random number of
        # Attention blocks from the legally allowed blocks
        for plan_idx in range(max_plan_idx + 1):
            rnd_r_cnt = 0
            if plan_idx > 0:
                # set the row for all from_blocks starting from 0 to
                # plan_block_length[plan_idx-1]
                # column indx start fromm plan_block_length[plan_idx-1] and ends at
                # plan_block_length[plan_idx]
                if plan_num_rand_blocks[plan_idx] > 0:
                    rnd_r_cnt = int(np.sum(plan_num_rand_blocks[:plan_idx]))
                    curr_r_cnt = int(np.sum(plan_num_rand_blocks[: plan_idx + 1]))
                    for blk_rw_idx in range(global_block_top, plan_block_length[plan_idx - 1]):
                        for h in range(num_heads):
                            rand_attn[h][blk_rw_idx, rnd_r_cnt:curr_r_cnt] = self._get_single_block_row_attention(
                                block_id=blk_rw_idx,
                                to_start_block_id=plan_block_length[plan_idx - 1],
                                to_end_block_id=plan_block_length[plan_idx],
                                num_rand_blocks=plan_num_rand_blocks[plan_idx],
                                window_block_left=window_block_left,
                                window_block_right=window_block_right,
                                global_block_left=global_block_left,
                                global_block_right=global_block_right,
                            )

                for pl_id in range(plan_idx):
                    if plan_num_rand_blocks[pl_id] == 0:
                        continue
                    for blk_rw_idx in range(plan_block_length[plan_idx - 1], plan_block_length[plan_idx]):
                        rnd_r_cnt = 0
                        to_start_block_id = 0
                        if pl_id > 0:
                            rnd_r_cnt = int(np.sum(plan_num_rand_blocks[:pl_id]))
                            to_start_block_id = plan_block_length[pl_id - 1]
                        curr_r_cnt = int(np.sum(plan_num_rand_blocks[: pl_id + 1]))
                        for h in range(num_heads):
                            rand_attn[h][blk_rw_idx, rnd_r_cnt:curr_r_cnt] = self._get_single_block_row_attention(
                                block_id=blk_rw_idx,
                                to_start_block_id=to_start_block_id,
                                to_end_block_id=plan_block_length[pl_id],
                                num_rand_blocks=plan_num_rand_blocks[pl_id],
                                window_block_left=window_block_left,
                                window_block_right=window_block_right,
                                global_block_left=global_block_left,
                                global_block_right=global_block_right,
                            )

            if plan_num_rand_blocks[plan_idx] == 0:
                continue
            curr_r_cnt = int(np.sum(plan_num_rand_blocks[: plan_idx + 1]))
            from_start_block_id = global_block_top
            to_start_block_id = 0
            if plan_idx > 0:
                rnd_r_cnt = int(np.sum(plan_num_rand_blocks[:plan_idx]))
                from_start_block_id = plan_block_length[plan_idx - 1]
                to_start_block_id = plan_block_length[plan_idx - 1]

            for blk_rw_idx in range(from_start_block_id, plan_block_length[plan_idx]):
                for h in range(num_heads):
                    rand_attn[h][blk_rw_idx, rnd_r_cnt:curr_r_cnt] = self._get_single_block_row_attention(
                        block_id=blk_rw_idx,
                        to_start_block_id=to_start_block_id,
                        to_end_block_id=plan_block_length[plan_idx],
                        num_rand_blocks=plan_num_rand_blocks[plan_idx],
                        window_block_left=window_block_left,
                        window_block_right=window_block_right,
                        global_block_left=global_block_left,
                        global_block_right=global_block_right,
                    )

        for nh in range(num_heads):
            rand_attn[nh] = rand_attn[nh][global_block_top : num_blocks - global_block_bottom, :]

        return rand_attn

    @staticmethod
    def _get_single_block_row_attention(
        block_id,
        to_start_block_id,
        to_end_block_id,
        num_rand_blocks,
        window_block_left=1,
        window_block_right=1,
        global_block_left=1,
        global_block_right=1,
    ):
        """
        For a single row block get random row attention.

        Args:
            block_id: int. block id of row.
            to_start_block_id: int. random attention column start id.
            to_end_block_id: int. random attention column end id.
            num_rand_blocks: int. number of random blocks to be selected.
            window_block_left: int. number of blocks of window to left of a block.
            window_block_right: int. number of blocks of window to right of a block.
            global_block_left: int. Number of blocks globally used to the left.
            global_block_right: int. Number of blocks globally used to the right.

        Returns:
            row containing the random attention vector of size num_rand_blocks.
        """
        # list of to_blocks from which to choose random attention
        to_block_list = np.arange(to_start_block_id, to_end_block_id, dtype=np.int32)
        # permute the blocks
        perm_block = np.random.permutation(to_block_list)

        # illegal blocks for the current block id, using window
        illegal_blocks = list(range(block_id - window_block_left, block_id + window_block_right + 1))

        # Add blocks at the start and at the end
        illegal_blocks.extend(list(range(global_block_left)))
        illegal_blocks.extend(list(range(to_end_block_id - global_block_right, to_end_block_id)))

        # The second from_block cannot choose random attention on second last to_block
        if block_id == 1:
            illegal_blocks.append(to_end_block_id - 2)

        # The second last from_block cannot choose random attention on second to_block
        if block_id == to_end_block_id - 2:
            illegal_blocks.append(1)

        selected_random_blokcs = []

        for i in range(to_end_block_id - to_start_block_id):
            if perm_block[i] not in illegal_blocks:
                selected_random_blokcs.append(perm_block[i])
            if len(selected_random_blokcs) == num_rand_blocks:
                break
        return np.array(selected_random_blokcs, dtype=np.int32)


class BigBirdPegasusEncoderAttention(nn.Module):
    def __init__(self, config, seed=None):
        super().__init__()
        self.config = config
        self.seed = seed

        self.attention_type = config.attention_type

        if self.attention_type == "original_full":
            self.self = BigBirdPegasusSelfAttention(config)
        elif self.attention_type == "block_sparse":
            self.self = BigBirdPegasusBlockSparseAttention(config, seed)
        else:
            raise ValueError(
                f"attention_type can either be original_full or block_sparse, but is {self.config.attention_type}"
            )

        self.output = nn.Linear(config.hidden_size, config.hidden_size, bias=config.use_bias)

    def set_attention_type(self, value: str):
        if value not in ["original_full", "block_sparse"]:
            raise ValueError(
                f"attention_type can only be set to either 'original_full' or 'block_sparse', but is {value}"
            )
        # attention type is already correctly set
        if value == self.attention_type:
            return

        self.attention_type = value
        if value == "original_full":
            # copy all weights to new full attention class
            attn_weights = BigBirdPegasusSelfAttention(self.config)
        else:
            # copy all weights to new sparse attention class
            attn_weights = BigBirdPegasusBlockSparseAttention(self.config, self.seed)

        attn_weights.query = self.self.query
        attn_weights.value = self.self.value
        attn_weights.key = self.self.key
        self.self = attn_weights
        self.attention_type = value

        if not self.training:
            self.self.eval()

    def forward(
        self,
        hidden_states,
        attention_mask=None,
        head_mask=None,
        past_key_value=None,
        output_attentions=False,
        band_mask=None,
        from_mask=None,
        to_mask=None,
        from_blocked_mask=None,
        to_blocked_mask=None,
    ):
        # Expand dims to enable multiplication in the self-attention module
        head_mask = head_mask.reshape(1, -1, 1, 1) if head_mask is not None else None

        if self.config.attention_type == "original_full":
            self_outputs = self.self(
                hidden_states,
                attention_mask,
                head_mask,
                past_key_value=past_key_value,
                output_attentions=output_attentions,
            )
        else:
            self_outputs = self.self(
                hidden_states, band_mask, from_mask, to_mask, from_blocked_mask, to_blocked_mask, output_attentions
            )

        attention_output = self.output(self_outputs[0])
        outputs = (attention_output,) + self_outputs[1:]  # add attentions if we output them
        return outputs


# Copied from transformers.models.bart.modeling_bart.BartAttention with Bart->BigBirdPegasusDecoder
class BigBirdPegasusDecoderAttention(nn.Module):
    """Multi-headed attention from 'Attention Is All You Need' paper"""

    def __init__(
        self,
        embed_dim: int,
        num_heads: int,
        dropout: float = 0.0,
        is_decoder: bool = False,
        bias: bool = True,
    ):
        super().__init__()
        self.embed_dim = embed_dim
        self.num_heads = num_heads
        self.dropout = dropout
        self.head_dim = embed_dim // num_heads

        if (self.head_dim * num_heads) != self.embed_dim:
            raise ValueError(
                f"embed_dim must be divisible by num_heads (got `embed_dim`: {self.embed_dim}"
                f" and `num_heads`: {num_heads})."
            )
        self.scaling = self.head_dim**-0.5
        self.is_decoder = is_decoder

        self.k_proj = nn.Linear(embed_dim, embed_dim, bias=bias)
        self.v_proj = nn.Linear(embed_dim, embed_dim, bias=bias)
        self.q_proj = nn.Linear(embed_dim, embed_dim, bias=bias)
        self.out_proj = nn.Linear(embed_dim, embed_dim, bias=bias)

    def _shape(self, tensor: torch.Tensor, seq_len: int, bsz: int):
        return tensor.view(bsz, seq_len, self.num_heads, self.head_dim).transpose(1, 2).contiguous()

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

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

        bsz, tgt_len, _ = hidden_states.size()

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

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

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

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

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

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

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

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

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

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

        attn_output = torch.bmm(attn_probs, value_states)

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

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

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

        attn_output = self.out_proj(attn_output)

        return attn_output, attn_weights_reshaped, past_key_value


class BigBirdPegasusEncoderLayer(nn.Module):
    def __init__(self, config: BigBirdPegasusConfig, seed=None):
        super().__init__()
        self.attention_type = config.attention_type
        self.embed_dim = config.d_model
        self.self_attn = BigBirdPegasusEncoderAttention(config, seed=seed)
        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,
        band_mask=None,
        from_mask=None,
        to_mask=None,
        from_blocked_mask=None,
        to_blocked_mask=None,
        output_attentions: bool = False,
    ):
        """
        Args:
            hidden_states (`torch.FloatTensor`): input to the layer of shape `(batch, seq_len, embed_dim)`
            attention_mask (`torch.FloatTensor`): attention mask of size
                `(batch, 1, tgt_len, src_len)` where padding elements are indicated by very large negative values.
            output_attentions (`bool`, *optional*):
                Whether or not to return the attentions tensors of all attention layers. See `attentions` under
                returned tensors for more detail.
        """
        residual = hidden_states
        hidden_states = self.self_attn_layer_norm(hidden_states)

        self_attention_outputs = self.self_attn(
            hidden_states=hidden_states,
            attention_mask=attention_mask,
            head_mask=layer_head_mask,
            output_attentions=output_attentions,
            band_mask=band_mask,
            from_mask=from_mask,
            to_mask=to_mask,
            from_blocked_mask=from_blocked_mask,
            to_blocked_mask=to_blocked_mask,
        )
        hidden_states = self_attention_outputs[0]

        hidden_states = nn.functional.dropout(hidden_states, p=self.dropout, training=self.training)
        hidden_states = residual + hidden_states

        residual = hidden_states
        hidden_states = self.final_layer_norm(hidden_states)
        hidden_states = self.activation_fn(self.fc1(hidden_states))

        hidden_states = self.fc2(hidden_states)
        hidden_states = nn.functional.dropout(hidden_states, p=self.dropout, training=self.training)
        hidden_states = residual + hidden_states

        if hidden_states.dtype == torch.float16 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 += (self_attention_outputs[1],)

        return outputs

    def set_attention_type(self, value: str):
        if value not in ["original_full", "block_sparse"]:
            raise ValueError(
                f"attention_type can only be set to either 'original_full' or 'block_sparse', but is {value}"
            )
        # attention type is already correctly set
        if value == self.attention_type:
            return
        self.attention_type = value
        self.self_attn.set_attention_type(value)


class BigBirdPegasusDecoderLayer(nn.Module):
    def __init__(self, config: BigBirdPegasusConfig):
        super().__init__()
        self.embed_dim = config.d_model
        self.self_attn = BigBirdPegasusDecoderAttention(
            embed_dim=self.embed_dim,
            num_heads=config.decoder_attention_heads,
            dropout=config.attention_dropout,
            is_decoder=True,
            bias=config.use_bias,
        )
        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 = BigBirdPegasusDecoderAttention(
            self.embed_dim,
            config.decoder_attention_heads,
            dropout=config.attention_dropout,
            is_decoder=True,
            bias=config.use_bias,
        )
        self.encoder_attn_layer_norm = nn.LayerNorm(self.embed_dim)
        self.fc1 = nn.Linear(self.embed_dim, config.decoder_ffn_dim)
        self.fc2 = nn.Linear(config.decoder_ffn_dim, self.embed_dim)
        self.final_layer_norm = nn.LayerNorm(self.embed_dim)

    # Copied from transformers.models.mbart.modeling_mbart.MBartDecoderLayer.forward
    def forward(
        self,
        hidden_states: torch.Tensor,
        attention_mask: Optional[torch.Tensor] = None,
        encoder_hidden_states: Optional[torch.Tensor] = None,
        encoder_attention_mask: Optional[torch.Tensor] = None,
        layer_head_mask: Optional[torch.Tensor] = None,
        cross_attn_layer_head_mask: Optional[torch.Tensor] = None,
        past_key_value: Optional[Tuple[torch.Tensor]] = None,
        output_attentions: Optional[bool] = False,
        use_cache: Optional[bool] = True,
    ) -> torch.Tensor:
        """
        Args:
            hidden_states (`torch.FloatTensor`): input to the layer of shape `(batch, seq_len, embed_dim)`
            attention_mask (`torch.FloatTensor`): attention mask of size
                `(batch, 1, tgt_len, src_len)` where padding elements are indicated by very large negative values.
            encoder_hidden_states (`torch.FloatTensor`):
                cross attention input to the layer of shape `(batch, seq_len, embed_dim)`
            encoder_attention_mask (`torch.FloatTensor`): encoder attention mask of size
                `(batch, 1, tgt_len, src_len)` where padding elements are indicated by very large negative values.
            layer_head_mask (`torch.FloatTensor`): mask for attention heads in a given layer of size
                `(encoder_attention_heads,)`.
            cross_attn_layer_head_mask (`torch.FloatTensor`): mask for cross-attention heads in a given layer of
                size `(decoder_attention_heads,)`.
            past_key_value (`Tuple(torch.FloatTensor)`): cached past key and value projection states
            output_attentions (`bool`, *optional*):
                Whether or not to return the attentions tensors of all attention layers. See `attentions` under
                returned tensors for more detail.
        """
        residual = hidden_states
        hidden_states = self.self_attn_layer_norm(hidden_states)

        # Self Attention
        # 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

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

            # 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.final_layer_norm(hidden_states)
        hidden_states = self.activation_fn(self.fc1(hidden_states))
        hidden_states = nn.functional.dropout(hidden_states, p=self.activation_dropout, training=self.training)
        hidden_states = self.fc2(hidden_states)
        hidden_states = nn.functional.dropout(hidden_states, p=self.dropout, training=self.training)
        hidden_states = residual + hidden_states

        outputs = (hidden_states,)

        if output_attentions:
            outputs += (self_attn_weights, cross_attn_weights)

        if use_cache:
            outputs += (present_key_value,)

        return outputs


# Copied from transformers.models.bart.modeling_bart.BartClassificationHead with Bart->BigBirdPegasus
class BigBirdPegasusClassificationHead(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) -> 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 BigBirdPegasusPreTrainedModel(PreTrainedModel):
    config_class = BigBirdPegasusConfig
    base_model_prefix = "model"
    supports_gradient_checkpointing = True
    _no_split_modules = ["BigBirdPegasusEncoderLayer", "BigBirdPegasusDecoderLayer"]
    _skip_keys_device_placement = "past_key_values"

    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_()

    def _set_gradient_checkpointing(self, module, value=False):
        if isinstance(module, (BigBirdPegasusDecoder, BigBirdPegasusEncoder)):
            module.gradient_checkpointing = value

    @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


BIGBIRD_PEGASUS_START_DOCSTRING = r"""
    This model inherits from [`PreTrainedModel`]. Check the superclass documentation for the generic methods the
    library implements for all its model (such as downloading or saving, resizing the input embeddings 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 ([`BigBirdPegasusConfig`]):
            Model configuration class with all the parameters of the model. Initializing with a config file does not
            load the weights associated with the model, only the configuration. Check out the
            [`~PreTrainedModel.from_pretrained`] method to load the model weights.
"""

BIGBIRD_PEGASUS_GENERATION_EXAMPLE = r"""
    Summarization example:

    ```python
    >>> from transformers import AutoTokenizer, BigBirdPegasusForConditionalGeneration

    >>> model = BigBirdPegasusForConditionalGeneration.from_pretrained("google/bigbird-pegasus-large-arxiv")
    >>> tokenizer = AutoTokenizer.from_pretrained("google/bigbird-pegasus-large-arxiv")

    >>> ARTICLE_TO_SUMMARIZE = (
    ...     "The dominant sequence transduction models are based on complex recurrent or convolutional neural "
    ...     "networks in an encoder-decoder configuration. The best performing models also connect the encoder "
    ...     "and decoder through an attention mechanism. We propose a new simple network architecture, the Transformer, "
    ...     "based solely on attention mechanisms, dispensing with recurrence and convolutions entirely. "
    ...     "Experiments on two machine translation tasks show these models to be superior in quality "
    ...     "while being more parallelizable and requiring significantly less time to train."
    ... )
    >>> inputs = tokenizer([ARTICLE_TO_SUMMARIZE], max_length=4096, return_tensors="pt", truncation=True)

    >>> # Generate Summary
    >>> summary_ids = model.generate(inputs["input_ids"], num_beams=4, max_length=15)
    >>> tokenizer.batch_decode(summary_ids, skip_special_tokens=True, clean_up_tokenization_spaces=False)[0]
    'dominant sequence models are based on recurrent or convolutional neural networks .'
    ```
"""

BIGBIRD_PEGASUS_INPUTS_DOCSTRING = r"""
    Args:
        input_ids (`torch.LongTensor` of shape `(batch_size, sequence_length)`):
            Indices of input sequence tokens in the vocabulary. Padding will be ignored by default should you provide
            it.

            Indices can be obtained using [`AutoTokenizer`]. See [`PreTrainedTokenizer.encode`] and
            [`PreTrainedTokenizer.__call__`] for details.

            [What are input IDs?](../glossary#input-ids)
        attention_mask (`torch.Tensor` of shape `(batch_size, sequence_length)`, *optional*):
            Mask to avoid performing attention on padding token indices. Mask values selected in `[0, 1]`:

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

            [What are attention masks?](../glossary#attention-mask)
        decoder_input_ids (`torch.LongTensor` of shape `(batch_size, target_sequence_length)`, *optional*):
            Provide for translation and summarization training. By default, the model will create this tensor by
            shifting the `input_ids` to the right, following the paper.
        decoder_attention_mask (`torch.LongTensor` of shape `(batch_size, target_sequence_length)`, *optional*):
            Default behavior: generate a tensor that ignores pad tokens in `decoder_input_ids`. Causal mask will also
            be used by default.

            If you want to change padding behavior, you should read
            [`modeling_bigbird_pegasus._prepare_decoder_attention_mask`] and modify to your needs. See diagram 1 in
            [the paper](https://arxiv.org/abs/1910.13461) for more information on the default strategy.

        decoder_head_mask (`torch.Tensor` of shape `(num_layers, num_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**.

        encoder_outputs (`tuple(tuple(torch.FloatTensor)`, *optional*):
            Tuple consists of (`last_hidden_state`, *optional*: `hidden_states`, *optional*: `attentions`)
            `last_hidden_state` of shape `(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 (`tuple(tuple(torch.FloatTensor))`, *optional*, returned when `use_cache=True` is passed or when `config.use_cache=True`):
            Tuple of `tuple(torch.FloatTensor)` of length `config.n_layers`, with each tuple having 2 tensors of shape
            `(batch_size, num_heads, sequence_length, embed_size_per_head)`) and 2 additional tensors of shape
            `(batch_size, num_heads, encoder_sequence_length, embed_size_per_head)`.

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

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

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

BIGBIRD_PEGASUS_STANDALONE_INPUTS_DOCSTRING = r"""
    Args:
        input_ids (`torch.LongTensor` of shape `(batch_size, sequence_length)`):
            Indices of input sequence tokens in the vocabulary. Padding will be ignored by default should you provide
            it.

            Indices can be obtained using [`ProphetNetTokenizer`]. See [`PreTrainedTokenizer.encode`] and
            [`PreTrainedTokenizer.__call__`] for details.

            [What are input IDs?](../glossary#input-ids)
        attention_mask (`torch.Tensor` of shape `(batch_size, sequence_length)`, *optional*):
            Mask to avoid performing attention on padding token indices. Mask values selected in `[0, 1]`:

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

            [What are attention masks?](../glossary#attention-mask)
        output_attentions (`bool`, *optional*):
            Whether or not to return the attentions tensors of all attention layers. See `attentions` under returned
            tensors for more detail.
        output_hidden_states (`bool`, *optional*):
            Whether or not to return the hidden states of all layers. See `hidden_states` under returned tensors for
            more detail.
        return_dict (`bool`, *optional*):
            Whether or not to return a [`~utils.ModelOutput`] instead of a plain tuple.
"""


class BigBirdPegasusEncoder(BigBirdPegasusPreTrainedModel):
    """
    Transformer encoder consisting of *config.encoder_layers* self attention layers. Each layer is a
    [`BigBirdPegasusEncoderLayer`].

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

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

        self.attention_type = config.attention_type
        self.block_size = config.block_size

        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

        self.embed_tokens = nn.Embedding(config.vocab_size, embed_dim, self.padding_idx)

        if embed_tokens is not None:
            self.embed_tokens.weight = embed_tokens.weight

        self.embed_positions = BigBirdPegasusLearnedPositionalEmbedding(
            config.max_position_embeddings,
            embed_dim,
        )
        self.layers = nn.ModuleList([BigBirdPegasusEncoderLayer(config, seed=i) for i in range(config.encoder_layers)])
        self.layernorm_embedding = nn.LayerNorm(embed_dim)

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

    def forward(
        self,
        input_ids: Optional[torch.Tensor] = None,
        attention_mask: Optional[torch.Tensor] = None,
        head_mask: Optional[torch.Tensor] = None,
        inputs_embeds: Optional[torch.Tensor] = None,
        output_attentions: Optional[bool] = None,
        output_hidden_states: Optional[bool] = None,
        return_dict: Optional[bool] = None,
    ):
        r"""
        Args:
            input_ids (`torch.LongTensor` of shape `(batch_size, sequence_length)`):
                Indices of input sequence tokens in the vocabulary. Padding will be ignored by default should you
                provide it.

                Indices can be obtained using [`AutoTokenizer`]. See [`PreTrainedTokenizer.encode`] and
                [`PreTrainedTokenizer.__call__`] for details.

                [What are input IDs?](../glossary#input-ids)
            attention_mask (`torch.Tensor` of shape `(batch_size, sequence_length)`, *optional*):
                Mask to avoid performing attention on padding token indices. Mask values selected in `[0, 1]`:

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

                [What are attention masks?](../glossary#attention-mask)

            inputs_embeds (`torch.FloatTensor` of shape `(batch_size, sequence_length, hidden_size)`, *optional*):
                Optionally, instead of passing `input_ids` you can choose to directly pass an embedded representation.
                This is useful if you want more control over how to convert `input_ids` indices into associated vectors
                than the model's internal embedding lookup matrix.
            output_attentions (`bool`, *optional*):
                Whether or not to return the attentions tensors of all attention layers. See `attentions` under
                returned tensors for more detail.
            output_hidden_states (`bool`, *optional*):
                Whether or not to return the hidden states of all layers. See `hidden_states` under returned tensors
                for more detail.
            return_dict (`bool`, *optional*):
                Whether or not to return a [`~utils.ModelOutput`] instead of a plain tuple.
        """
        output_attentions = output_attentions if output_attentions is not None else self.config.output_attentions
        output_hidden_states = (
            output_hidden_states if output_hidden_states is not None else self.config.output_hidden_states
        )
        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:
            self.warn_if_padding_and_no_attention_mask(input_ids, attention_mask)
            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 = nn.functional.dropout(hidden_states, p=self.dropout, training=self.training)

        if attention_mask is None:
            attention_mask = torch.ones(input_shape, device=hidden_states.device)
        attention_mask = attention_mask.long()

        # in order to use block_sparse attention, sequence_length has to be at least
        # bigger than all global attentions: 2 * block_size
        # + sliding tokens: 3 * block_size
        # + random tokens: 2 * num_random_blocks * block_size
        max_tokens_to_attend = (5 + 2 * self.config.num_random_blocks) * self.config.block_size
        if self.attention_type == "block_sparse" and input_shape[1] <= max_tokens_to_attend:
            # change attention_type from block_sparse to original_full
            sequence_length = input_shape[1]
            logger.warning(
                "Attention type 'block_sparse' is not possible if sequence_length: "
                f"{sequence_length} <= num global tokens: 2 * config.block_size "
                "+ min. num sliding tokens: 3 * config.block_size "
                "+ config.num_random_blocks * config.block_size "
                "+ additional buffer: config.num_random_blocks * config.block_size "
                f"= {max_tokens_to_attend} with config.block_size "
                f"= {self.config.block_size}, config.num_random_blocks "
                f"= {self.config.num_random_blocks}. "
                "Changing attention type to 'original_full'..."
            )
            self.set_attention_type("original_full")

        if self.attention_type == "block_sparse":
            padding_len, hidden_states, attention_mask = self._pad_to_block_size(hidden_states, attention_mask)
        else:
            padding_len = 0

        # expand attention_mask
        if self.attention_type == "original_full":
            # [bsz, seq_len] -> [bsz, 1, tgt_seq_len, src_seq_len]
            attention_mask = _expand_mask(attention_mask, inputs_embeds.dtype)
            blocked_encoder_mask = band_mask = from_mask = to_mask = None
        elif self.attention_type == "block_sparse":
            blocked_encoder_mask, band_mask, from_mask, to_mask = self.create_masks_for_block_sparse_attn(
                attention_mask, self.block_size
            )
            attention_mask = None
        else:
            raise ValueError(
                f"attention_type can either be original_full or block_sparse, but is {self.attention_type}"
            )

        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:
            if head_mask.size()[0] != len(self.layers):
                raise ValueError(
                    f"The head_mask should be specified for {len(self.layers)} layers, but it is for"
                    f" {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)
            to_drop = False
            if self.training:
                dropout_probability = torch.rand([])
                if dropout_probability < self.layerdrop:  # skip the layer
                    to_drop = True

            if to_drop:
                layer_outputs = (None, None)
            else:
                if self.gradient_checkpointing 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),
                        band_mask,
                        from_mask,
                        to_mask,
                        blocked_encoder_mask,
                        blocked_encoder_mask,
                    )
                else:
                    layer_outputs = encoder_layer(
                        hidden_states,
                        attention_mask,
                        layer_head_mask=(head_mask[idx] if head_mask is not None else None),
                        band_mask=band_mask,
                        from_mask=from_mask,
                        to_mask=to_mask,
                        from_blocked_mask=blocked_encoder_mask,
                        to_blocked_mask=blocked_encoder_mask,
                        output_attentions=output_attentions,
                    )

                hidden_states = layer_outputs[0]

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

        hidden_states = self.layernorm_embedding(hidden_states)

        if output_hidden_states:
            encoder_states = encoder_states + (hidden_states,)

        if padding_len > 0:
            # unpad `sequence_output` because the calling function is expecting a length == input_ids.size(1)
            hidden_states = hidden_states[:, :-padding_len]

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

        self.encoder_o = hidden_states

        return BaseModelOutput(
            last_hidden_state=hidden_states, hidden_states=encoder_states, attentions=all_attentions
        )

    def set_attention_type(self, value: str):
        if value not in ["original_full", "block_sparse"]:
            raise ValueError(
                f"attention_type can only be set to either 'original_full' or 'block_sparse', but is {value}"
            )
        # attention type is already correctly set
        if value == self.attention_type:
            return
        self.attention_type = value
        for layer in self.layers:
            layer.set_attention_type(value)

    @staticmethod  # Copied from transformers.models.big_bird.modeling_big_bird.BigBirdModel.create_masks_for_block_sparse_attn
    def create_masks_for_block_sparse_attn(attention_mask: torch.Tensor, block_size: int):
        batch_size, seq_length = attention_mask.size()
        if seq_length % block_size != 0:
            raise ValueError(
                f"Sequence length must be multiple of block size, but sequence length is {seq_length}, while block"
                f" size is {block_size}."
            )

        def create_band_mask_from_inputs(from_blocked_mask, to_blocked_mask):
            """
            Create 3D attention mask from a 2D tensor mask.

            Args:
                from_blocked_mask: 2D Tensor of shape [batch_size,
                from_seq_length//from_block_size, from_block_size].
                to_blocked_mask: int32 Tensor of shape [batch_size,
                to_seq_length//to_block_size, to_block_size].

            Returns:
                float Tensor of shape [batch_size, 1, from_seq_length//from_block_size-4, from_block_size,
                3*to_block_size].
            """
            exp_blocked_to_pad = torch.cat(
                [to_blocked_mask[:, 1:-3], to_blocked_mask[:, 2:-2], to_blocked_mask[:, 3:-1]], dim=2
            )
            band_mask = torch.einsum("blq,blk->blqk", from_blocked_mask[:, 2:-2], exp_blocked_to_pad)
            band_mask.unsqueeze_(1)
            return band_mask

        blocked_encoder_mask = attention_mask.view(batch_size, seq_length // block_size, block_size)
        band_mask = create_band_mask_from_inputs(blocked_encoder_mask, blocked_encoder_mask)

        from_mask = attention_mask.view(batch_size, 1, seq_length, 1)
        to_mask = attention_mask.view(batch_size, 1, 1, seq_length)

        return blocked_encoder_mask, band_mask, from_mask, to_mask

    def _pad_to_block_size(self, hidden_states: torch.Tensor, attention_mask: torch.Tensor):
        """A helper function to pad tokens and mask to work with implementation of BigBird block-sparse attention."""
        # padding
        block_size = self.config.block_size
        batch_size, seq_len = hidden_states.shape[:2]

        padding_len = (block_size - seq_len % block_size) % block_size
        if padding_len > 0:
            logger.info(
                f"Input ids are automatically padded from {seq_len} to {seq_len + padding_len} to be a multiple of "
                f"`config.block_size`: {block_size}"
            )
            pad_id = self.config.pad_token_id
            device = hidden_states.device
            input_ids_padding = torch.ones((batch_size, padding_len), dtype=torch.long, device=device) * pad_id
            inputs_embeds_padding = self.embed_tokens(input_ids_padding)
            hidden_states = torch.cat([hidden_states, inputs_embeds_padding], dim=-2)

            attention_mask = nn.functional.pad(
                attention_mask, (0, padding_len), value=0
            )  # no attention on the padding tokens

        return padding_len, hidden_states, attention_mask


class BigBirdPegasusDecoder(BigBirdPegasusPreTrainedModel):
    """
    Transformer decoder consisting of *config.decoder_layers* layers. Each layer is a [`BigBirdPegasusDecoderLayer`]

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

    def __init__(self, config: BigBirdPegasusConfig, 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

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

        if embed_tokens is not None:
            self.embed_tokens.weight = embed_tokens.weight

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

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

    def get_input_embeddings(self):
        return self.embed_tokens

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

    # Copied from transformers.models.bart.modeling_bart.BartDecoder._prepare_decoder_attention_mask
    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,
                device=inputs_embeds.device,
                past_key_values_length=past_key_values_length,
            )

        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]).to(
                inputs_embeds.device
            )
            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: Optional[torch.Tensor] = None,
        attention_mask: Optional[torch.Tensor] = None,
        encoder_hidden_states: Optional[torch.Tensor] = None,
        encoder_attention_mask: Optional[torch.Tensor] = None,
        head_mask: Optional[torch.Tensor] = None,
        cross_attn_head_mask: Optional[torch.Tensor] = None,
        past_key_values: Optional[List[torch.FloatTensor]] = None,
        inputs_embeds: Optional[torch.Tensor] = None,
        use_cache: Optional[bool] = None,
        output_attentions: Optional[bool] = None,
        output_hidden_states: Optional[bool] = None,
        return_dict: Optional[bool] = None,
    ):
        r"""
        Args:
            input_ids (`torch.LongTensor` of shape `(batch_size, sequence_length)`):
                Indices of input sequence tokens in the vocabulary. Padding will be ignored by default should you
                provide it.

                Indices can be obtained using [`AutoTokenizer`]. See [`PreTrainedTokenizer.encode`] and
                [`PreTrainedTokenizer.__call__`] for details.

                [What are input IDs?](../glossary#input-ids)
            attention_mask (`torch.Tensor` of shape `(batch_size, sequence_length)`, *optional*):
                Mask to avoid performing attention on padding token indices. Mask values selected in `[0, 1]`:

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

                [What are attention masks?](../glossary#attention-mask)
            encoder_hidden_states (`torch.FloatTensor` of shape `(batch_size, encoder_sequence_length, hidden_size)`, *optional*):
                Sequence of hidden-states at the output of the last layer of the encoder. Used in the cross-attention
                of the decoder.
            encoder_attention_mask (`torch.LongTensor` of shape `(batch_size, encoder_sequence_length)`, *optional*):
                Mask to avoid performing cross-attention on padding tokens indices of encoder input_ids. Mask values
                selected in `[0, 1]`:

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

                [What are attention masks?](../glossary#attention-mask)
            head_mask (`torch.Tensor` of shape `(decoder_layers, decoder_attention_heads)`, *optional*):
                Mask to nullify selected heads of the attention modules. Mask values selected in `[0, 1]`:

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

            cross_attn_head_mask (`torch.Tensor` of shape `(decoder_layers, decoder_attention_heads)`, *optional*):
                Mask to nullify selected heads of the cross-attention modules in 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 (`tuple(tuple(torch.FloatTensor))`, *optional*, returned when `use_cache=True` is passed or when `config.use_cache=True`):
                Tuple of `tuple(torch.FloatTensor)` of length `config.n_layers`, with each tuple having 2 tensors of
                shape `(batch_size, num_heads, sequence_length, embed_size_per_head)`) and 2 additional tensors of
                shape `(batch_size, num_heads, encoder_sequence_length, embed_size_per_head)`.

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

                If `past_key_values` are used, the user can optionally input only the last `decoder_input_ids` (those
                that don't have their past key value states given to this model) of shape `(batch_size, 1)` instead of
                all `decoder_input_ids` of shape `(batch_size, sequence_length)`. inputs_embeds (`torch.FloatTensor` of
                shape `(batch_size, sequence_length, hidden_size)`, *optional*): Optionally, instead of passing
                `input_ids` you can choose to directly pass an embedded representation. This is useful if you want more
                control over how to convert `input_ids` indices into associated vectors than the model's internal
                embedding lookup matrix.
            output_attentions (`bool`, *optional*):
                Whether or not to return the attentions tensors of all attention layers. See `attentions` under
                returned tensors for more detail.
            output_hidden_states (`bool`, *optional*):
                Whether or not to return the hidden states of all layers. See `hidden_states` under returned tensors
                for more detail.
            return_dict (`bool`, *optional*):
                Whether or not to return a [`~utils.ModelOutput`] instead of a plain tuple.
        """
        output_attentions = output_attentions if output_attentions is not None else self.config.output_attentions
        output_hidden_states = (
            output_hidden_states if output_hidden_states is not None else self.config.output_hidden_states
        )
        use_cache = use_cache if use_cache is not None else self.config.use_cache
        return_dict = return_dict if return_dict is not None else self.config.use_return_dict

        # retrieve input_ids and inputs_embeds
        if input_ids is not None and inputs_embeds is not None:
            raise ValueError("You cannot specify both decoder_input_ids and decoder_inputs_embeds at the same time")
        elif input_ids is not None:
            input_shape = input_ids.size()
            input_ids = input_ids.view(-1, input_shape[-1])
        elif inputs_embeds is not None:
            input_shape = inputs_embeds.size()[:-1]
        else:
            raise ValueError("You have to specify either decoder_input_ids or decoder_inputs_embeds")

        # 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)
        positions = positions.to(inputs_embeds.device)

        hidden_states = inputs_embeds + positions

        hidden_states = nn.functional.dropout(hidden_states, p=self.dropout, training=self.training)

        if self.gradient_checkpointing and self.training:
            if use_cache:
                logger.warning_once(
                    "`use_cache=True` is incompatible with gradient checkpointing. Setting `use_cache=False`..."
                )
                use_cache = False

        # 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:
                if attn_mask.size()[0] != len(self.layers):
                    raise ValueError(
                        f"The `{mask_name}` should be specified for {len(self.layers)} layers, but it is for"
                        f" {head_mask.size()[0]}."
                    )
        for idx, decoder_layer in enumerate(self.layers):
            # add LayerDrop (see https://arxiv.org/abs/1909.11556 for description)
            if output_hidden_states:
                all_hidden_states += (hidden_states,)
            if self.training:
                dropout_probability = torch.rand([])
                if dropout_probability < self.layerdrop:
                    continue

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

            if self.gradient_checkpointing and self.training:

                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],)

        hidden_states = self.layernorm_embedding(hidden_states)

        # 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 BigBirdPegasus Model outputting raw hidden-states without any specific head on top.",
    BIGBIRD_PEGASUS_START_DOCSTRING,
)
class BigBirdPegasusModel(BigBirdPegasusPreTrainedModel):
    _tied_weights_keys = ["encoder.embed_tokens.weight", "decoder.embed_tokens.weight"]

    def __init__(self, config: BigBirdPegasusConfig):
        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 = BigBirdPegasusEncoder(config, self.shared)
        self.decoder = BigBirdPegasusDecoder(config, self.shared)

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

    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(BIGBIRD_PEGASUS_INPUTS_DOCSTRING)
    @add_code_sample_docstrings(
        checkpoint=_CHECKPOINT_FOR_DOC,
        output_type=Seq2SeqModelOutput,
        config_class=_CONFIG_FOR_DOC,
        expected_output=_EXPECTED_OUTPUT_SHAPE,
    )
    # Copied from transformers.models.bart.modeling_bart.BartModel.forward with Bart->BigBirdPegasus
    def forward(
        self,
        input_ids: torch.LongTensor = None,
        attention_mask: Optional[torch.Tensor] = None,
        decoder_input_ids: Optional[torch.LongTensor] = None,
        decoder_attention_mask: Optional[torch.LongTensor] = None,
        head_mask: Optional[torch.Tensor] = None,
        decoder_head_mask: Optional[torch.Tensor] = None,
        cross_attn_head_mask: Optional[torch.Tensor] = None,
        encoder_outputs: Optional[List[torch.FloatTensor]] = None,
        past_key_values: Optional[List[torch.FloatTensor]] = None,
        inputs_embeds: Optional[torch.FloatTensor] = None,
        decoder_inputs_embeds: Optional[torch.FloatTensor] = None,
        use_cache: Optional[bool] = None,
        output_attentions: Optional[bool] = None,
        output_hidden_states: Optional[bool] = None,
        return_dict: Optional[bool] = None,
    ) -> Union[Tuple, Seq2SeqModelOutput]:
        # different to other models, BigBirdPegasus 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:
            if input_ids is None:
                raise ValueError(
                    "If no `decoder_input_ids` or `decoder_inputs_embeds` are "
                    "passed, `input_ids` cannot be `None`. Please pass either "
                    "`input_ids` or `decoder_input_ids` or `decoder_inputs_embeds`."
                )

            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 BigBirdPegasus Model with a language modeling head. Can be used for summarization.",
    BIGBIRD_PEGASUS_START_DOCSTRING,
)
# Copied from transformers.models.bart.modeling_bart.BartForConditionalGeneration with Bart->BigBirdPegasus, BART->BIGBIRD_PEGASUS
class BigBirdPegasusForConditionalGeneration(BigBirdPegasusPreTrainedModel):
    base_model_prefix = "model"
    _tied_weights_keys = ["encoder.embed_tokens.weight", "decoder.embed_tokens.weight", "lm_head.weight"]
    _keys_to_ignore_on_load_missing = ["final_logits_bias"]

    def __init__(self, config: BigBirdPegasusConfig):
        super().__init__(config)
        self.model = BigBirdPegasusModel(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)

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

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

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

    def resize_token_embeddings(self, new_num_tokens: int, pad_to_multiple_of: Optional[int] = None) -> nn.Embedding:
        new_embeddings = super().resize_token_embeddings(new_num_tokens, pad_to_multiple_of)
        self._resize_final_logits_bias(new_embeddings.weight.shape[0])
        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(BIGBIRD_PEGASUS_INPUTS_DOCSTRING)
    @replace_return_docstrings(output_type=Seq2SeqLMOutput, config_class=_CONFIG_FOR_DOC)
    @add_end_docstrings(BIGBIRD_PEGASUS_GENERATION_EXAMPLE)
    def forward(
        self,
        input_ids: torch.LongTensor = None,
        attention_mask: Optional[torch.Tensor] = None,
        decoder_input_ids: Optional[torch.LongTensor] = None,
        decoder_attention_mask: Optional[torch.LongTensor] = None,
        head_mask: Optional[torch.Tensor] = None,
        decoder_head_mask: Optional[torch.Tensor] = None,
        cross_attn_head_mask: Optional[torch.Tensor] = None,
        encoder_outputs: Optional[List[torch.FloatTensor]] = None,
        past_key_values: Optional[List[torch.FloatTensor]] = None,
        inputs_embeds: Optional[torch.FloatTensor] = None,
        decoder_inputs_embeds: Optional[torch.FloatTensor] = None,
        labels: Optional[torch.LongTensor] = None,
        use_cache: Optional[bool] = None,
        output_attentions: Optional[bool] = None,
        output_hidden_states: Optional[bool] = None,
        return_dict: Optional[bool] = None,
    ) -> Union[Tuple, Seq2SeqLMOutput]:
        r"""
        labels (`torch.LongTensor` of shape `(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 use_cache:
                logger.warning("The `use_cache` argument is changed to `False` since `labels` is provided.")
            use_cache = False
            if decoder_input_ids is None and decoder_inputs_embeds is None:
                decoder_input_ids = shift_tokens_right(
                    labels, self.config.pad_token_id, self.config.decoder_start_token_id
                )

        outputs = self.model(
            input_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])
        lm_logits = lm_logits + self.final_logits_bias.to(lm_logits.device)

        masked_lm_loss = None
        if labels is not None:
            labels = labels.to(lm_logits.device)
            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_key_values=None,
        attention_mask=None,
        decoder_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_key_values is used
        if past_key_values 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_key_values,
            "decoder_input_ids": decoder_input_ids,
            "attention_mask": attention_mask,
            "decoder_attention_mask": decoder_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_key_values, beam_idx):
        reordered_past = ()
        for layer_past in past_key_values:
            # 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.to(past_state.device)) for past_state in layer_past[:2])
                + layer_past[2:],
            )
        return reordered_past


@add_start_docstrings(
    """
    BigBirdPegasus model with a sequence classification/head on top (a linear layer on top of the pooled output) e.g.
    for GLUE tasks.
    """,
    BIGBIRD_PEGASUS_START_DOCSTRING,
)
class BigBirdPegasusForSequenceClassification(BigBirdPegasusPreTrainedModel):
    _tied_weights_keys = ["encoder.embed_tokens.weight", "decoder.embed_tokens.weight"]

    def __init__(self, config: BigBirdPegasusConfig, **kwargs):
        super().__init__(config, **kwargs)
        self.model = BigBirdPegasusModel(config)
        self.classification_head = BigBirdPegasusClassificationHead(
            config.d_model,
            config.d_model,
            config.num_labels,
            config.classifier_dropout,
        )

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

    @add_start_docstrings_to_model_forward(BIGBIRD_PEGASUS_INPUTS_DOCSTRING)
    @add_code_sample_docstrings(
        checkpoint=_CHECKPOINT_FOR_DOC,
        output_type=Seq2SeqSequenceClassifierOutput,
        config_class=_CONFIG_FOR_DOC,
    )
    # Copied from transformers.models.bart.modeling_bart.BartForSequenceClassification.forward
    def forward(
        self,
        input_ids: torch.LongTensor = None,
        attention_mask: Optional[torch.Tensor] = None,
        decoder_input_ids: Optional[torch.LongTensor] = None,
        decoder_attention_mask: Optional[torch.LongTensor] = None,
        head_mask: Optional[torch.Tensor] = None,
        decoder_head_mask: Optional[torch.Tensor] = None,
        cross_attn_head_mask: Optional[torch.Tensor] = None,
        encoder_outputs: Optional[List[torch.FloatTensor]] = None,
        inputs_embeds: Optional[torch.FloatTensor] = None,
        decoder_inputs_embeds: Optional[torch.FloatTensor] = None,
        labels: Optional[torch.LongTensor] = None,
        use_cache: Optional[bool] = None,
        output_attentions: Optional[bool] = None,
        output_hidden_states: Optional[bool] = None,
        return_dict: Optional[bool] = None,
    ) -> Union[Tuple, Seq2SeqSequenceClassifierOutput]:
        r"""
        labels (`torch.LongTensor` of shape `(batch_size,)`, *optional*):
            Labels for computing the sequence classification/regression loss. Indices should be in `[0, ...,
            config.num_labels - 1]`. If `config.num_labels > 1` a 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).to(hidden_states.device)

        if len(torch.unique_consecutive(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:
            labels = labels.to(logits.device)
            if self.config.problem_type is None:
                if self.config.num_labels == 1:
                    self.config.problem_type = "regression"
                elif self.config.num_labels > 1 and (labels.dtype == torch.long or labels.dtype == torch.int):
                    self.config.problem_type = "single_label_classification"
                else:
                    self.config.problem_type = "multi_label_classification"

            if self.config.problem_type == "regression":
                loss_fct = MSELoss()
                if self.config.num_labels == 1:
                    loss = loss_fct(logits.squeeze(), labels.squeeze())
                else:
                    loss = loss_fct(logits, labels)
            elif self.config.problem_type == "single_label_classification":
                loss_fct = CrossEntropyLoss()
                loss = loss_fct(logits.view(-1, self.config.num_labels), labels.view(-1))
            elif self.config.problem_type == "multi_label_classification":
                loss_fct = BCEWithLogitsLoss()
                loss = loss_fct(logits, labels)
        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(
    """
    BigBirdPegasus 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`).
    """,
    BIGBIRD_PEGASUS_START_DOCSTRING,
)
class BigBirdPegasusForQuestionAnswering(BigBirdPegasusPreTrainedModel):
    _tied_weights_keys = ["encoder.embed_tokens.weight", "decoder.embed_tokens.weight"]

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

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

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

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

    @add_start_docstrings_to_model_forward(BIGBIRD_PEGASUS_INPUTS_DOCSTRING)
    @add_code_sample_docstrings(
        checkpoint=_CHECKPOINT_FOR_DOC,
        output_type=Seq2SeqQuestionAnsweringModelOutput,
        config_class=_CONFIG_FOR_DOC,
    )
    # Copied from transformers.models.bart.modeling_bart.BartForQuestionAnswering.forward
    def forward(
        self,
        input_ids: torch.Tensor = None,
        attention_mask: Optional[torch.Tensor] = None,
        decoder_input_ids: Optional[torch.LongTensor] = None,
        decoder_attention_mask: Optional[torch.LongTensor] = None,
        head_mask: Optional[torch.Tensor] = None,
        decoder_head_mask: Optional[torch.Tensor] = None,
        cross_attn_head_mask: Optional[torch.Tensor] = None,
        encoder_outputs: Optional[List[torch.FloatTensor]] = None,
        start_positions: Optional[torch.LongTensor] = None,
        end_positions: Optional[torch.LongTensor] = None,
        inputs_embeds: Optional[torch.FloatTensor] = None,
        decoder_inputs_embeds: Optional[torch.FloatTensor] = None,
        use_cache: Optional[bool] = None,
        output_attentions: Optional[bool] = None,
        output_hidden_states: Optional[bool] = None,
        return_dict: Optional[bool] = None,
    ) -> Union[Tuple, Seq2SeqQuestionAnsweringModelOutput]:
        r"""
        start_positions (`torch.LongTensor` of shape `(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 (`torch.LongTensor` of shape `(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,
        )


# Copied from transformers.models.pegasus.modeling_pegasus.PegasusDecoderWrapper with Pegasus->BigBirdPegasus
class BigBirdPegasusDecoderWrapper(BigBirdPegasusPreTrainedModel):
    """
    This wrapper class is a helper class to correctly load pretrained checkpoints when the causal language model is
    used in combination with the [`EncoderDecoderModel`] framework.
    """

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

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


class BigBirdPegasusForCausalLM(BigBirdPegasusPreTrainedModel):
    _tied_weights_keys = ["lm_head.weight"]

    def __init__(self, config):
        config = copy.deepcopy(config)
        config.is_decoder = True
        config.is_encoder_decoder = False
        super().__init__(config)
        self.model = BigBirdPegasusDecoderWrapper(config)

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

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

    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: torch.LongTensor = None,
        attention_mask: Optional[torch.Tensor] = None,
        encoder_hidden_states: Optional[torch.FloatTensor] = None,
        encoder_attention_mask: Optional[torch.FloatTensor] = None,
        head_mask: Optional[torch.Tensor] = None,
        cross_attn_head_mask: Optional[torch.Tensor] = None,
        past_key_values: Optional[Tuple[Tuple[torch.Tensor]]] = None,
        inputs_embeds: Optional[torch.FloatTensor] = None,
        labels: Optional[torch.LongTensor] = None,
        use_cache: Optional[bool] = None,
        output_attentions: Optional[bool] = None,
        output_hidden_states: Optional[bool] = None,
        return_dict: Optional[bool] = None,
    ) -> Union[Tuple, CausalLMOutputWithCrossAttentions]:
        r"""
        Args:
            input_ids (`torch.LongTensor` of shape `(batch_size, sequence_length)`):
                Indices of input sequence tokens in the vocabulary. Padding will be ignored by default should you
                provide it.

                Indices can be obtained using [`AutoTokenizer`]. See [`PreTrainedTokenizer.encode`] and
                [`PreTrainedTokenizer.__call__`] for details.

                [What are input IDs?](../glossary#input-ids)
            attention_mask (`torch.Tensor` of shape `(batch_size, sequence_length)`, *optional*):
                Mask to avoid performing attention on padding token indices. Mask values selected in `[0, 1]`:

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

                [What are attention masks?](../glossary#attention-mask)
            encoder_hidden_states  (`torch.FloatTensor` of shape `(batch_size, 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 (`torch.FloatTensor` of shape `(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 (`torch.Tensor` of shape `(decoder_layers, decoder_attention_heads)`, *optional*):
                Mask to nullify selected heads of the attention modules. Mask values selected in `[0, 1]`:

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

            cross_attn_head_mask (`torch.Tensor` of shape `(decoder_layers, decoder_attention_heads)`, *optional*):
                Mask to nullify selected heads of the 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 (`tuple(tuple(torch.FloatTensor))`, *optional*, returned when `use_cache=True` is passed or when `config.use_cache=True`):
                Tuple of `tuple(torch.FloatTensor)` of length `config.n_layers`, with each tuple having 2 tensors of
                shape `(batch_size, num_heads, sequence_length, embed_size_per_head)`) and 2 additional tensors of
                shape `(batch_size, num_heads, encoder_sequence_length, embed_size_per_head)`. 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 `past_key_values` input) to speed up sequential decoding.

                If `past_key_values` are used, the user can optionally input only the last `decoder_input_ids` (those
                that don't have their past key value states given to this model) of shape `(batch_size, 1)` instead of
                all `decoder_input_ids` of shape `(batch_size, sequence_length)`.
            labels (`torch.LongTensor` of shape `(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 (`bool`, *optional*):
                If set to `True`, `past_key_values` key value states are returned and can be used to speed up decoding
                (see `past_key_values`).

                - 1 for tokens that are **not masked**,
                - 0 for tokens that are **masked**.
            output_attentions (`bool`, *optional*):
                Whether or not to return the attentions tensors of all attention layers. See `attentions` under
                returned tensors for more detail.
            output_hidden_states (`bool`, *optional*):
                Whether or not to return the hidden states of all layers. See `hidden_states` under returned tensors
                for more detail.
            return_dict (`bool`, *optional*):
                Whether or not to return a [`~utils.ModelOutput`] instead of a plain tuple.

        Returns:

        Example:

        ```python
        >>> from transformers import AutoTokenizer, BigBirdPegasusForCausalLM

        >>> tokenizer = AutoTokenizer.from_pretrained("google/bigbird-pegasus-large-arxiv")
        >>> model = BigBirdPegasusForCausalLM.from_pretrained(
        ...     "google/bigbird-pegasus-large-arxiv", 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)

        >>> logits = outputs.logits
        ```"""

        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_key_values=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_key_values:
            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_key_values,
            "use_cache": use_cache,
        }

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