ai-content-maker/.venv/Lib/site-packages/transformers/models/longt5/modeling_longt5.py

# coding=utf-8
# Copyright 2022 Google LLC., LongT5 Authors and HuggingFace Inc. team.
#
# Licensed under the Apache License, Version 2.0 (the "License");
# you may not use this file except in compliance with the License.
# You may obtain a copy of the License at
#
#     http://www.apache.org/licenses/LICENSE-2.0
#
# Unless required by applicable law or agreed to in writing, software
# distributed under the License is distributed on an "AS IS" BASIS,
# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
# See the License for the specific language governing permissions and
# limitations under the License.
""" PyTorch LongT5 model."""


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

import torch
from torch import nn
from torch.nn import CrossEntropyLoss

from ...activations import ACT2FN
from ...modeling_outputs import (
    BaseModelOutput,
    BaseModelOutputWithPastAndCrossAttentions,
    Seq2SeqLMOutput,
    Seq2SeqModelOutput,
)
from ...modeling_utils import PreTrainedModel
from ...pytorch_utils import ALL_LAYERNORM_LAYERS, find_pruneable_heads_and_indices, prune_linear_layer
from ...utils import (
    DUMMY_INPUTS,
    DUMMY_MASK,
    add_start_docstrings,
    add_start_docstrings_to_model_forward,
    is_torch_fx_proxy,
    logging,
    replace_return_docstrings,
)
from .configuration_longt5 import LongT5Config


logger = logging.get_logger(__name__)

_CONFIG_FOR_DOC = "LongT5Config"
_CHECKPOINT_FOR_DOC = "google/long-t5-local-base"

# TODO: Update before the merge

from ..deprecated._archive_maps import LONGT5_PRETRAINED_MODEL_ARCHIVE_LIST  # noqa: F401, E402


def _pad_to_multiple(x: torch.Tensor, block_len: int, dim: int, pad_value: int = 0) -> torch.Tensor:
    """Pad a tensor so that a sequence length will be a multiple of `block_len`"""
    pad_len = -x.shape[dim] % block_len
    # Handle cases when an empty input sequence is given
    if not all(x.shape):
        new_shape = list(x.shape)
        new_shape[dim] += pad_len
        return torch.zeros(new_shape, dtype=x.dtype)

    pad = [(0, 0)] * x.ndim
    pad[dim] = (0, pad_len)
    pad = sum(pad[::-1], ())
    x = nn.functional.pad(x, pad=pad, mode="constant", value=pad_value)
    return x


def _split_into_blocks(x: torch.Tensor, block_len: int, dim: int) -> torch.Tensor:
    """Split an input tensor into blocks of a given `block_len` along the given `dim`. If the dimension length
    is not a multiple of `block_len`, it will be padded first with selected `pad_value`.
    """
    # pad tensor to multiple of block_len
    if x.shape[dim] % block_len != 0:
        x = _pad_to_multiple(x, block_len, dim, pad_value=0)
    num_blocks = x.shape[dim] // block_len
    output_shape = x.shape[:dim] + (num_blocks, block_len) + x.shape[(dim + 1) :]
    # If 0 is in output_shape, we cannot apply reshape because of incompatibility with ONNX conversion
    if 0 in output_shape:
        return torch.empty(output_shape, dtype=x.dtype, device=x.device)
    return x.reshape(output_shape)


def _concatenate_3_blocks(x: torch.Tensor, block_dim: int, sequence_dim: int, pad_value: int = 0) -> torch.Tensor:
    """Concatenate three consecutive blocks for each input block for local attentiont.

    For more information, see: https://arxiv.org/pdf/2112.07916.pdf.
    """
    num_blocks = x.shape[block_dim]

    pad = [(0, 0)] * x.ndim
    pad[block_dim] = (1, 1)
    pad = sum(pad[::-1], ())
    # [batch_size, num_blocks, block_len] -> [batch_size, num_blocks + 2, block_len]
    x = nn.functional.pad(x, pad=pad, mode="constant", value=pad_value)

    blocks_list: List[torch.Tensor] = []
    for i in range(3):
        # We use indexing approach here:
        # https://numpy.org/doc/stable/user/basics.indexing.html#dealing-with-variable-numbers-of-indices-within-programs
        indices = [slice(0, None)] * x.ndim
        indices[block_dim] = slice(i, i + num_blocks)
        indices = tuple(indices)
        blocks_list.append(x[indices])
    # [batch_size, num_blocks, 3 * block_len, ...]
    return torch.cat(blocks_list, dim=sequence_dim)


def _make_3block_relative_position_ids(block_len: int) -> torch.Tensor:
    """Makes 3-blocked relative position ids for local attention."""
    position_ids = torch.arange(3 * block_len, dtype=torch.int32)
    center_position_ids = position_ids[block_len:-block_len]
    # [block_len, 3 * block_len]
    relative_position_ids = position_ids.unsqueeze(0) - center_position_ids.unsqueeze(1)
    return relative_position_ids


def _mask_local_attention_mask(local_attention_mask: torch.Tensor, block_len: int) -> torch.Tensor:
    """Mask local attention mask to enforce that tokens are not allowed to attend tokens farther than ``local_radius."""
    relative_position_ids = _make_3block_relative_position_ids(block_len)
    locality_mask = torch.abs(relative_position_ids) < block_len
    locality_mask = locality_mask[None, None, :, :]
    locality_mask = locality_mask.to(local_attention_mask.device)
    return torch.logical_and(local_attention_mask, locality_mask)


def _get_local_attention_mask(attention_mask: torch.Tensor, block_len: int, device: torch.device) -> torch.Tensor:
    """Prepare attention mask to be applied for a local attention."""
    # [batch_size, num_blocks, block_len]
    _blocked_attention_mask = _split_into_blocks(attention_mask, block_len, dim=1)
    # [batch_size, num_block, 3 * block_len]
    _3blocked_attention_mask = _concatenate_3_blocks(_blocked_attention_mask, block_dim=1, sequence_dim=2)

    _blocked_attention_mask = _blocked_attention_mask.unsqueeze(-1)
    _3blocked_attention_mask = _3blocked_attention_mask.unsqueeze(-2)
    # [batch_size, num_block, block_len, 3 * block_len]
    local_attention_mask = torch.logical_and(_blocked_attention_mask, _3blocked_attention_mask)
    local_attention_mask = _mask_local_attention_mask(local_attention_mask, block_len)
    # [batch_size, 1, num_block, block_len, 3 * block_len]
    return local_attention_mask.unsqueeze(1).to(device)


def _make_global_fixed_block_ids(
    attention_mask: torch.Tensor, global_block_size: int
) -> Tuple[torch.Tensor, torch.Tensor]:
    """Obtain the "fixed block" global id corresponding to each input token.

    This implementation is a simlified version of the original Flaxformr implementation adopted from:
    https://github.com/google/flaxformer/blob/main/flaxformer/architectures/longt5/long_attention.py.

    In our scenario, as we use this strategy only for a decoder, orphan tokens, i.e. those tokens which do not make for
    the whole fixed block, are assigned to the preceding block.

    Padding tokens from the original sequence are represented by -1.
    """
    batch_size, seq_len = attention_mask.shape[:2]

    def handle_orphan_tokens(block_ids: torch.Tensor) -> torch.Tensor:
        block_ends = (torch.arange(seq_len) % global_block_size) == global_block_size - 1
        block_ends = block_ends.to(block_ids.device)
        true_block_ends = torch.logical_and(block_ends, block_ids >= 0)
        full_blocks = true_block_ends.sum(-1).unsqueeze(-1).type(block_ids.dtype) - 1
        block_ids = torch.where(block_ids < full_blocks, block_ids, full_blocks)
        return block_ids

    fixed_block_mask = torch.ones_like(attention_mask, device=attention_mask.device) / global_block_size
    fixed_block_mask = torch.cumsum(fixed_block_mask, axis=1) - fixed_block_mask
    mask = torch.where(attention_mask != 0.0, 1.0, -1000.0).type(attention_mask.dtype)
    global_block_ids = torch.floor(mask + fixed_block_mask - 1.0).type(attention_mask.dtype)
    _global_block_ids_lower_bound = torch.tensor(-1, dtype=global_block_ids.dtype, device=global_block_ids.device)
    global_block_ids = torch.where(
        global_block_ids > _global_block_ids_lower_bound, global_block_ids, _global_block_ids_lower_bound
    )
    # set padding tokens to -1
    global_block_ids = (global_block_ids * attention_mask) + (attention_mask - 1)
    # [batch_size, seq_len]
    global_block_ids = handle_orphan_tokens(global_block_ids)
    num_globals = seq_len // global_block_size
    # [batch_size, seq_len // global_block_size]
    if num_globals > 0:
        _sequence_block_ids_max = torch.max(global_block_ids, dim=-1).values.repeat(num_globals, 1).transpose(0, 1)
    else:
        _sequence_block_ids_max = torch.zeros(
            batch_size, 0, dtype=global_block_ids.dtype, device=global_block_ids.device
        )
    global_segment_ids = torch.cumsum(torch.ones(batch_size, num_globals), dim=-1) - 1
    global_segment_ids = global_segment_ids.to(attention_mask.device)
    global_segment_ids = torch.where(global_segment_ids <= _sequence_block_ids_max, 1, 0)
    return global_block_ids.type(torch.int), global_segment_ids.type(torch.int)


def _make_side_relative_position_ids(attention_mask: torch.Tensor, global_block_size: int) -> torch.Tensor:
    """Create the relative position tensor for local -> global attention."""
    block_ids, global_segment_ids = _make_global_fixed_block_ids(attention_mask, global_block_size)
    global_seq_len = global_segment_ids.shape[-1]
    global_positions = torch.arange(global_seq_len, device=block_ids.device)
    side_relative_position = global_positions - block_ids[..., None]
    return side_relative_position.type(torch.int64)


def _create_global_aggregates(
    hidden_states: torch.Tensor, block_ids: torch.Tensor, global_seq_len: int
) -> torch.Tensor:
    """Compute individual block aggregates by summing over individual blocks."""
    # (batch..., seq_len, global_seq_len))
    block_ids = block_ids.where(
        block_ids >= 0, torch.tensor(global_seq_len, dtype=block_ids.dtype, device=block_ids.device)
    )
    one_hot_block_ids = nn.functional.one_hot(block_ids.type(torch.int64), global_seq_len + 1)[:, :, :-1]
    return torch.einsum("...nd,...ng->...gd", hidden_states, one_hot_block_ids.type(hidden_states.dtype))


# Copied from transformers.models.t5.modeling_t5.T5LayerNorm with T5->LongT5
class LongT5LayerNorm(nn.Module):
    def __init__(self, hidden_size, eps=1e-6):
        """
        Construct a layernorm module in the LongT5 style. No bias and no subtraction of mean.
        """
        super().__init__()
        self.weight = nn.Parameter(torch.ones(hidden_size))
        self.variance_epsilon = eps

    def forward(self, hidden_states):
        # LongT5 uses a layer_norm which only scales and doesn't shift, which is also known as Root Mean
        # Square Layer Normalization https://arxiv.org/abs/1910.07467 thus varience is calculated
        # w/o mean and there is no bias. Additionally we want to make sure that the accumulation for
        # half-precision inputs is done in fp32

        variance = hidden_states.to(torch.float32).pow(2).mean(-1, keepdim=True)
        hidden_states = hidden_states * torch.rsqrt(variance + self.variance_epsilon)

        # convert into half-precision if necessary
        if self.weight.dtype in [torch.float16, torch.bfloat16]:
            hidden_states = hidden_states.to(self.weight.dtype)

        return self.weight * hidden_states


try:
    from apex.normalization import FusedRMSNorm

    LongT5LayerNorm = FusedRMSNorm  # noqa

    logger.info("Discovered apex.normalization.FusedRMSNorm - will use it instead of LongT5LayerNorm")
except ImportError:
    # using the normal LongT5LayerNorm
    pass
except Exception:
    logger.warning("discovered apex but it failed to load, falling back to LongT5LayerNorm")
    pass

ALL_LAYERNORM_LAYERS.append(LongT5LayerNorm)


# Copied from transformers.models.t5.modeling_t5.T5DenseActDense with T5->LongT5
class LongT5DenseActDense(nn.Module):
    def __init__(self, config: LongT5Config):
        super().__init__()
        self.wi = nn.Linear(config.d_model, config.d_ff, bias=False)
        self.wo = nn.Linear(config.d_ff, config.d_model, bias=False)
        self.dropout = nn.Dropout(config.dropout_rate)
        self.act = ACT2FN[config.dense_act_fn]

    def forward(self, hidden_states):
        hidden_states = self.wi(hidden_states)
        hidden_states = self.act(hidden_states)
        hidden_states = self.dropout(hidden_states)
        if (
            isinstance(self.wo.weight, torch.Tensor)
            and hidden_states.dtype != self.wo.weight.dtype
            and self.wo.weight.dtype != torch.int8
        ):
            hidden_states = hidden_states.to(self.wo.weight.dtype)
        hidden_states = self.wo(hidden_states)
        return hidden_states


class LongT5DenseGatedActDense(nn.Module):
    def __init__(self, config: LongT5Config):
        super().__init__()
        self.wi_0 = nn.Linear(config.d_model, config.d_ff, bias=False)
        self.wi_1 = nn.Linear(config.d_model, config.d_ff, bias=False)
        self.wo = nn.Linear(config.d_ff, config.d_model, bias=False)
        self.dropout = nn.Dropout(config.dropout_rate)
        self.act = ACT2FN[config.dense_act_fn]

    def forward(self, hidden_states):
        hidden_gelu = self.act(self.wi_0(hidden_states))
        hidden_linear = self.wi_1(hidden_states)
        hidden_states = hidden_gelu * hidden_linear
        hidden_states = self.dropout(hidden_states)
        hidden_states = self.wo(hidden_states)
        return hidden_states


# Copied from transformers.models.t5.modeling_t5.T5LayerFF with T5->LongT5
class LongT5LayerFF(nn.Module):
    def __init__(self, config: LongT5Config):
        super().__init__()
        if config.is_gated_act:
            self.DenseReluDense = LongT5DenseGatedActDense(config)
        else:
            self.DenseReluDense = LongT5DenseActDense(config)

        self.layer_norm = LongT5LayerNorm(config.d_model, eps=config.layer_norm_epsilon)
        self.dropout = nn.Dropout(config.dropout_rate)

    def forward(self, hidden_states):
        forwarded_states = self.layer_norm(hidden_states)
        forwarded_states = self.DenseReluDense(forwarded_states)
        hidden_states = hidden_states + self.dropout(forwarded_states)
        return hidden_states


# Copied from transformers.models.t5.modeling_t5.T5Attention with T5->LongT5
class LongT5Attention(nn.Module):
    def __init__(self, config: LongT5Config, has_relative_attention_bias=False):
        super().__init__()
        self.is_decoder = config.is_decoder
        self.has_relative_attention_bias = has_relative_attention_bias
        self.relative_attention_num_buckets = config.relative_attention_num_buckets
        self.relative_attention_max_distance = config.relative_attention_max_distance
        self.d_model = config.d_model
        self.key_value_proj_dim = config.d_kv
        self.n_heads = config.num_heads
        self.dropout = config.dropout_rate
        self.inner_dim = self.n_heads * self.key_value_proj_dim

        # Mesh TensorFlow initialization to avoid scaling before softmax
        self.q = nn.Linear(self.d_model, self.inner_dim, bias=False)
        self.k = nn.Linear(self.d_model, self.inner_dim, bias=False)
        self.v = nn.Linear(self.d_model, self.inner_dim, bias=False)
        self.o = nn.Linear(self.inner_dim, self.d_model, bias=False)

        if self.has_relative_attention_bias:
            self.relative_attention_bias = nn.Embedding(self.relative_attention_num_buckets, self.n_heads)
        self.pruned_heads = set()
        self.gradient_checkpointing = False

    def prune_heads(self, heads):
        if len(heads) == 0:
            return
        heads, index = find_pruneable_heads_and_indices(
            heads, self.n_heads, self.key_value_proj_dim, self.pruned_heads
        )
        # Prune linear layers
        self.q = prune_linear_layer(self.q, index)
        self.k = prune_linear_layer(self.k, index)
        self.v = prune_linear_layer(self.v, index)
        self.o = prune_linear_layer(self.o, index, dim=1)
        # Update hyper params
        self.n_heads = self.n_heads - len(heads)
        self.inner_dim = self.key_value_proj_dim * self.n_heads
        self.pruned_heads = self.pruned_heads.union(heads)

    @staticmethod
    def _relative_position_bucket(relative_position, bidirectional=True, num_buckets=32, max_distance=128):
        """
        Adapted from Mesh Tensorflow:
        https://github.com/tensorflow/mesh/blob/0cb87fe07da627bf0b7e60475d59f95ed6b5be3d/mesh_tensorflow/transformer/transformer_layers.py#L593

        Translate relative position to a bucket number for relative attention. The relative position is defined as
        memory_position - query_position, i.e. the distance in tokens from the attending position to the attended-to
        position. If bidirectional=False, then positive relative positions are invalid. We use smaller buckets for
        small absolute relative_position and larger buckets for larger absolute relative_positions. All relative
        positions >=max_distance map to the same bucket. All relative positions <=-max_distance map to the same bucket.
        This should allow for more graceful generalization to longer sequences than the model has been trained on

        Args:
            relative_position: an int32 Tensor
            bidirectional: a boolean - whether the attention is bidirectional
            num_buckets: an integer
            max_distance: an integer

        Returns:
            a Tensor with the same shape as relative_position, containing int32 values in the range [0, num_buckets)
        """
        relative_buckets = 0
        if bidirectional:
            num_buckets //= 2
            relative_buckets += (relative_position > 0).to(torch.long) * num_buckets
            relative_position = torch.abs(relative_position)
        else:
            relative_position = -torch.min(relative_position, torch.zeros_like(relative_position))
        # now relative_position is in the range [0, inf)

        # half of the buckets are for exact increments in positions
        max_exact = num_buckets // 2
        is_small = relative_position < max_exact

        # The other half of the buckets are for logarithmically bigger bins in positions up to max_distance
        relative_position_if_large = max_exact + (
            torch.log(relative_position.float() / max_exact)
            / math.log(max_distance / max_exact)
            * (num_buckets - max_exact)
        ).to(torch.long)
        relative_position_if_large = torch.min(
            relative_position_if_large, torch.full_like(relative_position_if_large, num_buckets - 1)
        )

        relative_buckets += torch.where(is_small, relative_position, relative_position_if_large)
        return relative_buckets

    def compute_bias(self, query_length, key_length, device=None):
        """Compute binned relative position bias"""
        if device is None:
            device = self.relative_attention_bias.weight.device
        context_position = torch.arange(query_length, dtype=torch.long, device=device)[:, None]
        memory_position = torch.arange(key_length, dtype=torch.long, device=device)[None, :]
        relative_position = memory_position - context_position  # shape (query_length, key_length)
        relative_position_bucket = self._relative_position_bucket(
            relative_position,  # shape (query_length, key_length)
            bidirectional=(not self.is_decoder),
            num_buckets=self.relative_attention_num_buckets,
            max_distance=self.relative_attention_max_distance,
        )
        values = self.relative_attention_bias(relative_position_bucket)  # shape (query_length, key_length, num_heads)
        values = values.permute([2, 0, 1]).unsqueeze(0)  # shape (1, num_heads, query_length, key_length)
        return values

    def forward(
        self,
        hidden_states,
        mask=None,
        key_value_states=None,
        position_bias=None,
        past_key_value=None,
        layer_head_mask=None,
        query_length=None,
        use_cache=False,
        output_attentions=False,
    ):
        """
        Self-attention (if key_value_states is None) or attention over source sentence (provided by key_value_states).
        """
        # Input is (batch_size, seq_length, dim)
        # Mask is (batch_size, key_length) (non-causal) or (batch_size, key_length, key_length)
        # past_key_value[0] is (batch_size, n_heads, q_len - 1, dim_per_head)
        batch_size, seq_length = hidden_states.shape[:2]

        real_seq_length = seq_length

        if past_key_value is not None:
            if len(past_key_value) != 2:
                raise ValueError(
                    f"past_key_value should have 2 past states: keys and values. Got { len(past_key_value)} past states"
                )
            real_seq_length += past_key_value[0].shape[2] if query_length is None else query_length

        key_length = real_seq_length if key_value_states is None else key_value_states.shape[1]

        def shape(states):
            """projection"""
            return states.view(batch_size, -1, self.n_heads, self.key_value_proj_dim).transpose(1, 2)

        def unshape(states):
            """reshape"""
            return states.transpose(1, 2).contiguous().view(batch_size, -1, self.inner_dim)

        def project(hidden_states, proj_layer, key_value_states, past_key_value):
            """projects hidden states correctly to key/query states"""
            if key_value_states is None:
                # self-attn
                # (batch_size, n_heads, seq_length, dim_per_head)
                hidden_states = shape(proj_layer(hidden_states))
            elif past_key_value is None:
                # cross-attn
                # (batch_size, n_heads, seq_length, dim_per_head)
                hidden_states = shape(proj_layer(key_value_states))

            if past_key_value is not None:
                if key_value_states is None:
                    # self-attn
                    # (batch_size, n_heads, key_length, dim_per_head)
                    hidden_states = torch.cat([past_key_value, hidden_states], dim=2)
                elif past_key_value.shape[2] != key_value_states.shape[1]:
                    # checking that the `sequence_length` of the `past_key_value` is the same as
                    # the provided `key_value_states` to support prefix tuning
                    # cross-attn
                    # (batch_size, n_heads, seq_length, dim_per_head)
                    hidden_states = shape(proj_layer(key_value_states))
                else:
                    # cross-attn
                    hidden_states = past_key_value
            return hidden_states

        # get query states
        query_states = shape(self.q(hidden_states))  # (batch_size, n_heads, seq_length, dim_per_head)

        # get key/value states
        key_states = project(
            hidden_states, self.k, key_value_states, past_key_value[0] if past_key_value is not None else None
        )
        value_states = project(
            hidden_states, self.v, key_value_states, past_key_value[1] if past_key_value is not None else None
        )

        # compute scores
        scores = torch.matmul(
            query_states, key_states.transpose(3, 2)
        )  # equivalent of torch.einsum("bnqd,bnkd->bnqk", query_states, key_states), compatible with onnx op>9

        if position_bias is None:
            if not self.has_relative_attention_bias:
                position_bias = torch.zeros(
                    (1, self.n_heads, real_seq_length, key_length), device=scores.device, dtype=scores.dtype
                )
                if self.gradient_checkpointing and self.training:
                    position_bias.requires_grad = True
            else:
                position_bias = self.compute_bias(real_seq_length, key_length, device=scores.device)

            # if key and values are already calculated
            # we want only the last query position bias
            if past_key_value is not None:
                position_bias = position_bias[:, :, -hidden_states.size(1) :, :]

            if mask is not None:
                position_bias = position_bias + mask  # (batch_size, n_heads, seq_length, key_length)

        if self.pruned_heads:
            mask = torch.ones(position_bias.shape[1])
            mask[list(self.pruned_heads)] = 0
            position_bias_masked = position_bias[:, mask.bool()]
        else:
            position_bias_masked = position_bias

        scores += position_bias_masked
        attn_weights = nn.functional.softmax(scores.float(), dim=-1).type_as(
            scores
        )  # (batch_size, n_heads, seq_length, key_length)
        attn_weights = nn.functional.dropout(
            attn_weights, p=self.dropout, training=self.training
        )  # (batch_size, n_heads, seq_length, key_length)

        # Mask heads if we want to
        if layer_head_mask is not None:
            attn_weights = attn_weights * layer_head_mask

        attn_output = unshape(torch.matmul(attn_weights, value_states))  # (batch_size, seq_length, dim)
        attn_output = self.o(attn_output)

        present_key_value_state = (key_states, value_states) if (self.is_decoder and use_cache) else None
        outputs = (attn_output,) + (present_key_value_state,) + (position_bias,)

        if output_attentions:
            outputs = outputs + (attn_weights,)
        return outputs


class LongT5LocalAttention(nn.Module):
    def __init__(self, config: LongT5Config, has_relative_attention_bias: bool = False) -> None:
        super().__init__()
        self.is_decoder = config.is_decoder
        self.has_relative_attention_bias = has_relative_attention_bias
        self.relative_attention_num_buckets = config.relative_attention_num_buckets
        self.relative_attention_max_distance = config.relative_attention_max_distance
        self.d_model = config.d_model
        self.key_value_proj_dim = config.d_kv
        self.n_heads = config.num_heads
        self.local_radius = config.local_radius
        self.block_len = self.local_radius + 1
        self.dropout = config.dropout_rate
        self.inner_dim = self.n_heads * self.key_value_proj_dim

        # Mesh TensorFlow initialization to avoid scaling before softmax
        self.q = nn.Linear(self.d_model, self.inner_dim, bias=False)
        self.k = nn.Linear(self.d_model, self.inner_dim, bias=False)
        self.v = nn.Linear(self.d_model, self.inner_dim, bias=False)
        self.o = nn.Linear(self.inner_dim, self.d_model, bias=False)

        if self.has_relative_attention_bias:
            self.relative_attention_bias = nn.Embedding(self.relative_attention_num_buckets, self.n_heads)
        self.pruned_heads = set()
        self.gradient_checkpointing = False

    # Copied from transformers.models.t5.modeling_t5.T5Attention.prune_heads
    def prune_heads(self, heads):
        if len(heads) == 0:
            return
        heads, index = find_pruneable_heads_and_indices(
            heads, self.n_heads, self.key_value_proj_dim, self.pruned_heads
        )
        # Prune linear layers
        self.q = prune_linear_layer(self.q, index)
        self.k = prune_linear_layer(self.k, index)
        self.v = prune_linear_layer(self.v, index)
        self.o = prune_linear_layer(self.o, index, dim=1)
        # Update hyper params
        self.n_heads = self.n_heads - len(heads)
        self.inner_dim = self.key_value_proj_dim * self.n_heads
        self.pruned_heads = self.pruned_heads.union(heads)

    @staticmethod
    # Copied from transformers.models.t5.modeling_t5.T5Attention._relative_position_bucket
    def _relative_position_bucket(relative_position, bidirectional=True, num_buckets=32, max_distance=128):
        """
        Adapted from Mesh Tensorflow:
        https://github.com/tensorflow/mesh/blob/0cb87fe07da627bf0b7e60475d59f95ed6b5be3d/mesh_tensorflow/transformer/transformer_layers.py#L593

        Translate relative position to a bucket number for relative attention. The relative position is defined as
        memory_position - query_position, i.e. the distance in tokens from the attending position to the attended-to
        position. If bidirectional=False, then positive relative positions are invalid. We use smaller buckets for
        small absolute relative_position and larger buckets for larger absolute relative_positions. All relative
        positions >=max_distance map to the same bucket. All relative positions <=-max_distance map to the same bucket.
        This should allow for more graceful generalization to longer sequences than the model has been trained on

        Args:
            relative_position: an int32 Tensor
            bidirectional: a boolean - whether the attention is bidirectional
            num_buckets: an integer
            max_distance: an integer

        Returns:
            a Tensor with the same shape as relative_position, containing int32 values in the range [0, num_buckets)
        """
        relative_buckets = 0
        if bidirectional:
            num_buckets //= 2
            relative_buckets += (relative_position > 0).to(torch.long) * num_buckets
            relative_position = torch.abs(relative_position)
        else:
            relative_position = -torch.min(relative_position, torch.zeros_like(relative_position))
        # now relative_position is in the range [0, inf)

        # half of the buckets are for exact increments in positions
        max_exact = num_buckets // 2
        is_small = relative_position < max_exact

        # The other half of the buckets are for logarithmically bigger bins in positions up to max_distance
        relative_position_if_large = max_exact + (
            torch.log(relative_position.float() / max_exact)
            / math.log(max_distance / max_exact)
            * (num_buckets - max_exact)
        ).to(torch.long)
        relative_position_if_large = torch.min(
            relative_position_if_large, torch.full_like(relative_position_if_large, num_buckets - 1)
        )

        relative_buckets += torch.where(is_small, relative_position, relative_position_if_large)
        return relative_buckets

    def compute_bias(self, block_length: int):
        """Compute binned relative position bias"""
        target_device = (
            self.relative_attention_bias.weight.device
            if self.relative_attention_bias.weight.device.type != "meta"
            else None
        )
        memory_position = torch.arange(3 * block_length, dtype=torch.long, device=target_device)
        context_position = memory_position[block_length:-block_length]

        # (block_length, 3 * block_length)
        relative_position = memory_position[None, :] - context_position[:, None]
        relative_position_bucket = self._relative_position_bucket(
            relative_position,  # (block_length, 3 * block_length)
            bidirectional=(not self.is_decoder),
            num_buckets=self.relative_attention_num_buckets,
            max_distance=self.relative_attention_max_distance,
        )
        # (block_length, 3 * block_length, num_heads)
        values = self.relative_attention_bias(relative_position_bucket)
        # (1, 1, num_heads, block_length, 3 * block_length)
        values = values.permute([2, 0, 1]).unsqueeze(0).unsqueeze(0)
        return values

    def forward(
        self,
        hidden_states,
        mask=None,
        position_bias=None,
        layer_head_mask=None,
        output_attentions=False,
    ):
        batch_size, seq_length = hidden_states.shape[:2]

        def shape(states):
            """projection"""
            return states.view(batch_size, -1, self.n_heads, self.key_value_proj_dim)

        def unshape(states):
            """reshape"""
            return states.contiguous().view(batch_size, -1, self.inner_dim)

        # get query/key/value states -> (batch_size, seq_length, n_heads, dim_per_head)
        query_states = shape(self.q(hidden_states))
        key_states = shape(self.k(hidden_states))
        value_states = shape(self.v(hidden_states))

        # Split into blocks -> (batch_size, num_blocks, block_len, n_heads, dim_per_head)
        query_states = _split_into_blocks(query_states, self.block_len, dim=1)
        key_states = _split_into_blocks(key_states, self.block_len, dim=1)
        value_states = _split_into_blocks(value_states, self.block_len, dim=1)

        # Concatenate 3 blocks for keys and values -> (batch_size, num_blocks, 3 * block_len, n_heads, dim_per_head)
        key_states = _concatenate_3_blocks(key_states, block_dim=1, sequence_dim=2)
        value_states = _concatenate_3_blocks(value_states, block_dim=1, sequence_dim=2)

        # Compute scores
        scores = torch.einsum(
            "...qhd,...khd->...hqk", query_states, key_states
        )  # (batch_size, num_block, n_heads, block_len, 3 * block_len)

        if position_bias is None:
            # position_bias shape: # (1, 1, n_heads, block_len, 3 * block_len)
            if not self.has_relative_attention_bias:
                position_bias = torch.zeros(
                    (1, 1, self.n_heads, self.block_len, 3 * self.block_len), device=scores.device, dtype=scores.dtype
                )
                if self.gradient_checkpointing and self.training:
                    position_bias.requires_grad = True
            else:
                position_bias = self.compute_bias(self.block_len)

            if mask is not None:
                # Replace masked positions with -1e10 (according to the original implementation)
                mask = torch.where(mask > 0, 0.0, -1e10)
                # We need to adjust position bias shape to be sum with mask
                position_bias = position_bias + mask.transpose(1, 2)

        scores += position_bias
        # (batch_size, num_blocks, n_heads, block_len, 3 * block_len)
        attn_weights = nn.functional.softmax(scores.float(), dim=-1).type_as(scores)
        # (batch_size, num_blocks, n_heads, block_len, 3 * block_len)
        attn_weights = nn.functional.dropout(attn_weights, p=self.dropout, training=self.training)

        # Mask heads if we want to
        if layer_head_mask is not None:
            attn_weights = attn_weights * layer_head_mask
        attn_weights = attn_weights.type(value_states.dtype)
        attn_output = unshape(torch.einsum("...hqk,...khd->...qhd", attn_weights, value_states))
        attn_output = attn_output[:, :seq_length, :]
        attn_output = self.o(attn_output)

        present_key_value_state = None
        outputs = (attn_output,) + (present_key_value_state,) + (position_bias,)

        if output_attentions:
            outputs = outputs + (attn_weights,)
        return outputs


class LongT5TransientGlobalAttention(nn.Module):
    def __init__(self, config: LongT5Config, has_relative_attention_bias: bool = False) -> None:
        super().__init__()
        self.is_decoder = config.is_decoder
        self.has_relative_attention_bias = has_relative_attention_bias
        self.relative_attention_num_buckets = config.relative_attention_num_buckets
        self.relative_attention_max_distance = config.relative_attention_max_distance
        self.d_model = config.d_model
        self.key_value_proj_dim = config.d_kv
        self.n_heads = config.num_heads
        self.local_radius = config.local_radius
        self.block_len = self.local_radius + 1
        self.global_block_size = config.global_block_size
        self.dropout = config.dropout_rate
        self.inner_dim = self.n_heads * self.key_value_proj_dim

        # Mesh TensorFlow initialization to avoid scaling before softmax
        self.q = nn.Linear(self.d_model, self.inner_dim, bias=False)
        self.k = nn.Linear(self.d_model, self.inner_dim, bias=False)
        self.v = nn.Linear(self.d_model, self.inner_dim, bias=False)
        self.o = nn.Linear(self.inner_dim, self.d_model, bias=False)

        if self.has_relative_attention_bias:
            self.relative_attention_bias = nn.Embedding(self.relative_attention_num_buckets, self.n_heads)
        self.pruned_heads = set()

        # Relativen attention bias & Layer norm for global attention
        if self.has_relative_attention_bias:
            self.global_relative_attention_bias = nn.Embedding(self.relative_attention_num_buckets, self.n_heads)
        self.global_input_layer_norm = LongT5LayerNorm(config.d_model, eps=config.layer_norm_epsilon)

    # Copied from transformers.models.t5.modeling_t5.T5Attention.prune_heads
    def prune_heads(self, heads):
        if len(heads) == 0:
            return
        heads, index = find_pruneable_heads_and_indices(
            heads, self.n_heads, self.key_value_proj_dim, self.pruned_heads
        )
        # Prune linear layers
        self.q = prune_linear_layer(self.q, index)
        self.k = prune_linear_layer(self.k, index)
        self.v = prune_linear_layer(self.v, index)
        self.o = prune_linear_layer(self.o, index, dim=1)
        # Update hyper params
        self.n_heads = self.n_heads - len(heads)
        self.inner_dim = self.key_value_proj_dim * self.n_heads
        self.pruned_heads = self.pruned_heads.union(heads)

    @staticmethod
    # Copied from transformers.models.t5.modeling_t5.T5Attention._relative_position_bucket
    def _relative_position_bucket(relative_position, bidirectional=True, num_buckets=32, max_distance=128):
        """
        Adapted from Mesh Tensorflow:
        https://github.com/tensorflow/mesh/blob/0cb87fe07da627bf0b7e60475d59f95ed6b5be3d/mesh_tensorflow/transformer/transformer_layers.py#L593

        Translate relative position to a bucket number for relative attention. The relative position is defined as
        memory_position - query_position, i.e. the distance in tokens from the attending position to the attended-to
        position. If bidirectional=False, then positive relative positions are invalid. We use smaller buckets for
        small absolute relative_position and larger buckets for larger absolute relative_positions. All relative
        positions >=max_distance map to the same bucket. All relative positions <=-max_distance map to the same bucket.
        This should allow for more graceful generalization to longer sequences than the model has been trained on

        Args:
            relative_position: an int32 Tensor
            bidirectional: a boolean - whether the attention is bidirectional
            num_buckets: an integer
            max_distance: an integer

        Returns:
            a Tensor with the same shape as relative_position, containing int32 values in the range [0, num_buckets)
        """
        relative_buckets = 0
        if bidirectional:
            num_buckets //= 2
            relative_buckets += (relative_position > 0).to(torch.long) * num_buckets
            relative_position = torch.abs(relative_position)
        else:
            relative_position = -torch.min(relative_position, torch.zeros_like(relative_position))
        # now relative_position is in the range [0, inf)

        # half of the buckets are for exact increments in positions
        max_exact = num_buckets // 2
        is_small = relative_position < max_exact

        # The other half of the buckets are for logarithmically bigger bins in positions up to max_distance
        relative_position_if_large = max_exact + (
            torch.log(relative_position.float() / max_exact)
            / math.log(max_distance / max_exact)
            * (num_buckets - max_exact)
        ).to(torch.long)
        relative_position_if_large = torch.min(
            relative_position_if_large, torch.full_like(relative_position_if_large, num_buckets - 1)
        )

        relative_buckets += torch.where(is_small, relative_position, relative_position_if_large)
        return relative_buckets

    def compute_bias(self, block_length: int):
        """Compute binned relative position bias"""
        target_device = (
            self.relative_attention_bias.weight.device
            if self.relative_attention_bias.weight.device.type != "meta"
            else None
        )
        memory_position = torch.arange(3 * block_length, dtype=torch.long, device=target_device)
        context_position = memory_position[block_length:-block_length]

        # (block_length, 3 * block_length)
        relative_position = memory_position[None, :] - context_position[:, None]
        relative_position_bucket = self._relative_position_bucket(
            relative_position,  # (block_length, 3 * block_length)
            bidirectional=(not self.is_decoder),
            num_buckets=self.relative_attention_num_buckets,
            max_distance=self.relative_attention_max_distance,
        )
        # (block_length, 3 * block_length, num_heads)
        values = self.relative_attention_bias(relative_position_bucket)
        # (1, 1, num_heads, block_length, 3 * block_length)
        values = values.permute([2, 0, 1]).unsqueeze(0).unsqueeze(0)
        return values

    def compute_side_bias(self, mask: torch.Tensor, global_segment_ids: torch.Tensor) -> torch.Tensor:
        # (batch_size, 1, seq_len, global_seq_len)
        side_attention_mask = torch.eq(mask[..., None], global_segment_ids[:, None, :])[:, None, ...]
        attention_side_bias = torch.where(side_attention_mask > 0, 0.0, -1e10)
        # (batch_size, seq_len, global_seq_len)
        side_relative_position = _make_side_relative_position_ids(mask, self.global_block_size)
        side_relative_position_bucket = self._relative_position_bucket(
            side_relative_position,
            bidirectional=(not self.is_decoder),
            num_buckets=self.relative_attention_num_buckets,
            max_distance=self.relative_attention_max_distance,
        )
        # (batch_size, seq_len, global_seq_len, num_heads)
        side_bias = self.global_relative_attention_bias(side_relative_position_bucket)

        # (batch_size, num_heads, seq_len, global_seq_len)
        side_bias = side_bias.permute([0, 3, 1, 2])
        # (batch_size, num_heads, seq_len, global_seq_len)
        attention_side_bias = attention_side_bias + side_bias
        return attention_side_bias

    def forward(
        self,
        hidden_states,
        mask=None,
        position_bias=None,
        layer_head_mask=None,
        output_attentions=False,
    ):
        batch_size, seq_length = hidden_states.shape[:2]

        def shape(states):
            """projection"""
            return states.view(batch_size, -1, self.n_heads, self.key_value_proj_dim)

        def unshape(states):
            """reshape"""
            return states.contiguous().view(batch_size, -1, self.inner_dim)

        # Prepare components for transient-global attention
        # Obtain block_ids and global_segment_ids
        # global_seq_len := seq_len // self.global_block_size
        # shapes: (batch_size, seq_len) & (batch_size, global_seq_len)
        block_ids, global_segment_ids = _make_global_fixed_block_ids(
            mask if mask is not None else torch.ones(hidden_states.shape[:-1]),
            self.global_block_size,
        )
        # Create global inputs
        _global_seq_len = global_segment_ids.shape[-1]
        global_inputs = _create_global_aggregates(hidden_states, block_ids, _global_seq_len)
        global_inputs = self.global_input_layer_norm(global_inputs)

        # get query states -> (batch_size, seq_length, n_heads, dim_per_head)
        query_states = shape(self.q(hidden_states))
        key_states = shape(self.k(hidden_states))
        value_states = shape(self.v(hidden_states))
        # Get global/side key/value states  shape: (batch_size, global_seq_len, n_heads, dim_per_head)
        side_key_states = shape(self.k(global_inputs))
        side_value_states = shape(self.v(global_inputs))

        # Split into blocks -> (batch_size, num_blocks, block_len, n_heads, dim_per_head)
        query_states = _split_into_blocks(query_states, self.block_len, dim=1)
        key_states = _split_into_blocks(key_states, self.block_len, dim=1)
        value_states = _split_into_blocks(value_states, self.block_len, dim=1)

        # Concatenate 3 blocks for keys and values -> (batch_size, num_blocks, 3 * block_len, n_heads, dim_per_head)
        key_states = _concatenate_3_blocks(key_states, block_dim=1, sequence_dim=2)
        value_states = _concatenate_3_blocks(value_states, block_dim=1, sequence_dim=2)

        # Tile side inputs across local key/value blocks
        # New shape: (batch_size, num_blocks, global_seq_len, n_heads, dim_per_head)
        reps = [1] * (side_key_states.ndim + 1)
        reps[1] = key_states.shape[1]
        side_key_states = side_key_states.unsqueeze(1).repeat(reps)
        side_value_states = side_value_states.unsqueeze(1).repeat(reps)

        # Concatenate "local" and "side"/"global" key/value states to allow each token to attend global aggregated ones
        # New shape: (batch_size, num_blocks, 3 * block_len + global_seq_len, n_heads, dim_per_head)
        key_states = torch.cat([key_states, side_key_states], dim=2)
        value_states = torch.cat([value_states, side_value_states], dim=2)

        # Compute scores -> (batch_size, num_block, n_heads, block_len, 3 * block_len + global_seq_len)
        scores = torch.einsum("...qhd,...khd->...hqk", query_states, key_states)

        if mask is not None:
            # We need to adjust position bias shape to be sum with mask
            local_attention_mask = _get_local_attention_mask(mask, self.block_len, hidden_states.device)
            # Replace masked positions with -10_000 (according to the original implementation)
            local_attention_mask = torch.where(local_attention_mask > 0, 0.0, -1e10)
        else:
            local_attention_mask = None

        if position_bias is None:
            # position_bias shape: # (1, 1, n_heads, block_len, 3 * block_len)
            if not self.has_relative_attention_bias:
                position_bias = torch.zeros(
                    (1, 1, self.n_heads, self.block_len, 3 * self.block_len),
                    device=scores.device,
                    dtype=scores.dtype,
                )
                if self.gradient_checkpointing and self.training:
                    position_bias.requires_grad = True
            else:
                position_bias = self.compute_bias(self.block_len)

            if local_attention_mask is not None:
                # (batch_size, 1, n_heads, block_len, 3 * block_len)
                position_bias = position_bias + local_attention_mask.transpose(1, 2)
            position_bias = position_bias.type(scores.dtype)

            # Calculate global/side bias - shape: # (batch_size, num_heads, seq_len, global_seq_len)
            if mask is None:
                mask = torch.ones(batch_size, seq_length)
            # (batch_size, num_heads, seq_len, global_seq_len)
            side_position_bias = self.compute_side_bias(mask, global_segment_ids)
            # (batch_size, num_blocks, num_heads, block_len, global_seq_len)
            side_position_bias = _split_into_blocks(side_position_bias, self.block_len, dim=-2).transpose(1, 2)
            side_position_bias = side_position_bias.type(scores.dtype).to(scores.device)
            # (batch_size, num_blocks, num_heads, block_len, 3 * block_len + global_seq_len)
            position_bias = torch.cat([position_bias, side_position_bias], dim=-1)

        scores += position_bias
        # (batch_size, num_blocks, n_heads, block_len, 3 * block_len + global_seq_len)
        attn_weights = nn.functional.softmax(scores.float(), dim=-1).type_as(scores)
        attn_weights = nn.functional.dropout(attn_weights, p=self.dropout, training=self.training)

        # Mask heads if we want to
        if layer_head_mask is not None:
            attn_weights = attn_weights * layer_head_mask
        attn_weights = attn_weights.type(value_states.dtype)
        attn_output = unshape(torch.einsum("...hqk,...khd->...qhd", attn_weights, value_states))
        attn_output = attn_output[:, :seq_length, :]
        attn_output = self.o(attn_output)

        present_key_value_state = None
        outputs = (attn_output,) + (present_key_value_state,) + (position_bias,)

        if output_attentions:
            outputs = outputs + (attn_weights,)
        return outputs


# Copied from transformers.models.t5.modeling_t5.T5LayerSelfAttention with T5->LongT5
class LongT5LayerSelfAttention(nn.Module):
    def __init__(self, config, has_relative_attention_bias=False):
        super().__init__()
        self.SelfAttention = LongT5Attention(config, has_relative_attention_bias=has_relative_attention_bias)
        self.layer_norm = LongT5LayerNorm(config.d_model, eps=config.layer_norm_epsilon)
        self.dropout = nn.Dropout(config.dropout_rate)

    def forward(
        self,
        hidden_states,
        attention_mask=None,
        position_bias=None,
        layer_head_mask=None,
        past_key_value=None,
        use_cache=False,
        output_attentions=False,
    ):
        normed_hidden_states = self.layer_norm(hidden_states)
        attention_output = self.SelfAttention(
            normed_hidden_states,
            mask=attention_mask,
            position_bias=position_bias,
            layer_head_mask=layer_head_mask,
            past_key_value=past_key_value,
            use_cache=use_cache,
            output_attentions=output_attentions,
        )
        hidden_states = hidden_states + self.dropout(attention_output[0])
        outputs = (hidden_states,) + attention_output[1:]  # add attentions if we output them
        return outputs


class LongT5LayerLocalSelfAttention(nn.Module):
    """Local self attention used in encoder"""

    def __init__(self, config, has_relative_attention_bias=False):
        super().__init__()
        self.LocalSelfAttention = LongT5LocalAttention(config, has_relative_attention_bias=has_relative_attention_bias)
        self.layer_norm = LongT5LayerNorm(config.d_model, eps=config.layer_norm_epsilon)
        self.dropout = nn.Dropout(config.dropout_rate)

    def forward(
        self,
        hidden_states,
        attention_mask=None,
        position_bias=None,
        layer_head_mask=None,
        output_attentions=False,
        **kwargs: Any,  # to accept past_key_value and use_cache kwargs
    ):
        normed_hidden_states = self.layer_norm(hidden_states)
        attention_output = self.LocalSelfAttention(
            normed_hidden_states,
            mask=attention_mask,
            position_bias=position_bias,
            layer_head_mask=layer_head_mask,
            output_attentions=output_attentions,
        )
        hidden_states = hidden_states + self.dropout(attention_output[0])
        outputs = (hidden_states,) + attention_output[1:]  # add attentions if we output them
        return outputs


class LongT5LayerTransientGlobalSelfAttention(nn.Module):
    """Transient-Global self attention used in encoder"""

    def __init__(self, config, has_relative_attention_bias=False):
        super().__init__()
        self.TransientGlobalSelfAttention = LongT5TransientGlobalAttention(
            config, has_relative_attention_bias=has_relative_attention_bias
        )
        self.layer_norm = LongT5LayerNorm(config.d_model, eps=config.layer_norm_epsilon)
        self.dropout = nn.Dropout(config.dropout_rate)

    def forward(
        self,
        hidden_states,
        attention_mask=None,
        position_bias=None,
        layer_head_mask=None,
        output_attentions=False,
        **kwargs: Any,  # to accept past_key_value and use_cache kwargs
    ):
        normed_hidden_states = self.layer_norm(hidden_states)
        attention_output = self.TransientGlobalSelfAttention(
            normed_hidden_states,
            mask=attention_mask,
            position_bias=position_bias,
            layer_head_mask=layer_head_mask,
            output_attentions=output_attentions,
        )
        hidden_states = hidden_states + self.dropout(attention_output[0])
        outputs = (hidden_states,) + attention_output[1:]  # add attentions if we output them
        return outputs


# Copied from transformers.models.t5.modeling_t5.T5LayerCrossAttention with T5->LongT5
class LongT5LayerCrossAttention(nn.Module):
    def __init__(self, config):
        super().__init__()
        self.EncDecAttention = LongT5Attention(config, has_relative_attention_bias=False)
        self.layer_norm = LongT5LayerNorm(config.d_model, eps=config.layer_norm_epsilon)
        self.dropout = nn.Dropout(config.dropout_rate)

    def forward(
        self,
        hidden_states,
        key_value_states,
        attention_mask=None,
        position_bias=None,
        layer_head_mask=None,
        past_key_value=None,
        use_cache=False,
        query_length=None,
        output_attentions=False,
    ):
        normed_hidden_states = self.layer_norm(hidden_states)
        attention_output = self.EncDecAttention(
            normed_hidden_states,
            mask=attention_mask,
            key_value_states=key_value_states,
            position_bias=position_bias,
            layer_head_mask=layer_head_mask,
            past_key_value=past_key_value,
            use_cache=use_cache,
            query_length=query_length,
            output_attentions=output_attentions,
        )
        layer_output = hidden_states + self.dropout(attention_output[0])
        outputs = (layer_output,) + attention_output[1:]  # add attentions if we output them
        return outputs


class LongT5Block(nn.Module):
    def __init__(self, config, has_relative_attention_bias=False):
        super().__init__()
        self.is_decoder = config.is_decoder
        if config.is_decoder:
            attention_layer = LongT5LayerSelfAttention
        elif config.encoder_attention_type == "local":
            attention_layer = LongT5LayerLocalSelfAttention
        elif config.encoder_attention_type == "transient-global":
            attention_layer = LongT5LayerTransientGlobalSelfAttention
        else:
            raise ValueError(
                "For encoder attention mechanism, either `local` or `transient-global` attention type is expected, "
                f"but got {config.encoder_attention_type}."
            )
        self.layer = nn.ModuleList()
        self.layer.append(attention_layer(config, has_relative_attention_bias=has_relative_attention_bias))
        if self.is_decoder:
            self.layer.append(LongT5LayerCrossAttention(config))

        self.layer.append(LongT5LayerFF(config))

    def forward(
        self,
        hidden_states,
        attention_mask=None,
        position_bias=None,
        encoder_hidden_states=None,
        encoder_attention_mask=None,
        encoder_decoder_position_bias=None,
        layer_head_mask=None,
        cross_attn_layer_head_mask=None,
        past_key_value=None,
        use_cache=False,
        output_attentions=False,
        return_dict=True,
    ):
        if past_key_value is not None:
            if not self.is_decoder:
                logger.warning("`past_key_values` is passed to the encoder. Please make sure this is intended.")
            expected_num_past_key_values = 2 if encoder_hidden_states is None else 4

            if len(past_key_value) != expected_num_past_key_values:
                raise ValueError(
                    f"There should be {expected_num_past_key_values} past states. "
                    f"{'2 (past / key) for cross attention. ' if expected_num_past_key_values == 4 else ''}"
                    f"Got {len(past_key_value)} past key / value states"
                )

            self_attn_past_key_value = past_key_value[:2]
            cross_attn_past_key_value = past_key_value[2:]
        else:
            self_attn_past_key_value, cross_attn_past_key_value = None, None

        self_attention_outputs = self.layer[0](
            hidden_states,
            attention_mask=attention_mask,
            position_bias=position_bias,
            layer_head_mask=layer_head_mask,
            past_key_value=self_attn_past_key_value,
            use_cache=use_cache,
            output_attentions=output_attentions,
        )
        hidden_states, present_key_value_state = self_attention_outputs[:2]
        attention_outputs = self_attention_outputs[2:]  # Keep self-attention outputs and relative position weights

        # clamp inf values to enable fp16 inference - check https://github.com/huggingface/transformers/pull/19229/
        if hidden_states.dtype == torch.float16 and torch.isinf(hidden_states).any():
            clamp_value = torch.finfo(hidden_states.dtype).max - 1000
            hidden_states = torch.clamp(hidden_states, min=-clamp_value, max=clamp_value)

        do_cross_attention = self.is_decoder and encoder_hidden_states is not None
        if do_cross_attention:
            # the actual query length is unknown for cross attention
            # if using past key value states. Need to inject it here
            if present_key_value_state is not None:
                query_length = present_key_value_state[0].shape[2]
            else:
                query_length = None

            cross_attention_outputs = self.layer[1](
                hidden_states,
                key_value_states=encoder_hidden_states,
                attention_mask=encoder_attention_mask,
                position_bias=encoder_decoder_position_bias,
                layer_head_mask=cross_attn_layer_head_mask,
                past_key_value=cross_attn_past_key_value,
                query_length=query_length,
                use_cache=use_cache,
                output_attentions=output_attentions,
            )
            hidden_states = cross_attention_outputs[0]

            # clamp inf values to enable fp16 inference - check https://github.com/huggingface/transformers/pull/19229/
            if hidden_states.dtype == torch.float16 and torch.isinf(hidden_states).any():
                clamp_value = torch.finfo(hidden_states.dtype).max - 1000
                hidden_states = torch.clamp(hidden_states, min=-clamp_value, max=clamp_value)

            # Combine self attn and cross attn key value states
            if present_key_value_state is not None:
                present_key_value_state = present_key_value_state + cross_attention_outputs[1]

            # Keep cross-attention outputs and relative position weights
            attention_outputs = attention_outputs + cross_attention_outputs[2:]

        # Apply Feed Forward layer
        hidden_states = self.layer[-1](hidden_states)

        # clamp inf values to enable fp16 inference - check https://github.com/huggingface/transformers/pull/19229/
        if hidden_states.dtype == torch.float16 and torch.isinf(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 use_cache:
            outputs = outputs + (present_key_value_state,) + attention_outputs
        else:
            outputs = outputs + attention_outputs

        return outputs  # hidden-states, present_key_value_states, (self-attention position bias), (self-attention weights), (cross-attention position bias), (cross-attention weights)


class LongT5PreTrainedModel(PreTrainedModel):
    """
    An abstract class to handle weights initialization and a simple interface for downloading and loading pretrained
    models.
    """

    config_class = LongT5Config
    base_model_prefix = "transformer"
    supports_gradient_checkpointing = True
    _no_split_modules = ["LongT5Block"]

    @property
    # Copied from transformers.models.t5.modeling_t5.T5PreTrainedModel.dummy_inputs
    def dummy_inputs(self):
        input_ids = torch.tensor(DUMMY_INPUTS)
        input_mask = torch.tensor(DUMMY_MASK)
        dummy_inputs = {
            "decoder_input_ids": input_ids,
            "input_ids": input_ids,
            "decoder_attention_mask": input_mask,
        }
        return dummy_inputs

    def _init_weights(self, module):
        """Initialize the weights"""
        factor = self.config.initializer_factor  # Used for testing weights initialization
        if isinstance(module, LongT5LayerNorm):
            module.weight.data.fill_(factor * 1.0)
        elif isinstance(module, (LongT5Model, LongT5ForConditionalGeneration, LongT5EncoderModel)):
            # Mesh TensorFlow embeddings initialization
            # See https://github.com/tensorflow/mesh/blob/fa19d69eafc9a482aff0b59ddd96b025c0cb207d/mesh_tensorflow/layers.py#L1624
            module.shared.weight.data.normal_(mean=0.0, std=factor * 1.0)
            if hasattr(module, "lm_head") and not self.config.tie_word_embeddings:
                module.lm_head.weight.data.normal_(mean=0.0, std=factor * 1.0)
        elif isinstance(module, LongT5DenseActDense):
            # Mesh TensorFlow FF initialization
            # See https://github.com/tensorflow/mesh/blob/master/mesh_tensorflow/transformer/transformer_layers.py#L56
            # and https://github.com/tensorflow/mesh/blob/fa19d69eafc9a482aff0b59ddd96b025c0cb207d/mesh_tensorflow/layers.py#L89
            module.wi.weight.data.normal_(mean=0.0, std=factor * ((self.config.d_model) ** -0.5))
            if hasattr(module.wi, "bias") and module.wi.bias is not None:
                module.wi.bias.data.zero_()
            module.wo.weight.data.normal_(mean=0.0, std=factor * ((self.config.d_ff) ** -0.5))
            if hasattr(module.wo, "bias") and module.wo.bias is not None:
                module.wo.bias.data.zero_()
        elif isinstance(module, LongT5DenseGatedActDense):
            module.wi_0.weight.data.normal_(mean=0.0, std=factor * ((self.config.d_model) ** -0.5))
            if hasattr(module.wi_0, "bias") and module.wi_0.bias is not None:
                module.wi_0.bias.data.zero_()
            module.wi_1.weight.data.normal_(mean=0.0, std=factor * ((self.config.d_model) ** -0.5))
            if hasattr(module.wi_1, "bias") and module.wi_1.bias is not None:
                module.wi_1.bias.data.zero_()
            module.wo.weight.data.normal_(mean=0.0, std=factor * ((self.config.d_ff) ** -0.5))
            if hasattr(module.wo, "bias") and module.wo.bias is not None:
                module.wo.bias.data.zero_()
        elif isinstance(module, (LongT5Attention, LongT5LocalAttention, LongT5TransientGlobalAttention)):
            # Mesh TensorFlow attention initialization to avoid scaling before softmax
            # See https://github.com/tensorflow/mesh/blob/fa19d69eafc9a482aff0b59ddd96b025c0cb207d/mesh_tensorflow/transformer/attention.py#L136
            d_model = self.config.d_model
            key_value_proj_dim = self.config.d_kv
            n_heads = self.config.num_heads
            module.q.weight.data.normal_(mean=0.0, std=factor * ((d_model * key_value_proj_dim) ** -0.5))
            module.k.weight.data.normal_(mean=0.0, std=factor * (d_model**-0.5))
            module.v.weight.data.normal_(mean=0.0, std=factor * (d_model**-0.5))
            module.o.weight.data.normal_(mean=0.0, std=factor * ((n_heads * key_value_proj_dim) ** -0.5))
            if module.has_relative_attention_bias:
                module.relative_attention_bias.weight.data.normal_(mean=0.0, std=factor * ((d_model) ** -0.5))
                if isinstance(module, LongT5TransientGlobalAttention):
                    module.global_relative_attention_bias.weight.data.normal_(
                        mean=0.0, std=factor * ((d_model) ** -0.5)
                    )

    # Copied from transformers.models.t5.modeling_t5.T5PreTrainedModel._shift_right with T5->LongT5
    def _shift_right(self, input_ids):
        decoder_start_token_id = self.config.decoder_start_token_id
        pad_token_id = self.config.pad_token_id

        if decoder_start_token_id is None:
            raise ValueError(
                "self.model.config.decoder_start_token_id has to be defined. In LongT5 it is usually set to the pad_token_id. "
                "See LongT5 docs for more information."
            )

        # shift inputs to the right
        if is_torch_fx_proxy(input_ids):
            # Item assignment is not supported natively for proxies.
            shifted_input_ids = torch.full(input_ids.shape[:-1] + (1,), decoder_start_token_id)
            shifted_input_ids = torch.cat([shifted_input_ids, input_ids[..., :-1]], dim=-1)
        else:
            shifted_input_ids = input_ids.new_zeros(input_ids.shape)
            shifted_input_ids[..., 1:] = input_ids[..., :-1].clone()
            shifted_input_ids[..., 0] = decoder_start_token_id

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

        return shifted_input_ids


class LongT5Stack(LongT5PreTrainedModel):
    def __init__(self, config, embed_tokens=None):
        super().__init__(config)

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

        self.local_radius = config.local_radius
        self.block_len = self.local_radius + 1

        self.block = nn.ModuleList(
            [LongT5Block(config, has_relative_attention_bias=bool(i == 0)) for i in range(config.num_layers)]
        )
        self.final_layer_norm = LongT5LayerNorm(config.d_model, eps=config.layer_norm_epsilon)
        self.dropout = nn.Dropout(config.dropout_rate)

        self.gradient_checkpointing = False

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

    # Copied from transformers.models.t5.modeling_t5.T5Stack.get_input_embeddings
    def get_input_embeddings(self):
        return self.embed_tokens

    # Copied from transformers.models.t5.modeling_t5.T5Stack.set_input_embeddings
    def set_input_embeddings(self, new_embeddings):
        self.embed_tokens = new_embeddings

    def forward(
        self,
        input_ids=None,
        attention_mask=None,
        encoder_hidden_states=None,
        encoder_attention_mask=None,
        inputs_embeds=None,
        head_mask=None,
        cross_attn_head_mask=None,
        past_key_values=None,
        use_cache=None,
        output_attentions=None,
        output_hidden_states=None,
        return_dict=None,
    ):
        use_cache = use_cache if use_cache is not None else self.config.use_cache
        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

        if input_ids is not None and inputs_embeds is not None:
            err_msg_prefix = "decoder_" if self.is_decoder else ""
            raise ValueError(
                f"You cannot specify both {err_msg_prefix}input_ids and {err_msg_prefix}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:
            err_msg_prefix = "decoder_" if self.is_decoder else ""
            raise ValueError(f"You have to specify either {err_msg_prefix}input_ids or {err_msg_prefix}inputs_embeds")

        if inputs_embeds is None:
            assert self.embed_tokens is not None, "You have to initialize the model with valid token embeddings"
            inputs_embeds = self.embed_tokens(input_ids)

        batch_size, seq_length = input_shape

        # required mask seq length can be calculated via length of past
        mask_seq_length = past_key_values[0][0].shape[2] + seq_length if past_key_values is not None else seq_length

        if use_cache is True:
            assert self.is_decoder, f"`use_cache` can only be set to `True` if {self} is used as a decoder"

        if attention_mask is None:
            attention_mask = torch.ones(batch_size, mask_seq_length, device=inputs_embeds.device)

        # initialize past_key_values with `None` if past does not exist
        if past_key_values is None:
            past_key_values = [None] * len(self.block)

        # We can provide a self-attention mask of dimensions [batch_size, from_seq_length, to_seq_length]
        # ourselves in which case we just need to make it broadcastable to all heads.
        # We use local attention in encoder self-attention, otherwise standard self & cross attentions are used
        if self.is_decoder:
            extended_attention_mask = self.get_extended_attention_mask(
                attention_mask, input_shape, inputs_embeds.device
            )
        elif self.config.encoder_attention_type == "local":
            extended_attention_mask = _get_local_attention_mask(attention_mask, self.block_len, inputs_embeds.device)
        else:  # we need to use both local attention mask and standard extended mask for transient-global attention
            extended_attention_mask = attention_mask

        # If a 2D or 3D attention mask is provided for the cross-attention
        # we need to make broadcastable to [batch_size, num_heads, seq_length, seq_length]
        if self.is_decoder and encoder_hidden_states is not None:
            encoder_batch_size, encoder_sequence_length, _ = encoder_hidden_states.size()
            encoder_hidden_shape = (encoder_batch_size, encoder_sequence_length)
            if encoder_attention_mask is None:
                encoder_attention_mask = torch.ones(encoder_hidden_shape, device=inputs_embeds.device)
            encoder_extended_attention_mask = self.invert_attention_mask(encoder_attention_mask)
        else:
            encoder_extended_attention_mask = None

        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

        # Prepare head mask if needed
        head_mask = self.get_head_mask(head_mask, self.config.num_layers)
        cross_attn_head_mask = self.get_head_mask(cross_attn_head_mask, self.config.num_layers)
        present_key_value_states = () if use_cache else None
        all_hidden_states = () if output_hidden_states else None
        all_attentions = () if output_attentions else None
        all_cross_attentions = () if (output_attentions and self.is_decoder) else None
        position_bias = None
        encoder_decoder_position_bias = None

        hidden_states = self.dropout(inputs_embeds)

        for i, (layer_module, past_key_value) in enumerate(zip(self.block, past_key_values)):
            layer_head_mask = head_mask[i]
            cross_attn_layer_head_mask = cross_attn_head_mask[i]

            if output_hidden_states:
                all_hidden_states = all_hidden_states + (hidden_states,)

            if self.gradient_checkpointing and self.training:
                layer_outputs = self._gradient_checkpointing_func(
                    layer_module.forward,
                    hidden_states,
                    extended_attention_mask,
                    position_bias,
                    encoder_hidden_states,
                    encoder_extended_attention_mask,
                    encoder_decoder_position_bias,
                    layer_head_mask,
                    cross_attn_layer_head_mask,
                    None,  # past_key_value is always None with gradient checkpointing
                    use_cache,
                    output_attentions,
                )
            else:
                layer_outputs = layer_module(
                    hidden_states,
                    attention_mask=extended_attention_mask,
                    position_bias=position_bias,
                    encoder_hidden_states=encoder_hidden_states,
                    encoder_attention_mask=encoder_extended_attention_mask,
                    encoder_decoder_position_bias=encoder_decoder_position_bias,
                    layer_head_mask=layer_head_mask,
                    cross_attn_layer_head_mask=cross_attn_layer_head_mask,
                    past_key_value=past_key_value,
                    use_cache=use_cache,
                    output_attentions=output_attentions,
                )

            # layer_outputs is a tuple with:
            # hidden-states, key-value-states, (self-attention position bias), (self-attention weights), (cross-attention position bias), (cross-attention weights)
            if use_cache is False:
                layer_outputs = layer_outputs[:1] + (None,) + layer_outputs[1:]

            hidden_states, present_key_value_state = layer_outputs[:2]

            # We share the position biases between the layers - the first layer store them
            # layer_outputs = hidden-states, key-value-states (self-attention position bias), (self-attention weights),
            # (cross-attention position bias), (cross-attention weights)
            position_bias = layer_outputs[2]
            if self.is_decoder and encoder_hidden_states is not None:
                encoder_decoder_position_bias = layer_outputs[4 if output_attentions else 3]
            # append next layer key value states
            if use_cache:
                present_key_value_states = present_key_value_states + (present_key_value_state,)

            if output_attentions:
                all_attentions = all_attentions + (layer_outputs[3],)
                if self.is_decoder:
                    all_cross_attentions = all_cross_attentions + (layer_outputs[5],)

        hidden_states = self.final_layer_norm(hidden_states)
        hidden_states = self.dropout(hidden_states)

        # Add last layer
        if output_hidden_states:
            all_hidden_states = all_hidden_states + (hidden_states,)

        if not return_dict:
            return tuple(
                v
                for v in [
                    hidden_states,
                    present_key_value_states,
                    all_hidden_states,
                    all_attentions,
                    all_cross_attentions,
                ]
                if v is not None
            )
        return BaseModelOutputWithPastAndCrossAttentions(
            last_hidden_state=hidden_states,
            past_key_values=present_key_value_states,
            hidden_states=all_hidden_states,
            attentions=all_attentions,
            cross_attentions=all_cross_attentions,
        )


LONGT5_START_DOCSTRING = r"""

    The LongT5 model was proposed in [LongT5: Efficient Text-To-Text Transformer for Long
    Sequences](https://arxiv.org/abs/2112.07916) by Mandy Guo, Joshua Ainslie, David Uthus, Santiago Ontanon, Jianmo
    Ni, Yun-Hsuan Sung and Yinfei Yang. It's an encoder-decoder transformer pre-trained in a text-to-text denoising
    generative setting. LongT5 model is an extension of T5 model, and it enables using one of the two different
    efficient attention mechanisms - (1) Local attention, or (2) Transient-Global attention.

    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, pruning heads
    etc.)

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

    Parameters:
        config ([`LongT5Config`]): 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.
"""

LONGT5_INPUTS_DOCSTRING = r"""
    Args:
        input_ids (`torch.LongTensor` of shape `(batch_size, sequence_length)`):
            Indices of input sequence tokens in the vocabulary. LongT5 is a model with relative position embeddings so
            you should be able to pad the inputs on both the right and the left.

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

            [What are input IDs?](../glossary#input-ids)

            To know more on how to prepare `input_ids` for pretraining take a look a [LONGT5
            Training](./longt5#training).
        attention_mask (`torch.FloatTensor` 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*):
            Indices of decoder input sequence tokens in the vocabulary.

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

            [What are decoder input IDs?](../glossary#decoder-input-ids)

            LONGT5 uses the `pad_token_id` as the starting token for `decoder_input_ids` generation. If
            `past_key_values` is used, optionally only the last `decoder_input_ids` have to be input (see
            `past_key_values`).

            To know more on how to prepare `decoder_input_ids` for pretraining take a look at [LONGT5
            Training](./longt5#training).
        decoder_attention_mask (`torch.BoolTensor` 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.
        head_mask (`torch.FloatTensor` of shape `(num_heads,)` or `(num_layers, num_heads)`, *optional*):
            Mask to nullify selected heads of the self-attention modules in the encoder. Mask values selected in `[0,
            1]`:

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

        decoder_head_mask (`torch.FloatTensor` of shape `(num_heads,)` or `(num_layers, num_heads)`, *optional*):
            Mask to nullify selected heads of the self-attention modules in the decoder. Mask values selected in `[0,
            1]`:

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

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

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

        encoder_outputs (`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)` 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))` of length `config.n_layers` with each tuple having 4 tensors of shape `(batch_size, num_heads, sequence_length - 1, embed_size_per_head)`):
            Contains precomputed key and value hidden states of the attention blocks. Can be used to speed up decoding.

            If `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.
"""

LONGT5_ENCODER_INPUTS_DOCSTRING = r"""
    Args:
        input_ids (`torch.LongTensor` of shape `(batch_size, sequence_length)`):
            Indices of input sequence tokens in the vocabulary. LongT5 is a model with relative position embeddings so
            you should be able to pad the inputs on both the right and the left.

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

            To know more on how to prepare `input_ids` for pretraining take a look a [LONGT5
            Training](./longt5#training).
        attention_mask (`torch.FloatTensor` 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)
        head_mask (`torch.FloatTensor` of shape `(num_heads,)` or `(num_layers, num_heads)`, *optional*):
            Mask to nullify selected heads of the self-attention modules. Mask values selected in `[0, 1]`:

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

        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.
"""

# Warning message for FutureWarning: head_mask was separated into two input args - head_mask, decoder_head_mask
__HEAD_MASK_WARNING_MSG = """
The input argument `head_mask` was split into two arguments `head_mask` and `decoder_head_mask`. Currently,
`decoder_head_mask` is set to copy `head_mask`, but this feature is deprecated and will be removed in future versions.
If you do not want to use any `decoder_head_mask` now, please set `decoder_head_mask = torch.ones(num_layers,
num_heads)`.
"""


@add_start_docstrings(
    "The bare LONGT5 Model transformer outputting raw hidden-states without any specific head on top.",
    LONGT5_START_DOCSTRING,
)
class LongT5Model(LongT5PreTrainedModel):
    _keys_to_ignore_on_load_unexpected = [
        r"decoder.block.0.layer.1.EncDecAttention.relative_attention_bias.weight",
    ]
    _tied_weights_keys = ["encoder.embed_tokens.weight", "decoder.embed_tokens.weight"]

    def __init__(self, config: LongT5Config):
        super().__init__(config)
        self.shared = nn.Embedding(config.vocab_size, config.d_model)

        encoder_config = copy.deepcopy(config)
        encoder_config.is_decoder = False
        encoder_config.use_cache = False
        encoder_config.is_encoder_decoder = False
        self.encoder = LongT5Stack(encoder_config, self.shared)

        decoder_config = copy.deepcopy(config)
        decoder_config.is_decoder = True
        decoder_config.is_encoder_decoder = False
        decoder_config.num_layers = config.num_decoder_layers
        self.decoder = LongT5Stack(decoder_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, new_embeddings):
        self.shared = new_embeddings
        self.encoder.set_input_embeddings(new_embeddings)
        self.decoder.set_input_embeddings(new_embeddings)

    def _tie_weights(self):
        if self.config.tie_word_embeddings:
            self._tie_or_clone_weights(self.encoder.embed_tokens, self.shared)
            self._tie_or_clone_weights(self.decoder.embed_tokens, self.shared)

    def get_encoder(self):
        return self.encoder

    def get_decoder(self):
        return self.decoder

    def _prune_heads(self, heads_to_prune):
        """
        Prunes heads of the model. heads_to_prune: dict of {layer_num: list of heads to prune in this layer} See base
        class PreTrainedModel
        """
        for layer, heads in heads_to_prune.items():
            self.encoder.layer[layer].attention.prune_heads(heads)

    @add_start_docstrings_to_model_forward(LONGT5_INPUTS_DOCSTRING)
    @replace_return_docstrings(output_type=Seq2SeqModelOutput, config_class=_CONFIG_FOR_DOC)
    def forward(
        self,
        input_ids: Optional[torch.LongTensor] = None,
        attention_mask: Optional[torch.FloatTensor] = None,
        decoder_input_ids: Optional[torch.LongTensor] = None,
        decoder_attention_mask: Optional[torch.BoolTensor] = None,
        head_mask: Optional[torch.FloatTensor] = None,
        decoder_head_mask: Optional[torch.FloatTensor] = None,
        cross_attn_head_mask: Optional[torch.Tensor] = None,
        encoder_outputs: Optional[Tuple[Tuple[torch.FloatTensor]]] = None,
        past_key_values: Optional[Tuple[Tuple[torch.FloatTensor]]] = None,
        inputs_embeds: Optional[torch.Tensor] = None,
        decoder_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,
    ) -> Union[Tuple[torch.FloatTensor], Seq2SeqModelOutput]:
        r"""
        Returns:

        Example:

        ```python
        >>> from transformers import AutoTokenizer, LongT5Model

        >>> tokenizer = AutoTokenizer.from_pretrained("google/long-t5-local-base")
        >>> model = LongT5Model.from_pretrained("google/long-t5-local-base")

        >>> # Let's try a very long encoder input.
        >>> input_ids = tokenizer(
        ...     100 * "Studies have been shown that owning a dog is good for you", return_tensors="pt"
        ... ).input_ids  # Batch size 1

        >>> decoder_input_ids = tokenizer("Studies show that", return_tensors="pt").input_ids  # Batch size 1

        >>> # forward pass
        >>> outputs = model(input_ids=input_ids, decoder_input_ids=decoder_input_ids)
        >>> last_hidden_states = outputs.last_hidden_state
        ```"""
        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

        # FutureWarning: head_mask was separated into two input args - head_mask, decoder_head_mask
        if head_mask is not None and decoder_head_mask is None:
            if self.config.num_layers == self.config.num_decoder_layers:
                warnings.warn(__HEAD_MASK_WARNING_MSG, FutureWarning)
                decoder_head_mask = head_mask

        # Encode if needed (training, first prediction pass)
        if encoder_outputs is None:
            encoder_outputs = self.encoder(
                input_ids=input_ids,
                attention_mask=attention_mask,
                inputs_embeds=inputs_embeds,
                head_mask=head_mask,
                output_attentions=output_attentions,
                output_hidden_states=output_hidden_states,
                return_dict=return_dict,
            )
        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,
            )

        hidden_states = encoder_outputs[0]

        # Decode
        decoder_outputs = self.decoder(
            input_ids=decoder_input_ids,
            attention_mask=decoder_attention_mask,
            inputs_embeds=decoder_inputs_embeds,
            past_key_values=past_key_values,
            encoder_hidden_states=hidden_states,
            encoder_attention_mask=attention_mask,
            head_mask=decoder_head_mask,
            cross_attn_head_mask=cross_attn_head_mask,
            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("""LONGT5 Model with a `language modeling` head on top.""", LONGT5_START_DOCSTRING)
class LongT5ForConditionalGeneration(LongT5PreTrainedModel):
    _keys_to_ignore_on_load_unexpected = [
        r"decoder.block.0.layer.1.EncDecAttention.relative_attention_bias.weight",
    ]
    _tied_weights_keys = ["encoder.embed_tokens.weight", "decoder.embed_tokens.weight", "lm_head.weight"]

    def __init__(self, config: LongT5Config):
        super().__init__(config)
        self.model_dim = config.d_model

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

        encoder_config = copy.deepcopy(config)
        encoder_config.is_decoder = False
        encoder_config.use_cache = False
        encoder_config.is_encoder_decoder = False
        self.encoder = LongT5Stack(encoder_config, self.shared)

        decoder_config = copy.deepcopy(config)
        decoder_config.is_decoder = True
        decoder_config.is_encoder_decoder = False
        decoder_config.num_layers = config.num_decoder_layers
        self.decoder = LongT5Stack(decoder_config, self.shared)

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

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

    def get_input_embeddings(self):
        return self.shared

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

    def _tie_weights(self):
        if self.config.tie_word_embeddings:
            self._tie_or_clone_weights(self.encoder.embed_tokens, self.shared)
            self._tie_or_clone_weights(self.decoder.embed_tokens, self.shared)

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

    def get_output_embeddings(self):
        return self.lm_head

    def get_encoder(self):
        return self.encoder

    def get_decoder(self):
        return self.decoder

    @add_start_docstrings_to_model_forward(LONGT5_INPUTS_DOCSTRING)
    @replace_return_docstrings(output_type=Seq2SeqLMOutput, config_class=_CONFIG_FOR_DOC)
    def forward(
        self,
        input_ids: Optional[torch.LongTensor] = None,
        attention_mask: Optional[torch.FloatTensor] = None,
        decoder_input_ids: Optional[torch.LongTensor] = None,
        decoder_attention_mask: Optional[torch.BoolTensor] = None,
        head_mask: Optional[torch.FloatTensor] = None,
        decoder_head_mask: Optional[torch.FloatTensor] = None,
        cross_attn_head_mask: Optional[torch.Tensor] = None,
        encoder_outputs: Optional[Tuple[Tuple[torch.Tensor]]] = None,
        past_key_values: Optional[Tuple[Tuple[torch.Tensor]]] = 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[torch.FloatTensor], Seq2SeqLMOutput]:
        r"""
        labels (`torch.LongTensor` of shape `(batch_size,)`, *optional*):
            Labels for computing the sequence classification/regression loss. Indices should be in `[-100, 0, ...,
            config.vocab_size - 1]`. All labels set to `-100` are ignored (masked), the loss is only computed for
            labels in `[0, ..., config.vocab_size]`

        Returns:

        Examples:

        ```python
        >>> from transformers import AutoTokenizer, LongT5ForConditionalGeneration

        >>> tokenizer = AutoTokenizer.from_pretrained("Stancld/longt5-tglobal-large-16384-pubmed-3k_steps")
        >>> model = LongT5ForConditionalGeneration.from_pretrained(
        ...     "Stancld/longt5-tglobal-large-16384-pubmed-3k_steps"
        ... )

        >>> # Let's try a very long input.
        >>> inputs = tokenizer(100 * "studies have shown that owning a dog is good for you ", return_tensors="pt")
        >>> input_ids = inputs.input_ids

        >>> outputs = model.generate(input_ids)
        >>> print(tokenizer.decode(outputs[0], skip_special_tokens=True))
        abstractthe aim of this article is to provide an overview of the literature on the role of dog
        ```"""
        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

        # FutureWarning: head_mask was separated into two input args - head_mask, decoder_head_mask
        if head_mask is not None and decoder_head_mask is None:
            if self.config.num_layers == self.config.num_decoder_layers:
                warnings.warn(__HEAD_MASK_WARNING_MSG, FutureWarning)
                decoder_head_mask = head_mask

        # Encode if needed (training, first prediction pass)
        if encoder_outputs is None:
            # Convert encoder inputs in embeddings if needed
            encoder_outputs = self.encoder(
                input_ids=input_ids,
                attention_mask=attention_mask,
                inputs_embeds=inputs_embeds,
                head_mask=head_mask,
                output_attentions=output_attentions,
                output_hidden_states=output_hidden_states,
                return_dict=return_dict,
            )
        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,
            )

        hidden_states = encoder_outputs[0]

        if labels is not None and decoder_input_ids is None and decoder_inputs_embeds is None:
            # get decoder inputs from shifting lm labels to the right
            decoder_input_ids = self._shift_right(labels)

        # Decode
        decoder_outputs = self.decoder(
            input_ids=decoder_input_ids,
            attention_mask=decoder_attention_mask,
            inputs_embeds=decoder_inputs_embeds,
            past_key_values=past_key_values,
            encoder_hidden_states=hidden_states,
            encoder_attention_mask=attention_mask,
            head_mask=decoder_head_mask,
            cross_attn_head_mask=cross_attn_head_mask,
            use_cache=use_cache,
            output_attentions=output_attentions,
            output_hidden_states=output_hidden_states,
            return_dict=return_dict,
        )

        sequence_output = decoder_outputs[0]

        if self.config.tie_word_embeddings:
            # Rescale output before projecting on vocab
            # See https://github.com/tensorflow/mesh/blob/fa19d69eafc9a482aff0b59ddd96b025c0cb207d/mesh_tensorflow/transformer/transformer.py#L586
            sequence_output = sequence_output * (self.model_dim**-0.5)

        lm_logits = self.lm_head(sequence_output)

        loss = None
        if labels is not None:
            loss_fct = CrossEntropyLoss(ignore_index=-100)

            labels = labels.to(lm_logits.device)
            loss = loss_fct(lm_logits.view(-1, lm_logits.size(-1)), labels.view(-1))
            # TODO(thom): Add z_loss https://github.com/tensorflow/mesh/blob/fa19d69eafc9a482aff0b59ddd96b025c0cb207d/mesh_tensorflow/layers.py#L666

        if not return_dict:
            output = (lm_logits,) + decoder_outputs[1:] + encoder_outputs
            return ((loss,) + output) if loss is not None else output

        return Seq2SeqLMOutput(
            loss=loss,
            logits=lm_logits,
            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,
        )

    def prepare_inputs_for_generation(
        self,
        input_ids,
        past_key_values=None,
        attention_mask=None,
        head_mask=None,
        decoder_head_mask=None,
        cross_attn_head_mask=None,
        use_cache=None,
        encoder_outputs=None,
        **kwargs,
    ):
        # cut decoder_input_ids if past_key_values is used
        if past_key_values is not None:
            past_length = past_key_values[0][0].shape[2]

            # Some generation methods already pass only the last input ID
            if input_ids.shape[1] > past_length:
                remove_prefix_length = past_length
            else:
                # Default to old behavior: keep only final ID
                remove_prefix_length = input_ids.shape[1] - 1

            input_ids = input_ids[:, remove_prefix_length:]

        return {
            "decoder_input_ids": input_ids,
            "past_key_values": past_key_values,
            "encoder_outputs": encoder_outputs,
            "attention_mask": attention_mask,
            "head_mask": head_mask,
            "decoder_head_mask": decoder_head_mask,
            "cross_attn_head_mask": cross_attn_head_mask,
            "use_cache": use_cache,
        }

    def prepare_decoder_input_ids_from_labels(self, labels: torch.Tensor):
        return self._shift_right(labels)

    def _reorder_cache(self, past_key_values, beam_idx):
        # if decoder past is not included in output
        # speedy decoding is disabled and no need to reorder
        if past_key_values is None:
            logger.warning("You might want to consider setting `use_cache=True` to speed up decoding")
            return past_key_values

        reordered_decoder_past = ()
        for layer_past_states in past_key_values:
            # get the correct batch idx from layer past batch dim
            # batch dim of `past` is at 2nd position
            reordered_layer_past_states = ()
            for layer_past_state in layer_past_states:
                # need to set correct `past` for each of the four key / value states
                reordered_layer_past_states = reordered_layer_past_states + (
                    layer_past_state.index_select(0, beam_idx.to(layer_past_state.device)),
                )

            assert reordered_layer_past_states[0].shape == layer_past_states[0].shape
            assert len(reordered_layer_past_states) == len(layer_past_states)

            reordered_decoder_past = reordered_decoder_past + (reordered_layer_past_states,)
        return reordered_decoder_past


@add_start_docstrings(
    "The bare LONGT5 Model transformer outputting encoder's raw hidden-states without any specific head on top.",
    LONGT5_START_DOCSTRING,
)
class LongT5EncoderModel(LongT5PreTrainedModel):
    _tied_weights_keys = ["encoder.embed_tokens.weight"]
    _keys_to_ignore_on_load_unexpected = [r"decoder"]

    def __init__(self, config: LongT5Config):
        super().__init__(config)
        self.shared = nn.Embedding(config.vocab_size, config.d_model)

        encoder_config = copy.deepcopy(config)
        encoder_config.use_cache = False
        encoder_config.is_encoder_decoder = False
        self.encoder = LongT5Stack(encoder_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, new_embeddings):
        self.shared = new_embeddings
        self.encoder.set_input_embeddings(new_embeddings)

    def _tie_weights(self):
        if self.config.tie_word_embeddings:
            self._tie_or_clone_weights(self.encoder.embed_tokens, self.shared)

    def get_encoder(self):
        return self.encoder

    def _prune_heads(self, heads_to_prune):
        """
        Prunes heads of the model. heads_to_prune: dict of {layer_num: list of heads to prune in this layer} See base
        class PreTrainedModel
        """
        for layer, heads in heads_to_prune.items():
            self.encoder.layer[layer].attention.prune_heads(heads)

    @add_start_docstrings_to_model_forward(LONGT5_ENCODER_INPUTS_DOCSTRING)
    @replace_return_docstrings(output_type=BaseModelOutput, config_class=_CONFIG_FOR_DOC)
    def forward(
        self,
        input_ids: Optional[torch.LongTensor] = None,
        attention_mask: Optional[torch.FloatTensor] = None,
        head_mask: Optional[torch.FloatTensor] = None,
        inputs_embeds: Optional[torch.FloatTensor] = None,
        output_attentions: Optional[bool] = None,
        output_hidden_states: Optional[bool] = None,
        return_dict: Optional[bool] = None,
    ) -> Union[Tuple[torch.FloatTensor], BaseModelOutput]:
        r"""
        Returns:

        Example:

        ```python
        >>> from transformers import AutoTokenizer, LongT5ForConditionalGeneration

        >>> tokenizer = AutoTokenizer.from_pretrained("google/long-t5-local-base")
        >>> model = LongT5EncoderModel.from_pretrained("google/long-t5-local-base")
        >>> input_ids = tokenizer(
        ...     100 * "Studies have been shown that owning a dog is good for you ", return_tensors="pt"
        ... ).input_ids  # Batch size 1
        >>> outputs = model(input_ids=input_ids)
        >>> last_hidden_states = outputs.last_hidden_state
        ```"""
        return_dict = return_dict if return_dict is not None else self.config.use_return_dict

        encoder_outputs = self.encoder(
            input_ids=input_ids,
            attention_mask=attention_mask,
            inputs_embeds=inputs_embeds,
            head_mask=head_mask,
            output_attentions=output_attentions,
            output_hidden_states=output_hidden_states,
            return_dict=return_dict,
        )

        return encoder_outputs