modeling_custom_t5.py

import logging
import math
import os
from dataclasses import dataclass
from pathlib import Path
from typing import Optional, Tuple

import torch
from torch import nn
from torch.nn import CrossEntropyLoss
from torch.utils.checkpoint import checkpoint
from transformers import T5Config
from transformers.modeling_outputs import (
    BaseModelOutputWithPastAndCrossAttentions,
)
from transformers.utils import ModelOutput
from transformers.utils.model_parallel_utils import get_device_map, assert_device_map

from .configuration_custom_t5 import (
    POSITION_ENCODING_REL_T5_BIAS,
    POSITION_ENCODING_REL_TRANSFORMER_XL,
    POSITION_ENCODING_ROTARY,
    POSITION_ENCODING_ROTARY_NEW,
    POSITION_ENCODING_ABS_LEARNED,
    POSITION_ENCODING_ABS_SINUSOID,
    POSITION_ENCODING_ALiBi,
    POSITION_ENCODING_ALiBi_LEARNED,
    POSITION_ENCODING_NONE,
    POSITION_ENCODING_NONE_WINDOW,
    CustomT5Config,
)
from .modeling_t5 import (
    T5Stack,
    T5PreTrainedModel,
    T5Block,
    T5LayerNorm,
    T5LayerFF,
    T5LayerSelfAttention,
    T5Attention,
    T5LayerCrossAttention,
)

logger = logging.getLogger(__name__)


@dataclass
class CausalLMOutputWithPastAndLoss(ModelOutput):
    """
    Base class for causal language model (or autoregressive) outputs.

    Args:
        loss (`torch.FloatTensor` of shape `(1,)`, *optional*, returned when `labels` is provided):
            Language modeling loss (for next-token prediction).
        logits (`torch.FloatTensor` of shape `(batch_size, sequence_length, config.vocab_size)`):
            Prediction scores of the language modeling head (scores for each vocabulary token before SoftMax).
        past_key_values (`tuple(tuple(torch.FloatTensor))`, *optional*, returned when `use_cache=True` is passed or when `config.use_cache=True`):
            Tuple of `tuple(torch.FloatTensor)` of length `config.n_layers`, with each tuple having 2 tensors of shape
            `(batch_size, num_heads, sequence_length, embed_size_per_head)`)

            Contains pre-computed hidden-states (key and values in the self-attention blocks) that can be used (see
            `past_key_values` input) to speed up sequential decoding.
        hidden_states (`tuple(torch.FloatTensor)`, *optional*, returned when `output_hidden_states=True` is passed or when `config.output_hidden_states=True`):
            Tuple of `torch.FloatTensor` (one for the output of the embeddings, if the model has an embedding layer, +
            one for the output of each layer) of shape `(batch_size, sequence_length, hidden_size)`.

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

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

    loss: Optional[torch.FloatTensor] = None
    logits: torch.FloatTensor = None
    past_key_values: Optional[Tuple[Tuple[torch.FloatTensor]]] = None
    hidden_states: Optional[Tuple[torch.FloatTensor]] = None
    attentions: Optional[Tuple[torch.FloatTensor]] = None
    non_reduced_loss: Optional[torch.FloatTensor] = None


def fixed_pos_embedding(x, seq_dim=1, seq_len=None):
    dim = x.shape[-1]
    if seq_len is None:
        seq_len = x.shape[seq_dim]
    inv_freq = 1.0 / (10000 ** (torch.arange(0, dim, 2) / dim))
    sinusoid_inp = (
        torch.einsum("i , j -> i j", torch.arange(seq_len), inv_freq)
        .to(x.device)
        .float()
    )
    return torch.sin(sinusoid_inp), torch.cos(sinusoid_inp)


def rotate_every_two(x):
    """
    Example: [a, b, c, d] -> [-b, a, -d, c]
    """
    x1 = x[:, :, :, ::2]
    x2 = x[:, :, :, 1::2]
    x = torch.stack((-x2, x1), axis=-1)
    return x.flatten(-2)  # in einsum notation: rearrange(x, '... d j -> ... (d j)')


def apply_rotary_pos_emb(x, sincos, offset=0):
    sin, cos = map(
        lambda t: t[None, offset : x.shape[1] + offset, None, :].repeat_interleave(
            2, 3
        ),
        sincos,
    )
    # einsum notation for lambda t: repeat(t[offset:x.shape[1]+offset,:], "n d -> () n () (d j)", j=2)
    return (x * cos) + (rotate_every_two(x) * sin)


def apply_rotary_pos_emb_new(x, sincos, offset=0):
    sin, cos = map(
        lambda t: t[:, :, None, :].repeat_interleave(2, 3),
        sincos,
    )
    # einsum notation for lambda t: repeat(t[offset:x.shape[1]+offset,:], "n d -> () n () (d j)", j=2)
    return (x * cos) + (rotate_every_two(x) * sin)


class PositionalEmbedding(nn.Module):
    def __init__(self, demb):
        super().__init__()

        self.demb = demb

        inv_freq = 1 / (10000 ** (torch.arange(0.0, demb, 2.0) / demb))
        self.register_buffer("inv_freq", inv_freq)

    def forward(self, pos_seq, bsz=None):
        sinusoid_inp = torch.ger(pos_seq, self.inv_freq)
        pos_emb = torch.cat([sinusoid_inp.sin(), sinusoid_inp.cos()], dim=-1)

        if bsz is not None:
            return pos_emb[None, :, :].expand(bsz, -1, -1)
        else:
            return pos_emb[None, :, :]


class FixedAbsolutePositionalEmbedding(nn.Module):
    def __init__(self, dim):
        super().__init__()
        inv_freq = 1.0 / (10000 ** (torch.arange(0, dim, 2).float() / dim))
        t = torch.arange(16384).type_as(inv_freq)
        sinusoid_inp = torch.einsum("i , j -> i j", t, inv_freq)
        emb = torch.cat((sinusoid_inp.sin(), sinusoid_inp.cos()), dim=-1)
        self.embed = nn.Embedding.from_pretrained(emb, freeze=True)

    def forward(self, position_ids: torch.Tensor):
        return self.embed(position_ids.long())


class FixedRotaryPositionalEmbedding(nn.Module):
    def __init__(
        self, rotary_dim: int, rotary_base: int = 10000, max_position: int = 16384
    ):
        super().__init__()
        # This is an inverse frequency tensor
        # Each dimension has a higher denominator than the previous one
        # So, the frequency will be lower for higher dimensions
        inv_freq = 1.0 / (
            rotary_base ** (torch.arange(0, rotary_dim, 2).float() / rotary_dim)
        )  # [rotary_dim/2]

        # Now, we create frequencies for each position
        t = torch.arange(max_position, device=inv_freq.device, dtype=inv_freq.dtype)
        freqs = torch.einsum("i,j->ij", t, inv_freq)  # [max_position, rotary_dim/2]

        sins = torch.sin(freqs)
        coss = torch.cos(freqs)

        emb = torch.cat([sins, coss], dim=-1)  # [max_position, rotary_dim]
        self.embed = nn.Embedding.from_pretrained(emb, freeze=True)

    def forward(self, position_ids: torch.Tensor):
        return self.embed(position_ids.long())


class CustomT5Attention(T5Attention):
    def __init__(self, config: T5Config, has_relative_attention_bias=False):
        super(T5Attention, self).__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.d_model = config.d_model
        self.key_value_proj_dim = config.d_kv
        self.d_head = 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)

        self.position_encoding_type = getattr(
            config, "position_encoding_type", POSITION_ENCODING_REL_T5_BIAS
        )

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

        if self.position_encoding_type == POSITION_ENCODING_REL_TRANSFORMER_XL:
            self.r_r_bias = nn.Parameter(torch.FloatTensor(self.n_heads, self.d_head))
            nn.init.normal_(
                self.r_r_bias, mean=0.0, std=config.initializer_factor * 0.2
            )
            self.r_w_bias = nn.Parameter(torch.FloatTensor(self.n_heads, self.d_head))
            nn.init.normal_(
                self.r_w_bias, mean=0.0, std=config.initializer_factor * 0.2
            )
            self.r = nn.Linear(self.d_model, self.n_heads * self.d_head, bias=False)
            self.r.weight.data.normal_(
                mean=0.0, std=config.initializer_factor * (self.d_model**-0.5)
            )
            self.pos_emb = PositionalEmbedding(self.d_model)
            self.clamp_length = 1000

        if self.position_encoding_type == POSITION_ENCODING_ROTARY:
            self.rotary_dim = None
            if getattr(config, "rotary_dim", None) is not None:
                self.rotary_dim = config.rotary_dim
            self.rotary_dim = int(0.25 * self.d_head)

        if self.position_encoding_type == POSITION_ENCODING_ROTARY_NEW:
            # We hardcode the rotary dim to 25 percent of the head dim
            self.rotary_dim = self.d_head // 4

        self.pruned_heads = set()
        self.gradient_checkpointing = False

    def _rel_shift(self, x):
        zero_pad_shape = x.size()[:2] + (x.size(2), 1)
        zero_pad = torch.zeros(zero_pad_shape, device=x.device, dtype=x.dtype)
        x_padded = torch.cat([zero_pad, x], dim=3)
        x_padded_shape = x.size()[:2] + (x.size(3) + 1, x.size(2))
        x_padded = x_padded.view(*x_padded_shape)
        x = x_padded[:, :, 1:, :].view_as(x)
        return x

    def forward(
        self,
        hidden_states,
        mask=None,
        position_bias=None,
        key_value_states=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:
            assert (
                len(past_key_value) == 2
            ), 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)
                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)

        if self.position_encoding_type in [
            POSITION_ENCODING_ROTARY,
            POSITION_ENCODING_ROTARY_NEW,
        ]:
            key_states = shape(self.k(hidden_states))
        else:
            # 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,
        )

        attention_output_dict = {}

        if self.position_encoding_type == POSITION_ENCODING_REL_T5_BIAS:
            scores = torch.matmul(query_states, key_states.transpose(3, 2))
            attention_output_dict["scores_before"] = scores
            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)

                # 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)

            scores += position_bias
        elif self.position_encoding_type == POSITION_ENCODING_REL_TRANSFORMER_XL:
            if position_bias is None:
                pos_seq = torch.arange(
                    real_seq_length - 1,
                    -1,
                    -1.0,
                    device=hidden_states.device,
                    dtype=hidden_states.dtype,
                )
                if self.clamp_length > 0:
                    pos_seq = pos_seq.clamp_(max=self.clamp_length)
                position_bias = self.pos_emb(pos_seq)
                position_bias = nn.functional.dropout(
                    position_bias, p=self.dropout, training=self.training
                )

            position_embeds = position_bias  # position embeds: [1, seq_len, d_model]

            r_head_k = self.r(position_embeds)  # [1, seq_len, n_head*d_head]
            r_head_k = r_head_k.view(
                position_embeds.shape[1], self.n_heads, self.d_head
            )  # [seq_len, n_head, d_head]

            rw_head_q = query_states + self.r_w_bias[None, :, None, :]
            AC = torch.einsum("bnqd,bnkd->bnqk", (rw_head_q, key_states))

            rr_head_q = query_states + self.r_r_bias[None, :, None, :]
            BD = torch.einsum("bnid,jnd->bnij", (rr_head_q, r_head_k))
            BD = self._rel_shift(BD)

            scores = AC + BD

            if mask is not None:
                scores += mask
        elif self.position_encoding_type == POSITION_ENCODING_ROTARY:
            r_seq_len = hidden_states.shape[1]
            r_offset = 0

            if past_key_value is not None:
                r_offset = past_key_value[0].shape[2]
                r_seq_len += r_offset

            query_states = query_states.permute(0, 2, 1, 3)
            key_states = key_states.permute(0, 2, 1, 3)

            if self.rotary_dim is not None:
                k_rot = key_states[:, :, :, : self.rotary_dim]
                k_pass = key_states[:, :, :, self.rotary_dim :]

                q_rot = query_states[:, :, :, : self.rotary_dim]
                q_pass = query_states[:, :, :, self.rotary_dim :]

                sincos = fixed_pos_embedding(k_rot, 1, seq_len=r_seq_len)
                k_rot = apply_rotary_pos_emb(k_rot, sincos, offset=r_offset)
                q_rot = apply_rotary_pos_emb(q_rot, sincos, offset=r_offset)

                if output_attentions:
                    scores_pass = torch.matmul(
                        q_pass.permute(0, 2, 1, 3),
                        k_pass.permute(0, 2, 1, 3).transpose(3, 2),
                    )
                    attention_output_dict["scores_pass"] = scores_pass

                    scores_rot = torch.matmul(
                        q_rot.permute(0, 2, 1, 3),
                        k_rot.permute(0, 2, 1, 3).transpose(3, 2),
                    )
                    attention_output_dict["scores_rot"] = scores_rot

                key_states = torch.cat([k_rot, k_pass], dim=-1)
                query_states = torch.cat([q_rot, q_pass], dim=-1)
            else:
                sincos = fixed_pos_embedding(key_states, 1, seq_len=r_seq_len)
                key_states = apply_rotary_pos_emb(key_states, sincos, offset=r_offset)
                query_states = apply_rotary_pos_emb(
                    query_states, sincos, offset=r_offset
                )

            query_states = query_states.permute(0, 2, 1, 3)
            key_states = key_states.permute(0, 2, 1, 3)

            if past_key_value is not None:
                key_states = torch.cat([past_key_value[0], key_states], dim=2)

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

        elif self.position_encoding_type == POSITION_ENCODING_ROTARY_NEW:
            r_seq_len = hidden_states.shape[1]
            r_offset = 0

            if past_key_value is not None:
                r_offset = past_key_value[0].shape[2]
                r_seq_len += r_offset

            query_states = query_states.permute(0, 2, 1, 3)
            key_states = key_states.permute(0, 2, 1, 3)

            if self.rotary_dim is not None:
                k_rot = key_states[:, :, :, : self.rotary_dim]
                k_pass = key_states[:, :, :, self.rotary_dim :]

                q_rot = query_states[:, :, :, : self.rotary_dim]
                q_pass = query_states[:, :, :, self.rotary_dim :]

                sincos = position_bias
                # sincos is just vector created by torch.cat([sin, cos], dim=-1)
                # so we can just split it in half
                sin = sincos[:, :, : self.rotary_dim // 2]
                cos = sincos[:, :, self.rotary_dim // 2 :]

                # We don't need to pass offset here, because we already used
                # position_ids to retrieve correct sin and cos vectors
                k_rot = apply_rotary_pos_emb_new(k_rot, (sin, cos))
                q_rot = apply_rotary_pos_emb_new(q_rot, (sin, cos))

                key_states = torch.cat([k_rot, k_pass], dim=-1)
                query_states = torch.cat([q_rot, q_pass], dim=-1)
            else:
                raise ValueError("rotary_dim is None")

            query_states = query_states.permute(0, 2, 1, 3)
            key_states = key_states.permute(0, 2, 1, 3)

            if past_key_value is not None:
                key_states = torch.cat([past_key_value[0], key_states], dim=2)

            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 mask is not None:
                scores += mask  # (batch_size, n_heads, seq_length, key_length)
        elif self.position_encoding_type == POSITION_ENCODING_ALiBi:
            scores = torch.matmul(query_states, key_states.transpose(3, 2))
            attention_output_dict["scores_before"] = scores

            alibi = position_bias
            alibi = alibi.view(batch_size, self.n_heads, 1, key_length)

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

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

            scores += alibi
        else:
            assert (
                self.position_encoding_type == POSITION_ENCODING_NONE
            ), f"Unknown position encoding type: {self.position_encoding_type}"
            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 mask is not None:
                scores += mask  # (batch_size, n_heads, seq_length, key_length)

        attention_output_dict["scores"] = scores

        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

        attention_output_dict["probs"] = attn_weights

        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 + (attention_output_dict,)
        return outputs


class CustomT5LayerSelfAttention(T5LayerSelfAttention):
    def __init__(self, config, has_relative_attention_bias=False):
        super(T5LayerSelfAttention, self).__init__()
        self.SelfAttention = CustomT5Attention(
            config, has_relative_attention_bias=has_relative_attention_bias
        )
        self.layer_norm = T5LayerNorm(config.d_model, eps=config.layer_norm_epsilon)
        self.dropout = nn.Dropout(config.dropout_rate)


class CustomT5Block(T5Block):
    def __init__(self, config, has_relative_attention_bias=False):
        super(T5Block, self).__init__()
        self.is_decoder = config.is_decoder
        assert self.is_decoder
        self.layer = nn.ModuleList()
        self.layer.append(
            CustomT5LayerSelfAttention(
                config, has_relative_attention_bias=has_relative_attention_bias
            )
        )
        if self.is_decoder:
            self.layer.append(T5LayerCrossAttention(config))

        self.layer.append(T5LayerFF(config))


def _make_causal_mask(
    input_ids_shape: torch.Size, device: torch.device, past_key_values_length: int
) -> torch.BoolTensor:
    """
    Make causal mask used for self-attention.
    """
    batch_size, target_length = input_ids_shape
    mask = torch.empty(
        (target_length, target_length + past_key_values_length),
        dtype=torch.bool,
        device=device,
    )
    # ONNX doesn't support `torch.Tensor.triu` properly, thus we use this workaround
    seq_ids = torch.arange(target_length, device=device)
    mask[:, past_key_values_length:] = seq_ids[:, None] < seq_ids[None, :]

    if past_key_values_length > 0:
        mask[:, :past_key_values_length] = False

    expanded_mask = mask[None, None, :, :].expand(
        batch_size, 1, target_length, target_length + past_key_values_length
    )
    return expanded_mask


def _expand_mask(mask: torch.Tensor, tgt_length: int) -> torch.BoolTensor:
    """
    Expands attention_mask from `[batch_size, src_length]` to `[batch_size, 1, tgt_length, src_length]`.
    """
    batch_size, src_length = mask.shape
    tgt_length = tgt_length if tgt_length is not None else src_length

    expanded_mask = ~(mask[:, None, None, :].to(torch.bool))
    return expanded_mask.expand(batch_size, 1, tgt_length, src_length)


def build_alibi_tensor(
    attention_mask: torch.Tensor, num_heads: int, dtype: torch.dtype
) -> torch.Tensor:
    """
    Link to paper: https://arxiv.org/abs/2108.12409 Alibi tensor is not causal as the original paper mentions, it
    relies on a translation invariance of softmax for quick implementation: with l being a tensor, and a fixed value
    `softmax(l+a) = softmax(l)`. Based on
    https://github.com/ofirpress/attention_with_linear_biases/blob/a35aaca144e0eb6b789dfcb46784c4b8e31b7983/fairseq/models/transformer.py#L742
    TODO @thomasw21 this doesn't work as nicely due to the masking strategy, and so masking varies slightly.
    Args:
    Returns tensor shaped (batch_size * num_heads, 1, max_seq_len)
        attention_mask (`torch.Tensor`):
            Token-wise attention mask, this should be of shape (batch_size, max_seq_len).
        num_heads (`int`, *required*):
            number of heads
        dtype (`torch.dtype`, *optional*, default=`torch.bfloat16`):
            dtype of the output tensor
    """
    if len(attention_mask.shape) == 2:
        batch_size, seq_length = attention_mask.shape
    elif len(attention_mask.shape) == 3:
        batch_size, _, seq_length = attention_mask.shape
    closest_power_of_2 = 2 ** math.floor(math.log2(num_heads))
    base = torch.tensor(
        2 ** (-(2 ** -(math.log2(closest_power_of_2) - 3))),
        device=attention_mask.device,
        dtype=torch.float32,
    )
    powers = torch.arange(
        1, 1 + closest_power_of_2, device=attention_mask.device, dtype=torch.int32
    )
    slopes = torch.pow(base, powers)

    if closest_power_of_2 != num_heads:
        extra_base = torch.tensor(
            2 ** (-(2 ** -(math.log2(2 * closest_power_of_2) - 3))),
            device=attention_mask.device,
            dtype=torch.float32,
        )
        num_remaining_heads = min(closest_power_of_2, num_heads - closest_power_of_2)
        extra_powers = torch.arange(
            1,
            1 + 2 * num_remaining_heads,
            2,
            device=attention_mask.device,
            dtype=torch.int32,
        )
        slopes = torch.cat([slopes, torch.pow(extra_base, extra_powers)], dim=0)

    # Note: alibi will added to the attention bias that will be applied to the query, key product of attention
    # => therefore alibi will have to be of shape (batch_size, num_heads, query_length, key_length)
    # => here we set (batch_size=1, num_heads=num_heads, query_length=1, key_length=max_length)
    # => the query_length dimension will then be broadcasted correctly
    # This is more or less identical to T5's relative position bias:
    # https://github.com/huggingface/transformers/blob/f681437203baa7671de3174b0fa583c349d9d5e1/src/transformers/models/t5/modeling_t5.py#L527
    arange_tensor = ((attention_mask.cumsum(dim=-1) - 1) * attention_mask)[:, None, :]
    alibi = slopes[..., None] * arange_tensor
    return alibi.reshape(batch_size * num_heads, 1, seq_length).to(dtype)


class CustomT5Stack(T5Stack):
    def __init__(self, config, embed_tokens=None):
        super(T5Stack, self).__init__(config)

        self.embed_tokens = embed_tokens
        self.is_decoder = config.is_decoder
        self.position_encoding_type = getattr(
            config, "position_encoding_type", POSITION_ENCODING_REL_T5_BIAS
        )

        logger.info(f"position_encoding_type: {self.position_encoding_type}")

        self.block = nn.ModuleList(
            [
                CustomT5Block(config, has_relative_attention_bias=bool(i == 0))
                for i in range(config.num_layers)
            ]
        )
        self.final_layer_norm = T5LayerNorm(
            config.d_model, eps=config.layer_norm_epsilon
        )
        self.dropout = nn.Dropout(config.dropout_rate)

        if self.position_encoding_type == POSITION_ENCODING_ABS_LEARNED:
            self.wpe = nn.Embedding(2048, config.d_model)
            parent_dir = Path(os.path.dirname(os.path.abspath(__file__)))
            learned_embed_file = parent_dir / "gpt_neo_125m_pos_embed.npy"
            if learned_embed_file.exists():
                logger.info(
                    "Loading position embedding from {}".format(learned_embed_file)
                )
                import numpy as np

                weight = np.load(str(learned_embed_file))
                self.wpe.weight.data.copy_(torch.from_numpy(weight))
                self.wpe.weight.requires_grad = False
            else:
                self.wpe.weight.data.normal_(
                    mean=0.0, std=config.initializer_factor * 1.0
                )

        if self.position_encoding_type == POSITION_ENCODING_ABS_SINUSOID:
            self.wpe = FixedAbsolutePositionalEmbedding(config.d_model)

        if self.position_encoding_type == POSITION_ENCODING_ROTARY_NEW:
            # Rotary dim is X percentage of d_head
            # Right now, we just hardcode X here following:
            # https://github.com/huggingface/transformers/blob/v4.26.0/src/transformers/models/gpt_neox/configuration_gpt_neox.py
            rotary_dim = int(config.d_kv * 0.25)
            self.fixed_rotary_embedding = FixedRotaryPositionalEmbedding(
                rotary_dim, max_position=4096
            )

        if self.position_encoding_type in [
            POSITION_ENCODING_ALiBi,
            POSITION_ENCODING_ALiBi_LEARNED,
        ]:
            maxpos = 2048
            attn_heads = config.num_heads
            if self.position_encoding_type == POSITION_ENCODING_ALiBi_LEARNED:
                self.learned_logslopes = nn.Parameter(
                    torch.log(torch.Tensor(self.get_slopes(attn_heads)))
                )
            else:
                slopes = torch.Tensor(self.get_slopes(attn_heads))
                alibi = slopes.unsqueeze(1).unsqueeze(1) * torch.arange(
                    maxpos
                ).unsqueeze(0).unsqueeze(0).expand(attn_heads, -1, -1)
                alibi = alibi.view(attn_heads, 1, maxpos)
                self.register_buffer("alibi", alibi)

        # Initialize weights and apply final processing
        self.post_init()
        # Model parallel
        self.model_parallel = False
        self.device_map = None
        self.gradient_checkpointing = False

        self.window_size = 80  # only used for none_windowed

    def _alibi_prepare_attn_mask(
        self,
        attention_mask: torch.Tensor,
        input_shape: Tuple[int, int],
        past_key_values_length: int,
    ) -> torch.BoolTensor:
        # create causal mask
        # [batch_size, seq_length] -> [batch_size, 1, tgt_length, src_length]
        combined_attention_mask = None
        device = attention_mask.device
        _, src_length = input_shape

        if src_length > 1:
            combined_attention_mask = _make_causal_mask(
                input_shape,
                device=device,
                past_key_values_length=past_key_values_length,
            )

        # [batch_size, seq_length] -> [batch_size, 1, tgt_length, src_length]
        expanded_attn_mask = _expand_mask(attention_mask, tgt_length=src_length)
        combined_attention_mask = (
            expanded_attn_mask
            if combined_attention_mask is None
            else expanded_attn_mask | combined_attention_mask
        )

        return combined_attention_mask

    def get_slopes(self, n):
        def get_slopes_power_of_2(n):
            start = 2 ** (-(2 ** -(math.log2(n) - 3)))
            ratio = start
            return [start * ratio**i for i in range(n)]

        if math.log2(n).is_integer():
            return get_slopes_power_of_2(
                n
            )  # In the paper, we only train models that have 2^a heads for some a. This function has
        else:  # some good properties that only occur when the input is a power of 2. To maintain that even
            closest_power_of_2 = 2 ** math.floor(
                math.log2(n)
            )  # when the number of heads is not a power of 2, we use this workaround.
            return (
                get_slopes_power_of_2(closest_power_of_2)
                + self.get_slopes(2 * closest_power_of_2)[0::2][
                    : n - closest_power_of_2
                ]
            )

    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,
        position_ids=None,
        return_dict=None,
    ):
        # Model parallel
        if self.model_parallel:
            torch.cuda.set_device(self.first_device)
            self.embed_tokens = self.embed_tokens.to(self.first_device)
        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)

        if self.position_encoding_type in [
            POSITION_ENCODING_ABS_LEARNED,
            POSITION_ENCODING_ABS_SINUSOID,
        ]:
            if position_ids is not None:
                position_ids = position_ids.view(-1, input_shape[-1])

            if past_key_values is None:
                past_length = 0
            else:
                past_length = past_key_values[0][0].size(-2)

            device = input_ids.device if input_ids is not None else inputs_embeds.device
            if position_ids is None:
                position_ids = torch.arange(
                    past_length,
                    input_shape[-1] + past_length,
                    dtype=torch.long,
                    device=device,
                )
                position_ids = position_ids.unsqueeze(0).view(-1, input_shape[-1])

            position_embeds = self.wpe(position_ids)
            inputs_embeds += position_embeds

        batch_size, seq_length = input_shape

        # `position_bias` is a just tensor that is passed to all attention layers
        position_bias = None

        # 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).to(
                inputs_embeds.device
            )
        if (
            self.is_decoder
            and encoder_attention_mask is None
            and encoder_hidden_states is not None
        ):
            encoder_seq_length = encoder_hidden_states.shape[1]
            encoder_attention_mask = torch.ones(
                batch_size,
                encoder_seq_length,
                device=inputs_embeds.device,
                dtype=torch.long,
            )

        if self.position_encoding_type == POSITION_ENCODING_ROTARY_NEW:
            if position_ids is not None:
                position_ids = position_ids.view(-1, input_shape[-1])

            if past_key_values is None:
                past_length = 0
            else:
                past_length = past_key_values[0][0].size(-2)

            device = input_ids.device if input_ids is not None else inputs_embeds.device
            if position_ids is None:
                position_ids = torch.arange(
                    past_length,
                    input_shape[-1] + past_length,
                    dtype=torch.long,
                    device=device,
                )
                position_ids = position_ids.unsqueeze(0).view(-1, input_shape[-1])

            sinusoidal_pos = self.fixed_rotary_embedding(position_ids)
            position_bias = sinusoidal_pos

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

        if self.position_encoding_type == POSITION_ENCODING_NONE_WINDOW:
            indices = torch.arange(seq_length, device=inputs_embeds.device)
            causal_mask = indices[:, None] >= indices
            window_mask = (
                (indices.unsqueeze(0) - indices.unsqueeze(0).T)
                .abs()
                .less(self.window_size)
            )
            causal_mask = causal_mask & window_mask
            attention_mask = causal_mask.int()

            # Repeat the mask for each sample in the batch
            attention_mask = attention_mask[None, :, :].expand(
                batch_size, seq_length, seq_length
            )

        # 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.
        extended_attention_mask = self.get_extended_attention_mask(
            attention_mask, input_shape, inputs_embeds.device
        )

        if self.position_encoding_type == POSITION_ENCODING_ALiBi:
            num_heads = self.config.num_heads
            if len(attention_mask.shape) == 3:
                # We need to make a default attention mask
                alibi_attention_mask = torch.ones(batch_size, mask_seq_length).to(
                    inputs_embeds.device
                )
            else:
                alibi_attention_mask = attention_mask

            alibi = build_alibi_tensor(
                alibi_attention_mask, num_heads, dtype=inputs_embeds.dtype
            )
            position_bias = alibi
            del alibi_attention_mask

        if self.position_encoding_type in [POSITION_ENCODING_ALiBi_LEARNED]:
            if not hasattr(self, "alibi"):
                maxpos = 2048
                attn_heads = self.config.num_heads
                slopes = self.learned_logslopes.exp()
                alibi = slopes.unsqueeze(1).unsqueeze(1) * torch.arange(
                    maxpos, device=slopes.device
                ).unsqueeze(0).unsqueeze(0).expand(attn_heads, -1, -1)
                alibi = alibi.view(attn_heads, 1, maxpos)
            else:
                alibi = self.alibi

            alibi = alibi.unsqueeze(0).repeat(batch_size, 1, 1, 1)
            alibi = alibi[:, :, :, : attention_mask.shape[-1]]
            alibi = alibi.repeat(1, 1, extended_attention_mask.shape[2], 1)
            extended_attention_mask = torch.where(
                extended_attention_mask == 0,
                alibi,
                extended_attention_mask.repeat(1, self.config.num_heads, 1, 1),
            )

        # 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

        # 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]
            # Model parallel
            if self.model_parallel:
                torch.cuda.set_device(hidden_states.device)
                # Ensure that attention_mask is always on the same device as hidden_states
                if attention_mask is not None:
                    attention_mask = attention_mask.to(hidden_states.device)
                if position_bias is not None:
                    position_bias = position_bias.to(hidden_states.device)
                if encoder_hidden_states is not None:
                    encoder_hidden_states = encoder_hidden_states.to(
                        hidden_states.device
                    )
                if encoder_extended_attention_mask is not None:
                    encoder_extended_attention_mask = (
                        encoder_extended_attention_mask.to(hidden_states.device)
                    )
                if encoder_decoder_position_bias is not None:
                    encoder_decoder_position_bias = encoder_decoder_position_bias.to(
                        hidden_states.device
                    )
                if layer_head_mask is not None:
                    layer_head_mask = layer_head_mask.to(hidden_states.device)
                if cross_attn_layer_head_mask is not None:
                    cross_attn_layer_head_mask = cross_attn_layer_head_mask.to(
                        hidden_states.device
                    )
            if output_hidden_states:
                all_hidden_states = all_hidden_states + (hidden_states,)

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

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

                    return custom_forward

                layer_outputs = checkpoint(
                    create_custom_forward(layer_module),
                    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
                )
            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 + (None,)

            # Model Parallel: If it's the last layer for that device, put things on the next device
            if self.model_parallel:
                for k, v in self.device_map.items():
                    if i == v[-1] and "cuda:" + str(k) != self.last_device:
                        hidden_states = hidden_states.to("cuda:" + str(k + 1))

        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,
        )


class CustomDecoderOnlyT5(T5PreTrainedModel):
    config_class = CustomT5Config
    _keys_to_ignore_on_load_missing = [
        r"decoder\.embed_tokens\.weight",
        r"encoder",
        r"lm_head\.weight",
    ]
    _keys_to_ignore_on_load_unexpected = [
        r"decoder\.block\.0\.layer\.1\.EncDecAttention\.relative_attention_bias\.weight",
    ]

    def __init__(
        self,
        config=None,
        output_non_reduced_loss: bool = False,
        **kwargs,
    ):
        assert config is not None
        config.is_decoder = True
        config.is_encoder_decoder = False

        assert (
            config.position_encoding_type is not None
        ), "Position encoding type must be set"

        self.output_non_reduced_loss = output_non_reduced_loss
        self.main_input_name = "input_ids"

        super().__init__(config)

        self.model_dim = config.d_model

        self.shared = nn.Embedding(config.vocab_size, config.d_model)
        self.decoder = CustomT5Stack(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()

        # Model parallel
        self.model_parallel = False
        self.device_map = None
        #
        cross_attention_params = [
            p
            for n, p in self.decoder.named_parameters()
            if n.startswith("block.") and ".layer.1." in n
        ]
        for param in cross_attention_params:
            param.requires_grad = False

        # self.handle_tokenizer(tokenizer)

    def get_decoder(self):
        return self.decoder

    def parallelize(self, device_map=None):
        self.device_map = (
            get_device_map(len(self.encoder.block), range(torch.cuda.device_count()))
            if device_map is None
            else device_map
        )
        assert_device_map(self.device_map, len(self.encoder.block))
        self.encoder.parallelize(self.device_map)
        self.decoder.parallelize(self.device_map)
        self.lm_head = self.lm_head.to(self.decoder.first_device)
        self.model_parallel = True

    def deparallelize(self):
        self.encoder.deparallelize()
        self.decoder.deparallelize()
        self.encoder = self.encoder.to("cpu")
        self.decoder = self.decoder.to("cpu")
        self.lm_head = self.lm_head.to("cpu")
        self.model_parallel = False
        self.device_map = None
        torch.cuda.empty_cache()

    def get_input_embeddings(self):
        return self.shared

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

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

    def get_output_embeddings(self):
        return self.lm_head

    def prepare_inputs_for_generation(self, input_ids, past=None, **kwargs):
        token_type_ids = kwargs.get("token_type_ids", None)
        # only last token for inputs_ids if past is defined in kwargs
        if past:
            input_ids = input_ids[:, -1].unsqueeze(-1)
            if token_type_ids is not None:
                token_type_ids = token_type_ids[:, -1].unsqueeze(-1)

        attention_mask = kwargs.get("attention_mask", None)
        position_ids = kwargs.get("position_ids", None)

        if attention_mask is not None and position_ids is None:
            # create position_ids on the fly for batch generation
            position_ids = attention_mask.long().cumsum(-1) - 1
            position_ids.masked_fill_(attention_mask == 0, 1)
            if past:
                position_ids = position_ids[:, -1].unsqueeze(-1)
        else:
            position_ids = None

        return {
            "input_ids": input_ids,
            "past_key_values": past,
            "use_cache": kwargs.get("use_cache"),
            "attention_mask": attention_mask,
            "token_type_ids": token_type_ids,
            "position_ids": position_ids,
        }

    def forward(
        self,
        input_ids=None,
        past_key_values=None,
        attention_mask=None,
        token_type_ids=None,
        position_ids=None,
        head_mask=None,
        inputs_embeds=None,
        labels=None,
        use_cache=None,
        output_attentions=None,
        output_hidden_states=None,
        return_dict=None,
    ):
        return_dict = (
            return_dict if return_dict is not None else self.config.use_return_dict
        )

        if self.model_parallel:
            torch.cuda.set_device(self.decoder.first_device)

        if self.model_parallel:
            torch.cuda.set_device(self.decoder.first_device)
            if input_ids is not None:
                input_ids = input_ids.to(self.decoder.first_device)
            if attention_mask is not None:
                attention_mask = attention_mask.to(self.decoder.first_device)

        transformer_outputs = self.decoder(
            input_ids=input_ids,
            attention_mask=attention_mask,
            inputs_embeds=inputs_embeds,
            past_key_values=past_key_values,
            position_ids=position_ids,
            encoder_hidden_states=None,
            encoder_attention_mask=None,
            head_mask=head_mask,
            cross_attn_head_mask=None,
            use_cache=use_cache,
            output_attentions=output_attentions,
            output_hidden_states=output_hidden_states,
            return_dict=return_dict,
        )
        hidden_states = transformer_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
            hidden_states = hidden_states * (self.model_dim**-0.5)

        lm_logits = self.lm_head(hidden_states)

        loss = None
        non_reduced_loss = None
        if labels is not None:
            # Compute loss in fp32 to match with mesh-tf version
            # https://github.com/EleutherAI/gpt-neo/blob/89ce74164da2fb16179106f54e2269b5da8db333/models/gpt2/gpt2.py#L179
            lm_logits = lm_logits.to(torch.float32)

            # Shift so that tokens < n predict n
            shift_logits = lm_logits[..., :-1, :].contiguous()
            shift_labels = labels[..., 1:].contiguous()
            # Flatten the tokens
            loss_fct = CrossEntropyLoss()
            loss = loss_fct(
                shift_logits.view(-1, shift_logits.size(-1)), shift_labels.view(-1)
            )

            lm_logits = lm_logits.to(hidden_states.dtype)
            loss = loss.to(hidden_states.dtype)

            if self.output_non_reduced_loss:
                loss_fct = CrossEntropyLoss(reduction="none")
                non_reduced_loss = loss_fct(
                    shift_logits.view(-1, shift_logits.size(-1)), shift_labels.view(-1)
                )

                # Reshape to [batch_size, seq_length - 1]
                non_reduced_loss = non_reduced_loss.view(
                    shift_labels.shape[0], shift_labels.shape[1]
                )[:, -1].view(-1, 1)

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

        return CausalLMOutputWithPastAndLoss(
            loss=loss,
            logits=lm_logits,
            past_key_values=transformer_outputs.past_key_values,
            hidden_states=transformer_outputs.hidden_states,
            attentions=transformer_outputs.attentions,
            non_reduced_loss=non_reduced_loss,
        )

    @staticmethod
    def _reorder_cache(
        past: Tuple[Tuple[torch.Tensor]], beam_idx: torch.Tensor
    ) -> Tuple[Tuple[torch.Tensor]]:
        """
        This function is used to re-order the `past_key_values` cache if [`~PretrainedModel.beam_search`] or
        [`~PretrainedModel.beam_sample`] is called. This is required to match `past_key_values` with the correct
        beam_idx at every generation step.
        """
        return tuple(
            tuple(
                past_state.index_select(0, beam_idx.to(past_state.device))
                for past_state in layer_past
            )
            for layer_past in past
        )