backend/modeling_gpt2.py

# coding=utf-8
# Copyright 2018 The OpenAI Team Authors and HuggingFace Inc. team.
# Copyright (c) 2018, NVIDIA CORPORATION.  All rights reserved.
#
# Licensed under the Apache License, Version 2.0 (the "License");
# you may not use this file except in compliance with the License.
# You may obtain a copy of the License at
#
#     http://www.apache.org/licenses/LICENSE-2.0
#
# Unless required by applicable law or agreed to in writing, software
# distributed under the License is distributed on an "AS IS" BASIS,
# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
# See the License for the specific language governing permissions and
# limitations under the License.

"""
PyTorch OpenAI GPT-2 model.
Adapted from https://github.com/huggingface/transformers/blob/v4.0.1/src/transformers/models/gpt2/modeling_gpt2.py
and https://github.com/ghosthamlet/gpt2-ml-torch/blob/master/gpt2_ml_torch/modeling_gpt2.py
"""


import logging
import os
from dataclasses import dataclass
from typing import List, Optional, Tuple

import torch
import torch.nn as nn
from torch.nn import CrossEntropyLoss, MSELoss
from transformers import CONFIG_NAME, WEIGHTS_NAME, GPT2Config, GPT2Model
from transformers.activations import ACT2FN
from transformers.file_utils import (
    ModelOutput,
    add_code_sample_docstrings,
    add_start_docstrings,
    add_start_docstrings_to_model_forward,
    replace_return_docstrings,
)
from transformers.modeling_outputs import (
    BaseModelOutputWithPastAndCrossAttentions,
    CausalLMOutputWithCrossAttentions,
    SequenceClassifierOutputWithPast,
    TokenClassifierOutput,
)
from transformers.modeling_utils import (
    Conv1D,
    PreTrainedModel,
    SequenceSummary,
    find_pruneable_heads_and_indices,
    prune_conv1d_layer,
)
from transformers.utils.model_parallel_utils import assert_device_map, get_device_map

# THe Difference from Transformers is code under _USE_GROVER
_USE_GROVER = True

logger = logging.getLogger(__name__)

_CONFIG_FOR_DOC = "GPT2Config"
_TOKENIZER_FOR_DOC = "GPT2Tokenizer"

GPT2_PRETRAINED_MODEL_ARCHIVE_LIST = [
    "gpt2",
    "gpt2-medium",
    "gpt2-large",
    "gpt2-xl",
    "distilgpt2",
    # See all GPT-2 models at https://huggingface.co/models?filter=gpt2
]

logger.setLevel(logging.INFO)
console = logging.StreamHandler()
console.setLevel(logging.INFO)
logger.addHandler(console)

_GPT2_ML_TF_TO_TORCH = {
    "LayerNorm_embed_norm": "emb_norm",
    "pos_embed": "wpe.weight",
    "word_embed": "wte.weight",
    "layer": "h",
    # Most importently This two layer norm must be put on the same position as gpt2-ml
    # or generated data is bad, just repeat the last token
    "LayerNorm_mlp_ln0": "ln_1",
    "LayerNorm_mlp_ln1": "ln_2",
    "intermediate": "mlp.c_fc",
    "output": "mlp.c_proj",
    "query_layer": "attn.c_attn",
    "key_layer": "attn.c_attn",
    "value_layer": "attn.c_attn",
    "context_projection_layer": "attn.c_proj",
    "gamma": "weight",
    "kernel": "weight",
    "beta": "bias",
    "bias": "bias",
}


def convert_gpt2_checkpoint_to_pytorch(
    gpt2_checkpoint_path, gpt2_config_file, pytorch_dump_folder_path
):
    # Construct model
    if gpt2_config_file == "":
        config = GPT2Config()
    else:
        config = GPT2Config.from_json_file(gpt2_config_file)
    model = GPT2Model(config)

    # Load weights from numpy
    load_tf_weights_in_gpt2(model, config, gpt2_checkpoint_path)

    # Save pytorch-model
    pytorch_weights_dump_path = pytorch_dump_folder_path + "/" + WEIGHTS_NAME
    pytorch_config_dump_path = pytorch_dump_folder_path + "/" + CONFIG_NAME
    print("Save PyTorch model to {}".format(pytorch_weights_dump_path))
    torch.save(model.state_dict(), pytorch_weights_dump_path)
    print("Save configuration file to {}".format(pytorch_config_dump_path))
    with open(pytorch_config_dump_path, "w", encoding="utf-8") as f:
        f.write(config.to_json_string())


# XXX: MUST do like: convert_gpt2_checkpoint_to_pytorch('./model.ckpt-100000', './mega.json', './')
#      https://github.com/tensorflow/models/issues/2675#issuecomment-516595597
def load_tf_weights_in_gpt2(model, config, gpt2_checkpoint_path):
    """Load tf checkpoints in a pytorch model"""
    try:
        import re

        import tensorflow as tf
    except ImportError:
        logger.error(
            "Loading a TensorFlow model in PyTorch, requires TensorFlow to be installed. Please see "
            "https://www.tensorflow.org/install/ for installation instructions."
        )
        raise
    tf_path = os.path.abspath(gpt2_checkpoint_path)
    logger.info("Converting TensorFlow checkpoint from {}".format(tf_path))
    # Load weights from TF model
    init_vars = tf.train.list_variables(tf_path)
    names = []
    arrays = []
    for name, shape in init_vars:
        logger.info("Loading TF weight {} with shape {}".format(name, shape))
        array = tf.train.load_variable(tf_path, name)
        names.append(name)
        arrays.append(array.squeeze())

    import copy

    orig_model = copy.deepcopy(model)

    for name, array in zip(names, arrays):
        name = name[6:]  # skip "model/"
        name = name.split("/")
        pointer = model

        attn_layer = ""
        for m_name in name:
            if re.fullmatch(r"[A-Za-z]+\d+", m_name):
                scope_names = re.split(r"(\d+)", m_name)
            else:
                scope_names = [m_name]
            sname = scope_names[0]

            if sname == "" or sname == "embeddings":
                continue
            elif sname not in _GPT2_ML_TF_TO_TORCH:
                print("=========================================================")
                logger.info("Skip var name {}".format(scope_names))
                pointer = None
                break
            else:
                tname = _GPT2_ML_TF_TO_TORCH[sname]
                if "." in tname:
                    parent, child = tname.split(".")
                    pointer = getattr(pointer, parent)
                    pointer = getattr(pointer, child)
                else:
                    pointer = getattr(pointer, tname)

                if tname == "attn.c_attn":
                    attn_layer = sname

            if len(scope_names) >= 2:
                num = int(scope_names[1])
                pointer = pointer[num]

        if pointer is None:
            continue
        if attn_layer == "":
            try:
                assert pointer.shape == array.shape
            except AssertionError as e:
                e.args += (pointer.shape, array.shape)
                raise
        logger.info(
            "Initialize PyTorch weight {}, {}, {}".format(
                name, array.mean(), pointer.mean()
            )
        )
        if attn_layer == "":
            pointer.data = torch.from_numpy(array)
        else:
            shape = pointer.shape
            d = torch.from_numpy(array)
            is_bias = len(shape) == 1
            end = int(shape[0 if is_bias else 1] / 3)
            m = dict(
                query_layer=0,
                key_layer=end,
                value_layer=end * 2,
            )
            start = m[attn_layer]
            end = start + end
            if is_bias:
                pointer.data[start:end] = d
            else:
                pointer.data[:, start:end] = d
        logger.info(
            "Initialize PyTorch weight {}, {}, {}".format(
                name, array.mean(), pointer.mean()
            )
        )

    for name, params in orig_model.named_parameters():
        for n, p in model.named_parameters():
            if name == n:
                if params.equal(p):
                    print("--------------------------")
                    print(" %s not changed!" % n)
    return model


class Attention(nn.Module):
    def __init__(self, nx, n_ctx, config, scale=False, is_cross_attention=False):
        super().__init__()

        n_state = nx  # in Attention: n_state=768 (nx=n_embd)
        # [switch nx => n_state from Block to Attention to keep identical to TF implem]
        assert n_state % config.n_head == 0
        self.register_buffer(
            "bias",
            torch.tril(torch.ones((n_ctx, n_ctx), dtype=torch.uint8)).view(
                1, 1, n_ctx, n_ctx
            ),
        )
        self.register_buffer("masked_bias", torch.tensor(-1e4))
        self.n_head = config.n_head
        self.split_size = n_state
        self.scale = scale
        self.is_cross_attention = is_cross_attention
        if self.is_cross_attention:
            self.c_attn = Conv1D(2 * n_state, nx)
            self.q_attn = Conv1D(n_state, nx)
        else:
            self.c_attn = Conv1D(3 * n_state, nx)
        self.c_proj = Conv1D(n_state, nx)
        self.attn_dropout = nn.Dropout(config.attn_pdrop)
        self.resid_dropout = nn.Dropout(config.resid_pdrop)
        self.pruned_heads = set()

    def prune_heads(self, heads):
        if len(heads) == 0:
            return
        heads, index = find_pruneable_heads_and_indices(
            heads, self.n_head, self.split_size // self.n_head, self.pruned_heads
        )
        index_attn = torch.cat(
            [index, index + self.split_size, index + (2 * self.split_size)]
        )

        # Prune conv1d layers
        self.c_attn = prune_conv1d_layer(self.c_attn, index_attn, dim=1)
        self.c_proj = prune_conv1d_layer(self.c_proj, index, dim=0)

        # Update hyper params
        self.split_size = (self.split_size // self.n_head) * (self.n_head - len(heads))
        self.n_head = self.n_head - len(heads)
        self.pruned_heads = self.pruned_heads.union(heads)

    def _attn(
        self, q, k, v, attention_mask=None, head_mask=None, output_attentions=False
    ):
        w = torch.matmul(q, k)
        if self.scale:
            w = w / (float(v.size(-1)) ** 0.5)
        nd, ns = w.size(-2), w.size(-1)

        if not self.is_cross_attention:
            # if only "normal" attention layer implements causal mask
            mask = self.bias[:, :, ns - nd : ns, :ns]
            w = torch.where(mask.bool(), w, self.masked_bias.to(w.dtype))

        if attention_mask is not None:
            # Apply the attention mask
            w = w + attention_mask

        w = nn.Softmax(dim=-1)(w)
        w = self.attn_dropout(w)

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

        outputs = [torch.matmul(w, v)]
        if output_attentions:
            outputs.append(w)
        return outputs

    def merge_heads(self, x):
        x = x.permute(0, 2, 1, 3).contiguous()
        new_x_shape = x.size()[:-2] + (x.size(-2) * x.size(-1),)
        return x.view(*new_x_shape)  # in Tensorflow implem: fct merge_states

    def split_heads(self, x, k=False):
        new_x_shape = x.size()[:-1] + (self.n_head, x.size(-1) // self.n_head)
        x = x.view(*new_x_shape)  # in Tensorflow implem: fct split_states
        if k:
            return x.permute(0, 2, 3, 1)  # (batch, head, head_features, seq_length)
        else:
            return x.permute(0, 2, 1, 3)  # (batch, head, seq_length, head_features)

    def forward(
        self,
        hidden_states,
        layer_past=None,
        attention_mask=None,
        head_mask=None,
        encoder_hidden_states=None,
        encoder_attention_mask=None,
        use_cache=False,
        output_attentions=False,
    ):
        if encoder_hidden_states is not None:
            assert hasattr(
                self, "q_attn"
            ), "If class is used as cross attention, the weights `q_attn` have to be defined. Please make sure to instantiate class with `Attention(..., is_cross_attention=True)`."
            query = self.q_attn(hidden_states)
            key, value = self.c_attn(encoder_hidden_states).split(
                self.split_size, dim=2
            )
            attention_mask = encoder_attention_mask
        else:
            query, key, value = self.c_attn(hidden_states).split(self.split_size, dim=2)

        query = self.split_heads(query)
        key = self.split_heads(key, k=True)
        value = self.split_heads(value)
        if layer_past is not None:
            past_key, past_value = (
                layer_past[0].transpose(-2, -1),
                layer_past[1],
            )  # transpose back cf below
            key = torch.cat((past_key, key), dim=-1)
            value = torch.cat((past_value, value), dim=-2)

        if use_cache is True:
            present = torch.stack(
                (key.transpose(-2, -1), value)
            )  # transpose to have same shapes for stacking
        else:
            present = (None,)

        attn_outputs = self._attn(
            query, key, value, attention_mask, head_mask, output_attentions
        )
        a = attn_outputs[0]

        a = self.merge_heads(a)
        a = self.c_proj(a)
        a = self.resid_dropout(a)

        outputs = [a, present] + attn_outputs[1:]
        return outputs  # a, present, (attentions)


class MLP(nn.Module):
    def __init__(self, n_state, config):  # in MLP: n_state=3072 (4 * n_embd)
        super().__init__()
        nx = config.n_embd
        self.c_fc = Conv1D(n_state, nx)
        self.c_proj = Conv1D(nx, n_state)
        self.act = ACT2FN[config.activation_function]
        self.dropout = nn.Dropout(config.resid_pdrop)

    def forward(self, x):
        h = self.act(self.c_fc(x))
        h2 = self.c_proj(h)
        return self.dropout(h2)


class Block(nn.Module):
    def __init__(self, n_ctx, config, scale=False):
        super().__init__()
        hidden_size = config.n_embd
        inner_dim = config.n_inner if config.n_inner is not None else 4 * hidden_size
        self.ln_1 = nn.LayerNorm(hidden_size, eps=config.layer_norm_epsilon)
        self.attn = Attention(hidden_size, n_ctx, config, scale)
        self.ln_2 = nn.LayerNorm(hidden_size, eps=config.layer_norm_epsilon)
        if config.add_cross_attention:
            self.crossattention = Attention(
                hidden_size, n_ctx, config, scale, is_cross_attention=True
            )
            self.ln_cross_attn = nn.LayerNorm(
                hidden_size, eps=config.layer_norm_epsilon
            )
        self.mlp = MLP(inner_dim, config)

    def forward(
        self,
        hidden_states,
        layer_past=None,
        attention_mask=None,
        head_mask=None,
        encoder_hidden_states=None,
        encoder_attention_mask=None,
        use_cache=False,
        output_attentions=False,
    ):
        attn_outputs = self.attn(
            hidden_states,
            layer_past=layer_past,
            attention_mask=attention_mask,
            head_mask=head_mask,
            use_cache=use_cache,
            output_attentions=output_attentions,
        )
        attn_output = attn_outputs[0]  # output_attn: a, present, (attentions)
        outputs = attn_outputs[1:]
        # residual connection
        hidden_states = attn_output + hidden_states

        if encoder_hidden_states is not None:
            # add one self-attention block for cross-attention
            assert hasattr(
                self, "crossattention"
            ), f"If `encoder_hidden_states` are passed, {self} has to be instantiated with cross-attention layers by setting `config.add_cross_attention=True`"
            cross_attn_outputs = self.crossattention(
                self.ln_cross_attn(hidden_states),
                attention_mask=attention_mask,
                head_mask=head_mask,
                encoder_hidden_states=encoder_hidden_states,
                encoder_attention_mask=encoder_attention_mask,
                output_attentions=output_attentions,
            )
            attn_output = cross_attn_outputs[0]
            # residual connection
            hidden_states = hidden_states + attn_output
            outputs = (
                outputs + cross_attn_outputs[2:]
            )  # add cross attentions if we output attention weights

        feed_forward_hidden_states = self.mlp(self.ln_1(hidden_states))
        # residual connection
        hidden_states = hidden_states + feed_forward_hidden_states

        hidden_states = self.ln_2(hidden_states)

        outputs = [hidden_states] + outputs
        return outputs  # hidden_states, present, (attentions, cross_attentions)


class GPT2PreTrainedModel(PreTrainedModel):
    """
    An abstract class to handle weights initialization and a simple interface for downloading and loading pretrained
    models.
    """

    config_class = GPT2Config
    load_tf_weights = load_tf_weights_in_gpt2
    base_model_prefix = "transformer"
    is_parallelizable = True

    def __init__(self, *inputs, **kwargs):
        super().__init__(*inputs, **kwargs)

    def _init_weights(self, module):
        """Initialize the weights."""
        if isinstance(module, (nn.Linear, nn.Embedding, Conv1D)):
            # Slightly different from the TF version which uses truncated_normal for initialization
            # cf https://github.com/pytorch/pytorch/pull/5617
            module.weight.data.normal_(mean=0.0, std=self.config.initializer_range)
            if isinstance(module, (nn.Linear, Conv1D)) and module.bias is not None:
                module.bias.data.zero_()
        elif isinstance(module, nn.LayerNorm):
            module.bias.data.zero_()
            module.weight.data.fill_(1.0)


@dataclass
class GPT2DoubleHeadsModelOutput(ModelOutput):
    """
    Base class for outputs of models predicting if two sentences are consecutive or not.

    Args:
        loss (:obj:`torch.FloatTensor` of shape :obj:`(1,)`, `optional`, returned when ``labels`` is provided):
            Language modeling loss.
        mc_loss (:obj:`torch.FloatTensor` of shape :obj:`(1,)`, `optional`, returned when :obj:`mc_labels` is provided):
            Multiple choice classification loss.
        logits (:obj:`torch.FloatTensor` of shape :obj:`(batch_size, num_choices, sequence_length, config.vocab_size)`):
            Prediction scores of the language modeling head (scores for each vocabulary token before SoftMax).
        mc_logits (:obj:`torch.FloatTensor` of shape :obj:`(batch_size, num_choices)`):
            Prediction scores of the multiple choice classification head (scores for each choice before SoftMax).
        past_key_values (:obj:`List[torch.FloatTensor]`, `optional`, returned when ``use_cache=True`` is passed or when ``config.use_cache=True``):
            List of :obj:`torch.FloatTensor` of length :obj:`config.n_layers`, with each tensor of shape :obj:`(2,
            batch_size, num_heads, sequence_length, embed_size_per_head)`).

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

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

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

    loss: Optional[torch.FloatTensor] = None
    mc_loss: Optional[torch.FloatTensor] = None
    logits: torch.FloatTensor = None
    mc_logits: torch.FloatTensor = None
    past_key_values: Optional[List[torch.FloatTensor]] = None
    hidden_states: Optional[Tuple[torch.FloatTensor]] = None
    attentions: Optional[Tuple[torch.FloatTensor]] = None


GPT2_START_DOCSTRING = r"""

    This model inherits from :class:`~transformers.PreTrainedModel`. Check the superclass documentation for the generic
    methods the library implements for all its model (such as downloading or saving, resizing the input embeddings,
    pruning heads etc.)

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

    Parameters:
        config (:class:`~transformers.GPT2Config`): Model configuration class with all the parameters of the model.
            Initializing with a config file does not load the weights associated with the model, only the
            configuration. Check out the :meth:`~transformers.PreTrainedModel.from_pretrained` method to load the model
            weights.
"""

GPT2_INPUTS_DOCSTRING = r"""
    Args:
        input_ids (:obj:`torch.LongTensor` of shape :obj:`(batch_size, input_ids_length)`):
            :obj:`input_ids_length` = ``sequence_length`` if :obj:`past_key_values` is ``None`` else
            ``past_key_values[0].shape[-2]`` (``sequence_length`` of input past key value states). Indices of input
            sequence tokens in the vocabulary.

            If :obj:`past_key_values` is used, only ``input_ids`` that do not have their past calculated should be
            passed as ``input_ids``.

            Indices can be obtained using :class:`~transformers.GPT2Tokenizer`. See
            :meth:`transformers.PreTrainedTokenizer.encode` and :meth:`transformers.PreTrainedTokenizer.__call__` for
            details.

            `What are input IDs? <../glossary.html#input-ids>`__
        past_key_values (:obj:`List[torch.FloatTensor]` of length :obj:`config.n_layers`):
            Contains precomputed hidden-states (key and values in the attention blocks) as computed by the model (see
            :obj:`past_key_values` output below). Can be used to speed up sequential decoding. The ``input_ids`` which
            have their past given to this model should not be passed as ``input_ids`` as they have already been
            computed.
        attention_mask (:obj:`torch.FloatTensor` of shape :obj:`(batch_size, sequence_length)`, `optional`):
            Mask to avoid performing attention on padding token indices. Mask values selected in ``[0, 1]``:

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

            `What are attention masks? <../glossary.html#attention-mask>`__
        token_type_ids (:obj:`torch.LongTensor` of shape :obj:`(batch_size, input_ids_length)`, `optional`):
            Segment token indices to indicate first and second portions of the inputs. Indices are selected in ``[0,
            1]``:

            - 0 corresponds to a `sentence A` token,
            - 1 corresponds to a `sentence B` token.

            `What are token type IDs? <../glossary.html#token-type-ids>`_
        position_ids (:obj:`torch.LongTensor` of shape :obj:`(batch_size, sequence_length)`, `optional`):
            Indices of positions of each input sequence tokens in the position embeddings. Selected in the range ``[0,
            config.max_position_embeddings - 1]``.

            `What are position IDs? <../glossary.html#position-ids>`_
        head_mask (:obj:`torch.FloatTensor` of shape :obj:`(num_heads,)` or :obj:`(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 (:obj:`torch.FloatTensor` of shape :obj:`(batch_size, sequence_length, hidden_size)`, `optional`):
            Optionally, instead of passing :obj:`input_ids` you can choose to directly pass an embedded representation.
            This is useful if you want more control over how to convert :obj:`input_ids` indices into associated
            vectors than the model's internal embedding lookup matrix.

            If :obj:`past_key_values` is used, optionally only the last :obj:`inputs_embeds` have to be input (see
            :obj:`past_key_values`).
        use_cache (:obj:`bool`, `optional`):
            If set to :obj:`True`, :obj:`past_key_values` key value states are returned and can be used to speed up
            decoding (see :obj:`past_key_values`).
        output_attentions (:obj:`bool`, `optional`):
            Whether or not to return the attentions tensors of all attention layers. See ``attentions`` under returned
            tensors for more detail.
        output_hidden_states (:obj:`bool`, `optional`):
            Whether or not to return the hidden states of all layers. See ``hidden_states`` under returned tensors for
            more detail.
        return_dict (:obj:`bool`, `optional`):
            Whether or not to return a :class:`~transformers.file_utils.ModelOutput` instead of a plain tuple.
"""

PARALLELIZE_DOCSTRING = r"""
    This is an experimental feature and is a subject to change at a moment's notice.

    Uses a device map to distribute attention modules of the model across several devices. If no device map is given,
    it will evenly distribute blocks across all devices.

    Args:
        device_map (:obj:`Dict[int, list]`, optional, defaults to None):
            A dictionary that maps attention modules to devices. Note that the embedding module and LMHead are always
            automatically mapped to the first device (for esoteric reasons). That means that the first device should
            have fewer attention modules mapped to it than other devices. For reference, the gpt2 models have the
            following number of attention modules:

                - gpt2: 12
                - gpt2-medium: 24
                - gpt2-large: 36
                - gpt2-xl: 48

    Example::

            # Here is an example of a device map on a machine with 4 GPUs using gpt2-xl, which has a total of 48 attention modules:
            model = GPT2LMHeadModel.from_pretrained('gpt2-xl')
            device_map = {0: [0, 1, 2, 3, 4, 5, 6, 7, 8],

                          1: [9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19, 20, 21],
                          2: [22, 23, 24, 25, 26, 27, 28, 29, 30, 31, 32, 33, 34],
                          3: [35, 36, 37, 38, 39, 40, 41, 42, 43, 44, 45, 46, 47]}
            model.parallelize(device_map)
"""
DEPARALLELIZE_DOCSTRING = r"""
    Moves the model to cpu from a model parallel state.

    Example::

        # On a 4 GPU machine with gpt2-large:
        model = GPT2LMHeadModel.from_pretrained('gpt2-large')
        device_map = {0: [0, 1, 2, 3, 4, 5, 6, 7],

                    1: [8, 9, 10, 11, 12, 13, 14, 15],
                    2: [16, 17, 18, 19, 20, 21, 22, 23],
                    3: [24, 25, 26, 27, 28, 29, 30, 31, 32, 33, 34, 35]}
        model.parallelize(device_map) # Splits the model across several devices
        model.deparallelize() # Put the model back on cpu and cleans memory by calling torch.cuda.empty_cache()
"""


@add_start_docstrings(
    "The bare GPT2 Model transformer outputting raw hidden-states without any specific head on top.",
    GPT2_START_DOCSTRING,
)
class GPT2Model(GPT2PreTrainedModel):
    def __init__(self, config):
        super().__init__(config)

        self.wte = nn.Embedding(config.vocab_size, config.n_embd)
        self.wpe = nn.Embedding(config.n_positions, config.n_embd)
        if _USE_GROVER:
            self.emb_norm = nn.LayerNorm(config.n_embd, eps=config.layer_norm_epsilon)

        self.drop = nn.Dropout(config.embd_pdrop)
        self.h = nn.ModuleList(
            [Block(config.n_ctx, config, scale=True) for _ in range(config.n_layer)]
        )
        if not _USE_GROVER:
            self.ln_f = nn.LayerNorm(config.n_embd, eps=config.layer_norm_epsilon)

        self.init_weights()

        # Model parallel
        self.model_parallel = False
        self.device_map = None

    @add_start_docstrings(PARALLELIZE_DOCSTRING)
    def parallelize(self, device_map=None):
        # Check validity of device_map
        self.device_map = (
            get_device_map(len(self.h), range(torch.cuda.device_count()))
            if device_map is None
            else device_map
        )
        assert_device_map(self.device_map, len(self.h))
        self.model_parallel = True
        self.first_device = (
            "cpu"
            if "cpu" in self.device_map.keys()
            else "cuda:" + str(min(self.device_map.keys()))
        )
        self.last_device = "cuda:" + str(max(self.device_map.keys()))
        self.wte = self.wte.to(self.first_device)
        self.wpe = self.wpe.to(self.first_device)
        # Load onto devices
        for k, v in self.device_map.items():
            for block in v:
                cuda_device = "cuda:" + str(k)
                self.h[block] = self.h[block].to(cuda_device)
        # ln_f to last
        self.ln_f = self.ln_f.to(self.last_device)

    @add_start_docstrings(DEPARALLELIZE_DOCSTRING)
    def deparallelize(self):
        self.model_parallel = False
        self.device_map = None
        self.first_device = "cpu"
        self.last_device = "cpu"
        self.wte = self.wte.to("cpu")
        self.wpe = self.wpe.to("cpu")
        for index in range(len(self.h)):
            self.h[index] = self.h[index].to("cpu")
        self.ln_f = self.ln_f.to("cpu")
        torch.cuda.empty_cache()

    def get_input_embeddings(self):
        return self.wte

    def set_input_embeddings(self, new_embeddings):
        self.wte = new_embeddings

    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}
        """
        for layer, heads in heads_to_prune.items():
            self.h[layer].attn.prune_heads(heads)

    @add_start_docstrings_to_model_forward(GPT2_INPUTS_DOCSTRING)
    @add_code_sample_docstrings(
        tokenizer_class=_TOKENIZER_FOR_DOC,
        checkpoint="gpt2",
        output_type=BaseModelOutputWithPastAndCrossAttentions,
        config_class=_CONFIG_FOR_DOC,
    )
    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,
        encoder_hidden_states=None,
        encoder_attention_mask=None,
        use_cache=None,
        output_attentions=None,
        output_hidden_states=None,
        return_dict=None,
    ):
        output_attentions = (
            output_attentions
            if output_attentions is not None
            else self.config.output_attentions
        )
        output_hidden_states = (
            output_hidden_states
            if output_hidden_states is not None
            else self.config.output_hidden_states
        )
        use_cache = use_cache if use_cache is not None else self.config.use_cache
        return_dict = (
            return_dict if return_dict is not None else self.config.use_return_dict
        )

        if input_ids is not None and inputs_embeds is not None:
            raise ValueError(
                "You cannot specify both input_ids and inputs_embeds at the same time"
            )
        elif input_ids is not None:
            input_shape = input_ids.size()
            input_ids = input_ids.view(-1, input_shape[-1])
            batch_size = input_ids.shape[0]
        elif inputs_embeds is not None:
            input_shape = inputs_embeds.size()[:-1]
            batch_size = inputs_embeds.shape[0]
        else:
            raise ValueError("You have to specify either input_ids or inputs_embeds")

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

        if past_key_values is None:
            past_length = 0
            past_key_values = [None] * len(self.h)
        else:
            past_length = past_key_values[0][0].size(-2)
        if position_ids is None:
            device = input_ids.device if input_ids is not None else inputs_embeds.device
            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])

        # Attention mask.
        if attention_mask is not None:
            if batch_size <= 0:
                raise ValueError("batch_size has to be defined and > 0")
            attention_mask = attention_mask.view(batch_size, -1)
            # We create a 3D attention mask from a 2D tensor mask.
            # Sizes are [batch_size, 1, 1, to_seq_length]
            # So we can broadcast to [batch_size, num_heads, from_seq_length, to_seq_length]
            # this attention mask is more simple than the triangular masking of causal attention
            # used in OpenAI GPT, we just need to prepare the broadcast dimension here.
            attention_mask = attention_mask[:, None, None, :]

            # Since attention_mask is 1.0 for positions we want to attend and 0.0 for
            # masked positions, this operation will create a tensor which is 0.0 for
            # positions we want to attend and -10000.0 for masked positions.
            # Since we are adding it to the raw scores before the softmax, this is
            # effectively the same as removing these entirely.
            attention_mask = attention_mask.to(dtype=self.dtype)  # fp16 compatibility
            attention_mask = (1.0 - attention_mask) * -10000.0

        # If a 2D ou 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.config.add_cross_attention 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=device)
            encoder_attention_mask = self.invert_attention_mask(encoder_attention_mask)
        else:
            encoder_attention_mask = None

        # Prepare head mask if needed
        # 1.0 in head_mask indicate we keep the head
        # attention_probs has shape bsz x n_heads x N x N
        # head_mask has shape n_layer x batch x n_heads x N x N
        head_mask = self.get_head_mask(head_mask, self.config.n_layer)

        if inputs_embeds is None:
            inputs_embeds = self.wte(input_ids)
        position_embeds = self.wpe(position_ids)
        hidden_states = inputs_embeds + position_embeds

        if token_type_ids is not None:
            token_type_embeds = self.wte(token_type_ids)
            hidden_states = hidden_states + token_type_embeds

        hidden_states = self.drop(hidden_states)
        if _USE_GROVER:
            hidden_states = self.emb_norm(hidden_states)
        output_shape = input_shape + (hidden_states.size(-1),)

        presents = () if use_cache else None
        all_self_attentions = () if output_attentions else None
        all_cross_attentions = (
            () if output_attentions and self.config.add_cross_attention else None
        )
        all_hidden_states = () if output_hidden_states else None
        for i, (block, layer_past) in enumerate(zip(self.h, past_key_values)):

            # Model parallel
            if self.model_parallel:
                torch.cuda.set_device(hidden_states.device)
                # Ensure layer_past is on same device as hidden_states (might not be correct)
                if layer_past is not None:
                    layer_past = tuple(
                        past_state.to(hidden_states.device) for past_state in layer_past
                    )
                # 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 isinstance(head_mask, torch.Tensor):
                    head_mask = head_mask.to(hidden_states.device)

            if output_hidden_states:
                all_hidden_states = all_hidden_states + (
                    hidden_states.view(*output_shape),
                )

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

                def create_custom_forward(module):
                    def custom_forward(*inputs):
                        # checkpointing only works with tuple returns, not with lists
                        return tuple(
                            output
                            for output in module(*inputs, use_cache, output_attentions)
                        )

                    return custom_forward

                outputs = torch.utils.checkpoint.checkpoint(
                    create_custom_forward(block),
                    hidden_states,
                    layer_past,
                    attention_mask,
                    head_mask[i],
                    encoder_hidden_states,
                    encoder_attention_mask,
                )
            else:
                outputs = block(
                    hidden_states,
                    layer_past=layer_past,
                    attention_mask=attention_mask,
                    head_mask=head_mask[i],
                    encoder_hidden_states=encoder_hidden_states,
                    encoder_attention_mask=encoder_attention_mask,
                    use_cache=use_cache,
                    output_attentions=output_attentions,
                )

            hidden_states, present = outputs[:2]
            if use_cache is True:
                presents = presents + (present,)

            if output_attentions:
                all_self_attentions = all_self_attentions + (
                    outputs[2 if use_cache else 1],
                )
                if self.config.add_cross_attention:
                    all_cross_attentions = all_cross_attentions + (
                        outputs[3 if use_cache else 2],
                    )

            # 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))

        if not _USE_GROVER:
            hidden_states = self.ln_f(hidden_states)

        hidden_states = hidden_states.view(*output_shape)
        # Add last hidden state
        if output_hidden_states:
            all_hidden_states = all_hidden_states + (hidden_states,)

        if not return_dict:
            return tuple(
                v
                for v in [
                    hidden_states,
                    presents,
                    all_hidden_states,
                    all_self_attentions,
                    all_cross_attentions,
                ]
                if v is not None
            )

        return BaseModelOutputWithPastAndCrossAttentions(
            last_hidden_state=hidden_states,
            past_key_values=presents,
            hidden_states=all_hidden_states,
            attentions=all_self_attentions,
            cross_attentions=all_cross_attentions,
        )


@add_start_docstrings(
    """
    The GPT2 Model transformer with a language modeling head on top (linear layer with weights tied to the input
    embeddings).
    """,
    GPT2_START_DOCSTRING,
)
class GPT2LMHeadModel(GPT2PreTrainedModel):
    _keys_to_ignore_on_load_missing = [r"h\.\d+\.attn\.masked_bias", r"lm_head\.weight"]

    def __init__(self, config):
        super().__init__(config)
        self.transformer = GPT2Model(config)
        self.lm_head = nn.Linear(config.n_embd, config.vocab_size, bias=False)

        self.init_weights()

        # Model parallel
        self.model_parallel = False
        self.device_map = None

    @add_start_docstrings(PARALLELIZE_DOCSTRING)
    def parallelize(self, device_map=None):
        self.device_map = (
            get_device_map(len(self.transformer.h), range(torch.cuda.device_count()))
            if device_map is None
            else device_map
        )
        assert_device_map(self.device_map, len(self.transformer.h))
        self.transformer.parallelize(self.device_map)
        self.lm_head = self.lm_head.to(self.transformer.first_device)
        self.model_parallel = True

    @add_start_docstrings(DEPARALLELIZE_DOCSTRING)
    def deparallelize(self):
        self.transformer.deparallelize()
        self.transformer = self.transformer.to("cpu")
        self.lm_head = self.lm_head.to("cpu")
        self.model_parallel = False
        torch.cuda.empty_cache()

    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"),
            "position_ids": position_ids,
            "attention_mask": attention_mask,
            "token_type_ids": token_type_ids,
        }

    @add_start_docstrings_to_model_forward(GPT2_INPUTS_DOCSTRING)
    @add_code_sample_docstrings(
        tokenizer_class=_TOKENIZER_FOR_DOC,
        checkpoint="gpt2",
        output_type=CausalLMOutputWithCrossAttentions,
        config_class=_CONFIG_FOR_DOC,
    )
    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,
        encoder_hidden_states=None,
        encoder_attention_mask=None,
        labels=None,
        use_cache=None,
        output_attentions=None,
        output_hidden_states=None,
        return_dict=None,
    ):
        r"""
        labels (:obj:`torch.LongTensor` of shape :obj:`(batch_size, sequence_length)`, `optional`):
            Labels for language modeling. Note that the labels **are shifted** inside the model, i.e. you can set
            ``labels = input_ids`` Indices are selected in ``[-100, 0, ..., config.vocab_size]`` All labels set to
            ``-100`` are ignored (masked), the loss is only computed for labels in ``[0, ..., config.vocab_size]``
        """
        return_dict = (
            return_dict if return_dict is not None else self.config.use_return_dict
        )

        transformer_outputs = self.transformer(
            input_ids,
            past_key_values=past_key_values,
            attention_mask=attention_mask,
            token_type_ids=token_type_ids,
            position_ids=position_ids,
            head_mask=head_mask,
            inputs_embeds=inputs_embeds,
            encoder_hidden_states=encoder_hidden_states,
            encoder_attention_mask=encoder_attention_mask,
            use_cache=use_cache,
            output_attentions=output_attentions,
            output_hidden_states=output_hidden_states,
            return_dict=return_dict,
        )
        hidden_states = transformer_outputs[0]

        # Set device for model parallelism
        if self.model_parallel:
            torch.cuda.set_device(self.transformer.first_device)
            hidden_states = hidden_states.to(self.lm_head.weight.device)

        lm_logits = self.lm_head(hidden_states)

        loss = None
        if labels is not None:
            # 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)
            )

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

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

    @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 :obj:`past_key_values` cache if
        :meth:`~transformers.PreTrainedModel.beam_search` or :meth:`~transformers.PreTrainedModel.beam_sample` is
        called. This is required to match :obj:`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
        )


@add_start_docstrings(
    """
The GPT2 Model transformer with a language modeling and a multiple-choice classification head on top e.g. for
RocStories/SWAG tasks. The two heads are two linear layers. The language modeling head has its weights tied to the
input embeddings, the classification head takes as input the input of a specified classification token index in the
input sequence).
""",
    GPT2_START_DOCSTRING,
)
class GPT2DoubleHeadsModel(GPT2PreTrainedModel):
    def __init__(self, config):
        super().__init__(config)
        config.num_labels = 1
        self.transformer = GPT2Model(config)
        self.lm_head = nn.Linear(config.n_embd, config.vocab_size, bias=False)
        self.multiple_choice_head = SequenceSummary(config)

        self.init_weights()

        # Model parallel
        self.model_parallel = False
        self.device_map = None

    @add_start_docstrings(PARALLELIZE_DOCSTRING)
    def parallelize(self, device_map=None):
        self.device_map = (
            get_device_map(len(self.transformer.h), range(torch.cuda.device_count()))
            if device_map is None
            else device_map
        )
        assert_device_map(self.device_map, len(self.transformer.h))
        self.transformer.parallelize(self.device_map)
        self.lm_head = self.lm_head.to(self.transformer.first_device)
        self.multiple_choice_head = self.multiple_choice_head.to(
            self.transformer.first_device
        )
        self.model_parallel = True

    @add_start_docstrings(DEPARALLELIZE_DOCSTRING)
    def deparallelize(self):
        self.transformer.deparallelize()
        self.transformer = self.transformer.to("cpu")
        self.lm_head = self.lm_head.to("cpu")
        self.multiple_choice_head = self.multiple_choice_head.to("cpu")
        self.model_parallel = False
        torch.cuda.empty_cache()

    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"),
            "position_ids": position_ids,
            "attention_mask": attention_mask,
            "token_type_ids": token_type_ids,
        }

    @add_start_docstrings_to_model_forward(GPT2_INPUTS_DOCSTRING)
    @replace_return_docstrings(
        output_type=GPT2DoubleHeadsModelOutput, config_class=_CONFIG_FOR_DOC
    )
    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,
        mc_token_ids=None,
        labels=None,
        mc_labels=None,
        use_cache=None,
        output_attentions=None,
        output_hidden_states=None,
        return_dict=None,
        **kwargs,
    ):
        r"""
        mc_token_ids (:obj:`torch.LongTensor` of shape :obj:`(batch_size, num_choices)`, `optional`, default to index of the last token of the input):
            Index of the classification token in each input sequence. Selected in the range ``[0, input_ids.size(-1) -
            1[``.
        labels (:obj:`torch.LongTensor` of shape :obj:`(batch_size, sequence_length)`, `optional`):
            Labels for language modeling. Note that the labels **are shifted** inside the model, i.e. you can set
            ``labels = input_ids`` Indices are selected in ``[-1, 0, ..., config.vocab_size]`` All labels set to
            ``-100`` are ignored (masked), the loss is only computed for labels in ``[0, ..., config.vocab_size]``
        mc_labels (:obj:`torch.LongTensor` of shape :obj:`(batch_size)`, `optional`):
            Labels for computing the multiple choice classification loss. Indices should be in ``[0, ...,
            num_choices]`` where `num_choices` is the size of the second dimension of the input tensors. (see
            `input_ids` above)

        Return:

        Example::

            >>> import torch
            >>> from transformers import GPT2Tokenizer, GPT2DoubleHeadsModel

            >>> tokenizer = GPT2Tokenizer.from_pretrained('gpt2')
            >>> model = GPT2DoubleHeadsModel.from_pretrained('gpt2')

            >>> # Add a [CLS] to the vocabulary (we should train it also!)
            >>> num_added_tokens = tokenizer.add_special_tokens({'cls_token': '[CLS]'})

            >>> embedding_layer = model.resize_token_embeddings(len(tokenizer))  # Update the model embeddings with the new vocabulary size

            >>> choices = ["Hello, my dog is cute [CLS]", "Hello, my cat is cute [CLS]"]
            >>> encoded_choices = [tokenizer.encode(s) for s in choices]
            >>> cls_token_location = [tokens.index(tokenizer.cls_token_id) for tokens in encoded_choices]

            >>> input_ids = torch.tensor(encoded_choices).unsqueeze(0)  # Batch size: 1, number of choices: 2
            >>> mc_token_ids = torch.tensor([cls_token_location])  # Batch size: 1

            >>> outputs = model(input_ids, mc_token_ids=mc_token_ids)
            >>> lm_logits = outputs.lm_logits
            >>> mc_logits = outputs.mc_logits

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

        transformer_outputs = self.transformer(
            input_ids,
            past_key_values=past_key_values,
            attention_mask=attention_mask,
            token_type_ids=token_type_ids,
            position_ids=position_ids,
            head_mask=head_mask,
            inputs_embeds=inputs_embeds,
            use_cache=use_cache,
            output_attentions=output_attentions,
            output_hidden_states=output_hidden_states,
            return_dict=return_dict,
        )

        hidden_states = transformer_outputs[0]

        # Set device for model parallelism
        if self.model_parallel:
            torch.cuda.set_device(self.transformer.first_device)
            hidden_states = hidden_states.to(self.lm_head.weight.device)

        lm_logits = self.lm_head(hidden_states)
        mc_logits = self.multiple_choice_head(hidden_states, mc_token_ids).squeeze(-1)

        mc_loss = None
        if mc_labels is not None:
            loss_fct = CrossEntropyLoss()
            mc_loss = loss_fct(
                mc_logits.view(-1, mc_logits.size(-1)), mc_labels.view(-1)
            )
        lm_loss = None
        if labels is not None:
            shift_logits = lm_logits[..., :-1, :].contiguous()
            shift_labels = labels[..., 1:].contiguous()
            loss_fct = CrossEntropyLoss()
            lm_loss = loss_fct(
                shift_logits.view(-1, shift_logits.size(-1)), shift_labels.view(-1)
            )

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

        return GPT2DoubleHeadsModelOutput(
            loss=lm_loss,
            mc_loss=mc_loss,
            logits=lm_logits,
            mc_logits=mc_logits,
            past_key_values=transformer_outputs.past_key_values,
            hidden_states=transformer_outputs.hidden_states,
            attentions=transformer_outputs.attentions,
        )

    @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 :obj:`past_key_values` cache if
        :meth:`~transformers.PreTrainedModel.beam_search` or :meth:`~transformers.PreTrainedModel.beam_sample` is
        called. This is required to match :obj:`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
        )


@add_start_docstrings(
    """
    The GPT2 Model transformer with a sequence classification head on top (linear layer).

    :class:`~transformers.GPT2ForSequenceClassification` uses the last token in order to do the classification, as
    other causal models (e.g. GPT-1) do.

    Since it does classification on the last token, it requires to know the position of the last token. If a
    :obj:`pad_token_id` is defined in the configuration, it finds the last token that is not a padding token in each
    row. If no :obj:`pad_token_id` is defined, it simply takes the last value in each row of the batch. Since it cannot
    guess the padding tokens when :obj:`inputs_embeds` are passed instead of :obj:`input_ids`, it does the same (take
    the last value in each row of the batch).
    """,
    GPT2_START_DOCSTRING,
)
class GPT2ForSequenceClassification(GPT2PreTrainedModel):
    _keys_to_ignore_on_load_missing = [r"h\.\d+\.attn\.masked_bias", r"lm_head\.weight"]

    def __init__(self, config):
        super().__init__(config)
        self.num_labels = config.num_labels
        self.transformer = GPT2Model(config)
        self.score = nn.Linear(config.n_embd, self.num_labels, bias=False)

        self.init_weights()

        # Model parallel
        self.model_parallel = False
        self.device_map = None

    @add_start_docstrings_to_model_forward(GPT2_INPUTS_DOCSTRING)
    @add_code_sample_docstrings(
        tokenizer_class=_TOKENIZER_FOR_DOC,
        checkpoint="microsoft/dialogrpt",
        output_type=SequenceClassifierOutputWithPast,
        config_class=_CONFIG_FOR_DOC,
    )
    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,
    ):
        r"""
        labels (:obj:`torch.LongTensor` of shape :obj:`(batch_size,)`, `optional`):
            Labels for computing the sequence classification/regression loss. Indices should be in :obj:`[0, ...,
            config.num_labels - 1]`. If :obj:`config.num_labels == 1` a regression loss is computed (Mean-Square loss),
            If :obj:`config.num_labels > 1` a classification loss is computed (Cross-Entropy).
        """
        return_dict = (
            return_dict if return_dict is not None else self.config.use_return_dict
        )

        transformer_outputs = self.transformer(
            input_ids,
            past_key_values=past_key_values,
            attention_mask=attention_mask,
            token_type_ids=token_type_ids,
            position_ids=position_ids,
            head_mask=head_mask,
            inputs_embeds=inputs_embeds,
            use_cache=use_cache,
            output_attentions=output_attentions,
            output_hidden_states=output_hidden_states,
            return_dict=return_dict,
        )
        hidden_states = transformer_outputs[0]
        logits = self.score(hidden_states)

        if input_ids is not None:
            batch_size, sequence_length = input_ids.shape[:2]
        else:
            batch_size, sequence_length = inputs_embeds.shape[:2]

        assert (
            self.config.pad_token_id is not None or batch_size == 1
        ), "Cannot handle batch sizes > 1 if no padding token is defined."
        if self.config.pad_token_id is None:
            sequence_lengths = -1
        else:
            if input_ids is not None:
                sequence_lengths = (
                    torch.ne(input_ids, self.config.pad_token_id).sum(-1) - 1
                )
            else:
                sequence_lengths = -1
                logger.warning(
                    f"{self.__class__.__name__} will not detect padding tokens in `inputs_embeds`. Results may be "
                    f"unexpected if using padding tokens in conjunction with `inputs_embeds.`"
                )

        pooled_logits = logits[range(batch_size), sequence_lengths]

        loss = None
        if labels is not None:
            if self.num_labels == 1:
                #  We are doing regression
                loss_fct = MSELoss()
                loss = loss_fct(pooled_logits.view(-1), labels.to(self.dtype).view(-1))
            else:
                loss_fct = CrossEntropyLoss()
                loss = loss_fct(
                    pooled_logits.view(-1, self.num_labels), labels.view(-1)
                )

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

        return SequenceClassifierOutputWithPast(
            loss=loss,
            logits=pooled_logits,
            past_key_values=transformer_outputs.past_key_values,
            hidden_states=transformer_outputs.hidden_states,
            attentions=transformer_outputs.attentions,
        )


@add_start_docstrings(
    """
    GPT2 Model with a token classification head on top (a linear layer on top of the hidden-states output) e.g. for
    Named-Entity-Recognition (NER) tasks.
    """,
    GPT2_START_DOCSTRING,
)
class GPT2ForTokenClassification(GPT2PreTrainedModel):
    def __init__(self, config):
        super().__init__(config)
        self.num_labels = config.num_labels

        self.transformer = GPT2Model(config)
        if (
            hasattr(config, "classifier_dropout")
            and config.classifier_dropout is not None
        ):
            classifier_dropout = config.classifier_dropout
        elif hasattr(config, "hidden_dropout") and config.hidden_dropout is not None:
            classifier_dropout = config.hidden_dropout
        else:
            classifier_dropout = 0.1
        self.dropout = nn.Dropout(classifier_dropout)
        self.classifier = nn.Linear(config.hidden_size, config.num_labels)

        self.init_weights()

        # Model parallel
        self.model_parallel = False
        self.device_map = None

    @add_start_docstrings_to_model_forward(GPT2_INPUTS_DOCSTRING)
    @add_code_sample_docstrings(
        tokenizer_class=_TOKENIZER_FOR_DOC,
        checkpoint="microsoft/DialogRPT-updown",
        output_type=TokenClassifierOutput,
        config_class=_CONFIG_FOR_DOC,
    )
    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,
    ):
        r"""
        labels (:obj:`torch.LongTensor` of shape :obj:`(batch_size,)`, `optional`):
            Labels for computing the sequence classification/regression loss. Indices should be in :obj:`[0, ...,
            config.num_labels - 1]`. If :obj:`config.num_labels == 1` a regression loss is computed (Mean-Square loss),
            If :obj:`config.num_labels > 1` a classification loss is computed (Cross-Entropy).
        """
        return_dict = (
            return_dict if return_dict is not None else self.config.use_return_dict
        )

        transformer_outputs = self.transformer(
            input_ids,
            past_key_values=past_key_values,
            attention_mask=attention_mask,
            token_type_ids=token_type_ids,
            position_ids=position_ids,
            head_mask=head_mask,
            inputs_embeds=inputs_embeds,
            use_cache=use_cache,
            output_attentions=output_attentions,
            output_hidden_states=output_hidden_states,
            return_dict=return_dict,
        )

        hidden_states = transformer_outputs[0]
        hidden_states = self.dropout(hidden_states)
        logits = self.classifier(hidden_states)

        loss = None
        if labels is not None:
            loss_fct = CrossEntropyLoss()
            # Only keep active parts of the loss
            if attention_mask is not None:
                active_loss = attention_mask.view(-1) == 1
                active_logits = logits.view(-1, self.num_labels)
                active_labels = torch.where(
                    active_loss,
                    labels.view(-1),
                    torch.tensor(loss_fct.ignore_index).type_as(labels),
                )
                loss = loss_fct(active_logits, active_labels)
            else:
                loss = loss_fct(logits.view(-1, self.num_labels), labels.view(-1))

        if not return_dict:
            output = (logits,) + transformer_outputs[2:]
            return ((loss,) + output) if loss is not None else output

        return TokenClassifierOutput(
            loss=loss,
            logits=logits,
            hidden_states=transformer_outputs.hidden_states,
            attentions=transformer_outputs.attentions,
        )