# Copyright 2019 The Google AI Language Team Authors and The HuggingFace Inc. team.
#
# Licensed under the Apache License, Version 2.0 (the "License");
# you may not use this file except in compliance with the License.
# You may obtain a copy of the License at
#
#     http://www.apache.org/licenses/LICENSE-2.0
#
# Unless required by applicable law or agreed to in writing, software
# distributed under the License is distributed on an "AS IS" BASIS,
# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
# See the License for the specific language governing permissions and
# limitations under the License.
"""PyTorch ELECTRA model."""

from collections.abc import Callable
from dataclasses import dataclass

import torch
from torch import nn
from torch.nn import BCEWithLogitsLoss, CrossEntropyLoss, MSELoss

from ... import initialization as init
from ...activations import ACT2FN, get_activation
from ...cache_utils import Cache, DynamicCache, EncoderDecoderCache
from ...generation import GenerationMixin
from ...masking_utils import create_bidirectional_mask, create_causal_mask
from ...modeling_layers import GradientCheckpointingLayer
from ...modeling_outputs import (
    BaseModelOutputWithCrossAttentions,
    BaseModelOutputWithPastAndCrossAttentions,
    CausalLMOutputWithCrossAttentions,
    MaskedLMOutput,
    MultipleChoiceModelOutput,
    QuestionAnsweringModelOutput,
    SequenceClassifierOutput,
    TokenClassifierOutput,
)
from ...modeling_utils import ALL_ATTENTION_FUNCTIONS, PreTrainedModel
from ...processing_utils import Unpack
from ...pytorch_utils import apply_chunking_to_forward
from ...utils import (
    ModelOutput,
    TransformersKwargs,
    auto_docstring,
    logging,
)
from ...utils.generic import can_return_tuple, merge_with_config_defaults
from ...utils.output_capturing import capture_outputs
from .configuration_electra import ElectraConfig


logger = logging.get_logger(__name__)


class ElectraEmbeddings(nn.Module):
    """Construct the embeddings from word, position and token_type embeddings."""

    def __init__(self, config):
        super().__init__()
        self.word_embeddings = nn.Embedding(config.vocab_size, config.embedding_size, padding_idx=config.pad_token_id)
        self.position_embeddings = nn.Embedding(config.max_position_embeddings, config.embedding_size)
        self.token_type_embeddings = nn.Embedding(config.type_vocab_size, config.embedding_size)

        self.LayerNorm = nn.LayerNorm(config.embedding_size, eps=config.layer_norm_eps)
        self.dropout = nn.Dropout(config.hidden_dropout_prob)

        # position_ids (1, len position emb) is contiguous in memory and exported when serialized
        self.register_buffer(
            "position_ids", torch.arange(config.max_position_embeddings).expand((1, -1)), persistent=False
        )
        self.register_buffer(
            "token_type_ids", torch.zeros(self.position_ids.size(), dtype=torch.long), persistent=False
        )

    # Copied from transformers.models.bert.modeling_bert.BertEmbeddings.forward
    def forward(
        self,
        input_ids: torch.LongTensor | None = None,
        token_type_ids: torch.LongTensor | None = None,
        position_ids: torch.LongTensor | None = None,
        inputs_embeds: torch.FloatTensor | None = None,
        past_key_values_length: int = 0,
    ) -> torch.Tensor:
        if input_ids is not None:
            input_shape = input_ids.size()
        else:
            input_shape = inputs_embeds.size()[:-1]

        batch_size, seq_length = input_shape

        if position_ids is None:
            position_ids = self.position_ids[:, past_key_values_length : seq_length + past_key_values_length]

        # Setting the token_type_ids to the registered buffer in constructor where it is all zeros, which usually occurs
        # when its auto-generated, registered buffer helps users when tracing the model without passing token_type_ids, solves
        # issue #5664
        if token_type_ids is None:
            if hasattr(self, "token_type_ids"):
                # NOTE: We assume either pos ids to have bsz == 1 (broadcastable) or bsz == effective bsz (input_shape[0])
                buffered_token_type_ids = self.token_type_ids.expand(position_ids.shape[0], -1)
                buffered_token_type_ids = torch.gather(buffered_token_type_ids, dim=1, index=position_ids)
                token_type_ids = buffered_token_type_ids.expand(batch_size, seq_length)
            else:
                token_type_ids = torch.zeros(input_shape, dtype=torch.long, device=self.position_ids.device)

        if inputs_embeds is None:
            inputs_embeds = self.word_embeddings(input_ids)
        token_type_embeddings = self.token_type_embeddings(token_type_ids)
        embeddings = inputs_embeds + token_type_embeddings

        position_embeddings = self.position_embeddings(position_ids)
        embeddings = embeddings + position_embeddings

        embeddings = self.LayerNorm(embeddings)
        embeddings = self.dropout(embeddings)
        return embeddings


# Copied from transformers.models.bert.modeling_bert.eager_attention_forward
def eager_attention_forward(
    module: nn.Module,
    query: torch.Tensor,
    key: torch.Tensor,
    value: torch.Tensor,
    attention_mask: torch.Tensor | None,
    scaling: float | None = None,
    dropout: float = 0.0,
    **kwargs: Unpack[TransformersKwargs],
):
    if scaling is None:
        scaling = query.size(-1) ** -0.5

    # Take the dot product between "query" and "key" to get the raw attention scores.
    attn_weights = torch.matmul(query, key.transpose(2, 3)) * scaling

    if attention_mask is not None:
        attn_weights = attn_weights + attention_mask

    attn_weights = nn.functional.softmax(attn_weights, dim=-1)
    attn_weights = nn.functional.dropout(attn_weights, p=dropout, training=module.training)

    attn_output = torch.matmul(attn_weights, value)
    attn_output = attn_output.transpose(1, 2).contiguous()

    return attn_output, attn_weights


# Copied from transformers.models.bert.modeling_bert.BertSelfAttention with Bert->Electra
class ElectraSelfAttention(nn.Module):
    def __init__(self, config, is_causal=False, layer_idx=None):
        super().__init__()
        if config.hidden_size % config.num_attention_heads != 0 and not hasattr(config, "embedding_size"):
            raise ValueError(
                f"The hidden size ({config.hidden_size}) is not a multiple of the number of attention "
                f"heads ({config.num_attention_heads})"
            )
        self.config = config

        self.num_attention_heads = config.num_attention_heads
        self.attention_head_size = int(config.hidden_size / config.num_attention_heads)
        self.all_head_size = self.num_attention_heads * self.attention_head_size
        self.scaling = self.attention_head_size**-0.5

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

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

        self.is_decoder = config.is_decoder
        self.is_causal = is_causal
        self.layer_idx = layer_idx

    def forward(
        self,
        hidden_states: torch.Tensor,
        attention_mask: torch.FloatTensor | None = None,
        past_key_values: Cache | None = None,
        cache_position: torch.Tensor | None = None,
        **kwargs: Unpack[TransformersKwargs],
    ) -> tuple[torch.Tensor]:
        input_shape = hidden_states.shape[:-1]
        hidden_shape = (*input_shape, -1, self.attention_head_size)

        # get all proj
        query_layer = self.query(hidden_states).view(*hidden_shape).transpose(1, 2)
        key_layer = self.key(hidden_states).view(*hidden_shape).transpose(1, 2)
        value_layer = self.value(hidden_states).view(*hidden_shape).transpose(1, 2)

        if past_key_values is not None:
            # decoder-only bert can have a simple dynamic cache for example
            current_past_key_values = past_key_values
            if isinstance(past_key_values, EncoderDecoderCache):
                current_past_key_values = past_key_values.self_attention_cache

            # save all key/value_layer to cache to be re-used for fast auto-regressive generation
            key_layer, value_layer = current_past_key_values.update(
                key_layer,
                value_layer,
                self.layer_idx,
                {"cache_position": cache_position},
            )

        attention_interface: Callable = ALL_ATTENTION_FUNCTIONS.get_interface(
            self.config._attn_implementation, eager_attention_forward
        )

        attn_output, attn_weights = attention_interface(
            self,
            query_layer,
            key_layer,
            value_layer,
            attention_mask,
            dropout=0.0 if not self.training else self.dropout.p,
            scaling=self.scaling,
            **kwargs,
        )
        attn_output = attn_output.reshape(*input_shape, -1).contiguous()
        return attn_output, attn_weights


# Copied from transformers.models.bert.modeling_bert.BertCrossAttention with Bert->Electra
class ElectraCrossAttention(nn.Module):
    def __init__(self, config, is_causal=False, layer_idx=None):
        super().__init__()
        if config.hidden_size % config.num_attention_heads != 0 and not hasattr(config, "embedding_size"):
            raise ValueError(
                f"The hidden size ({config.hidden_size}) is not a multiple of the number of attention "
                f"heads ({config.num_attention_heads})"
            )
        self.config = config

        self.num_attention_heads = config.num_attention_heads
        self.attention_head_size = int(config.hidden_size / config.num_attention_heads)
        self.all_head_size = self.num_attention_heads * self.attention_head_size
        self.scaling = self.attention_head_size**-0.5

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

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

        self.is_causal = is_causal
        self.layer_idx = layer_idx

    def forward(
        self,
        hidden_states: torch.Tensor,
        encoder_hidden_states: torch.FloatTensor | None = None,
        attention_mask: torch.FloatTensor | None = None,
        past_key_values: EncoderDecoderCache | None = None,
        **kwargs: Unpack[TransformersKwargs],
    ) -> tuple[torch.Tensor]:
        # determine input shapes
        bsz, tgt_len = hidden_states.shape[:-1]
        src_len = encoder_hidden_states.shape[1]

        q_input_shape = (bsz, tgt_len, -1, self.attention_head_size)
        kv_input_shape = (bsz, src_len, -1, self.attention_head_size)

        # get query proj
        query_layer = self.query(hidden_states).view(*q_input_shape).transpose(1, 2)

        is_updated = past_key_values.is_updated.get(self.layer_idx) if past_key_values is not None else False
        if past_key_values is not None and is_updated:
            # reuse k,v, cross_attentions
            key_layer = past_key_values.cross_attention_cache.layers[self.layer_idx].keys
            value_layer = past_key_values.cross_attention_cache.layers[self.layer_idx].values
        else:
            key_layer = self.key(encoder_hidden_states).view(*kv_input_shape).transpose(1, 2)
            value_layer = self.value(encoder_hidden_states).view(*kv_input_shape).transpose(1, 2)

            if past_key_values is not None:
                # save all states to the cache
                key_layer, value_layer = past_key_values.cross_attention_cache.update(
                    key_layer, value_layer, self.layer_idx
                )
                # set flag that curr layer for cross-attn is already updated so we can re-use in subsequent calls
                past_key_values.is_updated[self.layer_idx] = True

        attention_interface: Callable = ALL_ATTENTION_FUNCTIONS.get_interface(
            self.config._attn_implementation, eager_attention_forward
        )

        attn_output, attn_weights = attention_interface(
            self,
            query_layer,
            key_layer,
            value_layer,
            attention_mask,
            dropout=0.0 if not self.training else self.dropout.p,
            scaling=self.scaling,
            **kwargs,
        )
        attn_output = attn_output.reshape(bsz, tgt_len, -1).contiguous()
        return attn_output, attn_weights


# Copied from transformers.models.bert.modeling_bert.BertSelfOutput
class ElectraSelfOutput(nn.Module):
    def __init__(self, config):
        super().__init__()
        self.dense = nn.Linear(config.hidden_size, config.hidden_size)
        self.LayerNorm = nn.LayerNorm(config.hidden_size, eps=config.layer_norm_eps)
        self.dropout = nn.Dropout(config.hidden_dropout_prob)

    def forward(self, hidden_states: torch.Tensor, input_tensor: torch.Tensor) -> torch.Tensor:
        hidden_states = self.dense(hidden_states)
        hidden_states = self.dropout(hidden_states)
        hidden_states = self.LayerNorm(hidden_states + input_tensor)
        return hidden_states


# Copied from transformers.models.bert.modeling_bert.BertAttention with Bert->Electra,BERT->ELECTRA
class ElectraAttention(nn.Module):
    def __init__(self, config, is_causal=False, layer_idx=None, is_cross_attention=False):
        super().__init__()
        self.is_cross_attention = is_cross_attention
        attention_class = ElectraCrossAttention if is_cross_attention else ElectraSelfAttention
        self.self = attention_class(config, is_causal=is_causal, layer_idx=layer_idx)
        self.output = ElectraSelfOutput(config)

    def forward(
        self,
        hidden_states: torch.Tensor,
        attention_mask: torch.FloatTensor | None = None,
        encoder_hidden_states: torch.FloatTensor | None = None,
        encoder_attention_mask: torch.FloatTensor | None = None,
        past_key_values: Cache | None = None,
        cache_position: torch.Tensor | None = None,
        **kwargs: Unpack[TransformersKwargs],
    ) -> tuple[torch.Tensor]:
        attention_mask = attention_mask if not self.is_cross_attention else encoder_attention_mask
        attention_output, attn_weights = self.self(
            hidden_states,
            encoder_hidden_states=encoder_hidden_states,
            attention_mask=attention_mask,
            past_key_values=past_key_values,
            cache_position=cache_position,
            **kwargs,
        )
        attention_output = self.output(attention_output, hidden_states)
        return attention_output, attn_weights


# Copied from transformers.models.bert.modeling_bert.BertIntermediate
class ElectraIntermediate(nn.Module):
    def __init__(self, config):
        super().__init__()
        self.dense = nn.Linear(config.hidden_size, config.intermediate_size)
        if isinstance(config.hidden_act, str):
            self.intermediate_act_fn = ACT2FN[config.hidden_act]
        else:
            self.intermediate_act_fn = config.hidden_act

    def forward(self, hidden_states: torch.Tensor) -> torch.Tensor:
        hidden_states = self.dense(hidden_states)
        hidden_states = self.intermediate_act_fn(hidden_states)
        return hidden_states


# Copied from transformers.models.bert.modeling_bert.BertOutput
class ElectraOutput(nn.Module):
    def __init__(self, config):
        super().__init__()
        self.dense = nn.Linear(config.intermediate_size, config.hidden_size)
        self.LayerNorm = nn.LayerNorm(config.hidden_size, eps=config.layer_norm_eps)
        self.dropout = nn.Dropout(config.hidden_dropout_prob)

    def forward(self, hidden_states: torch.Tensor, input_tensor: torch.Tensor) -> torch.Tensor:
        hidden_states = self.dense(hidden_states)
        hidden_states = self.dropout(hidden_states)
        hidden_states = self.LayerNorm(hidden_states + input_tensor)
        return hidden_states


# Copied from transformers.models.bert.modeling_bert.BertLayer with Bert->Electra
class ElectraLayer(GradientCheckpointingLayer):
    def __init__(self, config, layer_idx=None):
        super().__init__()
        self.chunk_size_feed_forward = config.chunk_size_feed_forward
        self.seq_len_dim = 1
        self.attention = ElectraAttention(config, is_causal=config.is_decoder, layer_idx=layer_idx)
        self.is_decoder = config.is_decoder
        self.add_cross_attention = config.add_cross_attention
        if self.add_cross_attention:
            if not self.is_decoder:
                raise ValueError(f"{self} should be used as a decoder model if cross attention is added")
            self.crossattention = ElectraAttention(
                config,
                is_causal=False,
                layer_idx=layer_idx,
                is_cross_attention=True,
            )
        self.intermediate = ElectraIntermediate(config)
        self.output = ElectraOutput(config)

    def forward(
        self,
        hidden_states: torch.Tensor,
        attention_mask: torch.FloatTensor | None = None,
        encoder_hidden_states: torch.FloatTensor | None = None,
        encoder_attention_mask: torch.FloatTensor | None = None,
        past_key_values: Cache | None = None,
        cache_position: torch.Tensor | None = None,
        **kwargs: Unpack[TransformersKwargs],
    ) -> tuple[torch.Tensor]:
        self_attention_output, _ = self.attention(
            hidden_states,
            attention_mask,
            past_key_values=past_key_values,
            cache_position=cache_position,
            **kwargs,
        )
        attention_output = self_attention_output

        if self.is_decoder and encoder_hidden_states is not None:
            if not hasattr(self, "crossattention"):
                raise ValueError(
                    f"If `encoder_hidden_states` are passed, {self} has to be instantiated with cross-attention layers"
                    " by setting `config.add_cross_attention=True`"
                )

            cross_attention_output, _ = self.crossattention(
                self_attention_output,
                None,  # attention_mask
                encoder_hidden_states,
                encoder_attention_mask,
                past_key_values=past_key_values,
                **kwargs,
            )
            attention_output = cross_attention_output

        layer_output = apply_chunking_to_forward(
            self.feed_forward_chunk, self.chunk_size_feed_forward, self.seq_len_dim, attention_output
        )
        return layer_output

    def feed_forward_chunk(self, attention_output):
        intermediate_output = self.intermediate(attention_output)
        layer_output = self.output(intermediate_output, attention_output)
        return layer_output


# Copied from transformers.models.bert.modeling_bert.BertEncoder with Bert->Electra
class ElectraEncoder(nn.Module):
    def __init__(self, config):
        super().__init__()
        self.config = config
        self.layer = nn.ModuleList([ElectraLayer(config, layer_idx=i) for i in range(config.num_hidden_layers)])

    def forward(
        self,
        hidden_states: torch.Tensor,
        attention_mask: torch.FloatTensor | None = None,
        encoder_hidden_states: torch.FloatTensor | None = None,
        encoder_attention_mask: torch.FloatTensor | None = None,
        past_key_values: Cache | None = None,
        use_cache: bool | None = None,
        cache_position: torch.Tensor | None = None,
        **kwargs: Unpack[TransformersKwargs],
    ) -> tuple[torch.Tensor] | BaseModelOutputWithPastAndCrossAttentions:
        for i, layer_module in enumerate(self.layer):
            hidden_states = layer_module(
                hidden_states,
                attention_mask,
                encoder_hidden_states,  # as a positional argument for gradient checkpointing
                encoder_attention_mask=encoder_attention_mask,
                past_key_values=past_key_values,
                cache_position=cache_position,
                **kwargs,
            )

        return BaseModelOutputWithPastAndCrossAttentions(
            last_hidden_state=hidden_states,
            past_key_values=past_key_values if use_cache else None,
        )


class ElectraDiscriminatorPredictions(nn.Module):
    """Prediction module for the discriminator, made up of two dense layers."""

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

        self.dense = nn.Linear(config.hidden_size, config.hidden_size)
        self.activation = get_activation(config.hidden_act)
        self.dense_prediction = nn.Linear(config.hidden_size, 1)
        self.config = config

    def forward(self, discriminator_hidden_states):
        hidden_states = self.dense(discriminator_hidden_states)
        hidden_states = self.activation(hidden_states)
        logits = self.dense_prediction(hidden_states).squeeze(-1)

        return logits


class ElectraGeneratorPredictions(nn.Module):
    """Prediction module for the generator, made up of two dense layers."""

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

        self.activation = get_activation("gelu")
        self.LayerNorm = nn.LayerNorm(config.embedding_size, eps=config.layer_norm_eps)
        self.dense = nn.Linear(config.hidden_size, config.embedding_size)

    def forward(self, generator_hidden_states):
        hidden_states = self.dense(generator_hidden_states)
        hidden_states = self.activation(hidden_states)
        hidden_states = self.LayerNorm(hidden_states)

        return hidden_states


@auto_docstring
class ElectraPreTrainedModel(PreTrainedModel):
    config_class = ElectraConfig
    base_model_prefix = "electra"
    supports_gradient_checkpointing = True
    _supports_flash_attn = True
    _supports_sdpa = True
    _supports_flex_attn = True
    _supports_attention_backend = True
    _can_record_outputs = {
        "hidden_states": ElectraLayer,
        "attentions": ElectraSelfAttention,
        "cross_attentions": ElectraCrossAttention,
    }

    def _init_weights(self, module):
        super()._init_weights(module)
        if isinstance(module, ElectraEmbeddings):
            init.copy_(module.position_ids, torch.arange(module.position_ids.shape[-1]).expand((1, -1)))
            init.zeros_(module.token_type_ids)


@dataclass
@auto_docstring(
    custom_intro="""
    Output type of [`ElectraForPreTraining`].
    """
)
class ElectraForPreTrainingOutput(ModelOutput):
    r"""
    loss (*optional*, returned when `labels` is provided, `torch.FloatTensor` of shape `(1,)`):
        Total loss of the ELECTRA objective.
    logits (`torch.FloatTensor` of shape `(batch_size, sequence_length)`):
        Prediction scores of the head (scores for each token before SoftMax).
    """

    loss: torch.FloatTensor | None = None
    logits: torch.FloatTensor | None = None
    hidden_states: tuple[torch.FloatTensor] | None = None
    attentions: tuple[torch.FloatTensor] | None = None


@auto_docstring
class ElectraModel(ElectraPreTrainedModel):
    def __init__(self, config):
        super().__init__(config)
        self.embeddings = ElectraEmbeddings(config)

        if config.embedding_size != config.hidden_size:
            self.embeddings_project = nn.Linear(config.embedding_size, config.hidden_size)

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

    def get_input_embeddings(self):
        return self.embeddings.word_embeddings

    def set_input_embeddings(self, value):
        self.embeddings.word_embeddings = value

    @merge_with_config_defaults
    @capture_outputs
    @auto_docstring
    def forward(
        self,
        input_ids: torch.Tensor | None = None,
        attention_mask: torch.Tensor | None = None,
        token_type_ids: torch.Tensor | None = None,
        position_ids: torch.Tensor | None = None,
        inputs_embeds: torch.Tensor | None = None,
        encoder_hidden_states: torch.Tensor | None = None,
        encoder_attention_mask: torch.Tensor | None = None,
        past_key_values: list[torch.FloatTensor] | None = None,
        use_cache: bool | None = None,
        cache_position: torch.Tensor | None = None,
        **kwargs: Unpack[TransformersKwargs],
    ) -> tuple[torch.Tensor] | BaseModelOutputWithCrossAttentions:
        if self.config.is_decoder:
            use_cache = use_cache if use_cache is not None else self.config.use_cache
        else:
            use_cache = False

        if use_cache and past_key_values is None:
            past_key_values = (
                EncoderDecoderCache(DynamicCache(config=self.config), DynamicCache(config=self.config))
                if encoder_hidden_states is not None or self.config.is_encoder_decoder
                else DynamicCache(config=self.config)
            )

        if (input_ids is None) ^ (inputs_embeds is not None):
            raise ValueError("You must specify exactly one of input_ids or inputs_embeds")

        if input_ids is not None:
            device = input_ids.device
            input_shape = input_ids.shape
        else:
            device = inputs_embeds.device
            input_shape = inputs_embeds.shape[:-1]

        seq_length = input_shape[1]
        past_key_values_length = past_key_values.get_seq_length() if past_key_values is not None else 0
        if cache_position is None:
            cache_position = torch.arange(past_key_values_length, past_key_values_length + seq_length, device=device)

        embedding_output = self.embeddings(
            input_ids=input_ids,
            position_ids=position_ids,
            token_type_ids=token_type_ids,
            inputs_embeds=inputs_embeds,
            past_key_values_length=past_key_values_length,
        )
        if hasattr(self, "embeddings_project"):
            embedding_output = self.embeddings_project(embedding_output)

        attention_mask, encoder_attention_mask = self._create_attention_masks(
            attention_mask=attention_mask,
            encoder_attention_mask=encoder_attention_mask,
            embedding_output=embedding_output,
            encoder_hidden_states=encoder_hidden_states,
            cache_position=cache_position,
            past_key_values=past_key_values,
        )

        encoder_outputs = self.encoder(
            embedding_output,
            attention_mask=attention_mask,
            encoder_hidden_states=encoder_hidden_states,
            encoder_attention_mask=encoder_attention_mask,
            past_key_values=past_key_values,
            use_cache=use_cache,
            position_ids=position_ids,
            **kwargs,
        )

        return BaseModelOutputWithPastAndCrossAttentions(
            last_hidden_state=encoder_outputs.last_hidden_state,
            past_key_values=encoder_outputs.past_key_values,
        )

    # Copied from transformers.models.bert.modeling_bert.BertModel._create_attention_masks
    def _create_attention_masks(
        self,
        attention_mask,
        encoder_attention_mask,
        embedding_output,
        encoder_hidden_states,
        cache_position,
        past_key_values,
    ):
        if self.config.is_decoder:
            attention_mask = create_causal_mask(
                config=self.config,
                inputs_embeds=embedding_output,
                attention_mask=attention_mask,
                cache_position=cache_position,
                past_key_values=past_key_values,
            )
        else:
            attention_mask = create_bidirectional_mask(
                config=self.config,
                inputs_embeds=embedding_output,
                attention_mask=attention_mask,
            )

        if encoder_attention_mask is not None:
            encoder_attention_mask = create_bidirectional_mask(
                config=self.config,
                inputs_embeds=embedding_output,
                attention_mask=encoder_attention_mask,
                encoder_hidden_states=encoder_hidden_states,
            )

        return attention_mask, encoder_attention_mask


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

    def __init__(self, config):
        super().__init__()
        self.dense = nn.Linear(config.hidden_size, config.hidden_size)
        classifier_dropout = (
            config.classifier_dropout if config.classifier_dropout is not None else config.hidden_dropout_prob
        )
        self.activation = get_activation("gelu")
        self.dropout = nn.Dropout(classifier_dropout)
        self.out_proj = nn.Linear(config.hidden_size, config.num_labels)

    def forward(self, features, **kwargs):
        x = features[:, 0, :]  # take <s> token (equiv. to [CLS])
        x = self.dropout(x)
        x = self.dense(x)
        x = self.activation(x)  # although BERT uses tanh here, it seems Electra authors used gelu here
        x = self.dropout(x)
        x = self.out_proj(x)
        return x


# Copied from transformers.models.xlm.modeling_xlm.XLMSequenceSummary with XLM->Electra
class ElectraSequenceSummary(nn.Module):
    r"""
    Compute a single vector summary of a sequence hidden states.

    Args:
        config ([`ElectraConfig`]):
            The config used by the model. Relevant arguments in the config class of the model are (refer to the actual
            config class of your model for the default values it uses):

            - **summary_type** (`str`) -- The method to use to make this summary. Accepted values are:

                - `"last"` -- Take the last token hidden state (like XLNet)
                - `"first"` -- Take the first token hidden state (like Bert)
                - `"mean"` -- Take the mean of all tokens hidden states
                - `"cls_index"` -- Supply a Tensor of classification token position (GPT/GPT-2)
                - `"attn"` -- Not implemented now, use multi-head attention

            - **summary_use_proj** (`bool`) -- Add a projection after the vector extraction.
            - **summary_proj_to_labels** (`bool`) -- If `True`, the projection outputs to `config.num_labels` classes
              (otherwise to `config.hidden_size`).
            - **summary_activation** (`Optional[str]`) -- Set to `"tanh"` to add a tanh activation to the output,
              another string or `None` will add no activation.
            - **summary_first_dropout** (`float`) -- Optional dropout probability before the projection and activation.
            - **summary_last_dropout** (`float`)-- Optional dropout probability after the projection and activation.
    """

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

        self.summary_type = getattr(config, "summary_type", "last")
        if self.summary_type == "attn":
            # We should use a standard multi-head attention module with absolute positional embedding for that.
            # Cf. https://github.com/zihangdai/xlnet/blob/master/modeling.py#L253-L276
            # We can probably just use the multi-head attention module of PyTorch >=1.1.0
            raise NotImplementedError

        self.summary = nn.Identity()
        if hasattr(config, "summary_use_proj") and config.summary_use_proj:
            if hasattr(config, "summary_proj_to_labels") and config.summary_proj_to_labels and config.num_labels > 0:
                num_classes = config.num_labels
            else:
                num_classes = config.hidden_size
            self.summary = nn.Linear(config.hidden_size, num_classes)

        activation_string = getattr(config, "summary_activation", None)
        self.activation: Callable = get_activation(activation_string) if activation_string else nn.Identity()

        self.first_dropout = nn.Identity()
        if hasattr(config, "summary_first_dropout") and config.summary_first_dropout > 0:
            self.first_dropout = nn.Dropout(config.summary_first_dropout)

        self.last_dropout = nn.Identity()
        if hasattr(config, "summary_last_dropout") and config.summary_last_dropout > 0:
            self.last_dropout = nn.Dropout(config.summary_last_dropout)

    def forward(
        self, hidden_states: torch.FloatTensor, cls_index: torch.LongTensor | None = None
    ) -> torch.FloatTensor:
        """
        Compute a single vector summary of a sequence hidden states.

        Args:
            hidden_states (`torch.FloatTensor` of shape `[batch_size, seq_len, hidden_size]`):
                The hidden states of the last layer.
            cls_index (`torch.LongTensor` of shape `[batch_size]` or `[batch_size, ...]` where ... are optional leading dimensions of `hidden_states`, *optional*):
                Used if `summary_type == "cls_index"` and takes the last token of the sequence as classification token.

        Returns:
            `torch.FloatTensor`: The summary of the sequence hidden states.
        """
        if self.summary_type == "last":
            output = hidden_states[:, -1]
        elif self.summary_type == "first":
            output = hidden_states[:, 0]
        elif self.summary_type == "mean":
            output = hidden_states.mean(dim=1)
        elif self.summary_type == "cls_index":
            if cls_index is None:
                cls_index = torch.full_like(
                    hidden_states[..., :1, :],
                    hidden_states.shape[-2] - 1,
                    dtype=torch.long,
                )
            else:
                cls_index = cls_index.unsqueeze(-1).unsqueeze(-1)
                cls_index = cls_index.expand((-1,) * (cls_index.dim() - 1) + (hidden_states.size(-1),))
            # shape of cls_index: (bsz, XX, 1, hidden_size) where XX are optional leading dim of hidden_states
            output = hidden_states.gather(-2, cls_index).squeeze(-2)  # shape (bsz, XX, hidden_size)
        elif self.summary_type == "attn":
            raise NotImplementedError

        output = self.first_dropout(output)
        output = self.summary(output)
        output = self.activation(output)
        output = self.last_dropout(output)

        return output


@auto_docstring(
    custom_intro="""
    ELECTRA Model transformer with a sequence classification/regression head on top (a linear layer on top of the
    pooled output) e.g. for GLUE tasks.
    """
)
class ElectraForSequenceClassification(ElectraPreTrainedModel):
    def __init__(self, config):
        super().__init__(config)
        self.num_labels = config.num_labels
        self.config = config
        self.electra = ElectraModel(config)
        self.classifier = ElectraClassificationHead(config)

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

    @can_return_tuple
    @auto_docstring
    def forward(
        self,
        input_ids: torch.Tensor | None = None,
        attention_mask: torch.Tensor | None = None,
        token_type_ids: torch.Tensor | None = None,
        position_ids: torch.Tensor | None = None,
        inputs_embeds: torch.Tensor | None = None,
        labels: torch.Tensor | None = None,
        **kwargs: Unpack[TransformersKwargs],
    ) -> tuple[torch.Tensor] | SequenceClassifierOutput:
        r"""
        labels (`torch.LongTensor` of shape `(batch_size,)`, *optional*):
            Labels for computing the sequence classification/regression loss. Indices should be in `[0, ...,
            config.num_labels - 1]`. If `config.num_labels == 1` a regression loss is computed (Mean-Square loss), If
            `config.num_labels > 1` a classification loss is computed (Cross-Entropy).
        """
        discriminator_hidden_states = self.electra(
            input_ids,
            attention_mask=attention_mask,
            token_type_ids=token_type_ids,
            position_ids=position_ids,
            inputs_embeds=inputs_embeds,
            return_dict=True,
            **kwargs,
        )

        sequence_output = discriminator_hidden_states[0]
        logits = self.classifier(sequence_output)

        loss = None
        if labels is not None:
            if self.config.problem_type is None:
                if self.num_labels == 1:
                    self.config.problem_type = "regression"
                elif self.num_labels > 1 and (labels.dtype == torch.long or labels.dtype == torch.int):
                    self.config.problem_type = "single_label_classification"
                else:
                    self.config.problem_type = "multi_label_classification"

            if self.config.problem_type == "regression":
                loss_fct = MSELoss()
                if self.num_labels == 1:
                    loss = loss_fct(logits.squeeze(), labels.squeeze())
                else:
                    loss = loss_fct(logits, labels)
            elif self.config.problem_type == "single_label_classification":
                loss_fct = CrossEntropyLoss()
                loss = loss_fct(logits.view(-1, self.num_labels), labels.view(-1))
            elif self.config.problem_type == "multi_label_classification":
                loss_fct = BCEWithLogitsLoss()
                loss = loss_fct(logits, labels)

        return SequenceClassifierOutput(
            loss=loss,
            logits=logits,
            hidden_states=discriminator_hidden_states.hidden_states,
            attentions=discriminator_hidden_states.attentions,
        )


@auto_docstring(
    custom_intro="""
    Electra model with a binary classification head on top as used during pretraining for identifying generated tokens.

    It is recommended to load the discriminator checkpoint into that model.
    """
)
class ElectraForPreTraining(ElectraPreTrainedModel):
    def __init__(self, config):
        super().__init__(config)

        self.electra = ElectraModel(config)
        self.discriminator_predictions = ElectraDiscriminatorPredictions(config)
        # Initialize weights and apply final processing
        self.post_init()

    @can_return_tuple
    @auto_docstring
    def forward(
        self,
        input_ids: torch.Tensor | None = None,
        attention_mask: torch.Tensor | None = None,
        token_type_ids: torch.Tensor | None = None,
        position_ids: torch.Tensor | None = None,
        inputs_embeds: torch.Tensor | None = None,
        labels: torch.Tensor | None = None,
        **kwargs: Unpack[TransformersKwargs],
    ) -> tuple[torch.Tensor] | ElectraForPreTrainingOutput:
        r"""
        labels (`torch.LongTensor` of shape `(batch_size, sequence_length)`, *optional*):
            Labels for computing the ELECTRA loss. Input should be a sequence of tokens (see `input_ids` docstring)
            Indices should be in `[0, 1]`:

            - 0 indicates the token is an original token,
            - 1 indicates the token was replaced.

        Examples:

        ```python
        >>> from transformers import ElectraForPreTraining, AutoTokenizer
        >>> import torch

        >>> discriminator = ElectraForPreTraining.from_pretrained("google/electra-base-discriminator")
        >>> tokenizer = AutoTokenizer.from_pretrained("google/electra-base-discriminator")

        >>> sentence = "The quick brown fox jumps over the lazy dog"
        >>> fake_sentence = "The quick brown fox fake over the lazy dog"

        >>> fake_tokens = tokenizer.tokenize(fake_sentence, add_special_tokens=True)
        >>> fake_inputs = tokenizer.encode(fake_sentence, return_tensors="pt")
        >>> discriminator_outputs = discriminator(fake_inputs)
        >>> predictions = torch.round((torch.sign(discriminator_outputs[0]) + 1) / 2)

        >>> fake_tokens
        ['[CLS]', 'the', 'quick', 'brown', 'fox', 'fake', 'over', 'the', 'lazy', 'dog', '[SEP]']

        >>> predictions.squeeze().tolist()
        [0.0, 0.0, 0.0, 0.0, 0.0, 1.0, 0.0, 0.0, 0.0, 0.0, 0.0]
        ```"""
        discriminator_hidden_states = self.electra(
            input_ids,
            attention_mask=attention_mask,
            token_type_ids=token_type_ids,
            position_ids=position_ids,
            inputs_embeds=inputs_embeds,
            return_dict=True,
            **kwargs,
        )
        discriminator_sequence_output = discriminator_hidden_states[0]

        logits = self.discriminator_predictions(discriminator_sequence_output)

        loss = None
        if labels is not None:
            loss_fct = nn.BCEWithLogitsLoss()
            if attention_mask is not None:
                active_loss = attention_mask.view(-1, discriminator_sequence_output.shape[1]) == 1
                active_logits = logits.view(-1, discriminator_sequence_output.shape[1])[active_loss]
                active_labels = labels[active_loss]
                loss = loss_fct(active_logits, active_labels.float())
            else:
                loss = loss_fct(logits.view(-1, discriminator_sequence_output.shape[1]), labels.float())

        return ElectraForPreTrainingOutput(
            loss=loss,
            logits=logits,
            hidden_states=discriminator_hidden_states.hidden_states,
            attentions=discriminator_hidden_states.attentions,
        )


@auto_docstring(
    custom_intro="""
    Electra model with a language modeling head on top.

    Even though both the discriminator and generator may be loaded into this model, the generator is the only model of
    the two to have been trained for the masked language modeling task.
    """
)
class ElectraForMaskedLM(ElectraPreTrainedModel):
    _tied_weights_keys = {"generator_lm_head.weight": "electra.embeddings.word_embeddings.weight"}

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

        self.electra = ElectraModel(config)
        self.generator_predictions = ElectraGeneratorPredictions(config)

        self.generator_lm_head = nn.Linear(config.embedding_size, config.vocab_size)
        # Initialize weights and apply final processing
        self.post_init()

    def get_output_embeddings(self):
        return self.generator_lm_head

    def set_output_embeddings(self, word_embeddings):
        self.generator_lm_head = word_embeddings

    @can_return_tuple
    @auto_docstring
    def forward(
        self,
        input_ids: torch.Tensor | None = None,
        attention_mask: torch.Tensor | None = None,
        token_type_ids: torch.Tensor | None = None,
        position_ids: torch.Tensor | None = None,
        inputs_embeds: torch.Tensor | None = None,
        labels: torch.Tensor | None = None,
        **kwargs: Unpack[TransformersKwargs],
    ) -> tuple[torch.Tensor] | MaskedLMOutput:
        r"""
        labels (`torch.LongTensor` of shape `(batch_size, sequence_length)`, *optional*):
            Labels for computing the masked language modeling loss. Indices should be in `[-100, 0, ...,
            config.vocab_size]` (see `input_ids` docstring) Tokens with indices set to `-100` are ignored (masked), the
            loss is only computed for the tokens with labels in `[0, ..., config.vocab_size]`
        """
        generator_hidden_states = self.electra(
            input_ids,
            attention_mask=attention_mask,
            token_type_ids=token_type_ids,
            position_ids=position_ids,
            inputs_embeds=inputs_embeds,
            return_dict=True,
            **kwargs,
        )
        generator_sequence_output = generator_hidden_states[0]

        prediction_scores = self.generator_predictions(generator_sequence_output)
        prediction_scores = self.generator_lm_head(prediction_scores)

        loss = None
        # Masked language modeling softmax layer
        if labels is not None:
            loss_fct = nn.CrossEntropyLoss()  # -100 index = padding token
            loss = loss_fct(prediction_scores.view(-1, self.config.vocab_size), labels.view(-1))

        return MaskedLMOutput(
            loss=loss,
            logits=prediction_scores,
            hidden_states=generator_hidden_states.hidden_states,
            attentions=generator_hidden_states.attentions,
        )


@auto_docstring(
    custom_intro="""
    Electra model with a token classification head on top.

    Both the discriminator and generator may be loaded into this model.
    """
)
class ElectraForTokenClassification(ElectraPreTrainedModel):
    def __init__(self, config):
        super().__init__(config)
        self.num_labels = config.num_labels

        self.electra = ElectraModel(config)
        classifier_dropout = (
            config.classifier_dropout if config.classifier_dropout is not None else config.hidden_dropout_prob
        )
        self.dropout = nn.Dropout(classifier_dropout)
        self.classifier = nn.Linear(config.hidden_size, config.num_labels)
        # Initialize weights and apply final processing
        self.post_init()

    @can_return_tuple
    @auto_docstring
    def forward(
        self,
        input_ids: torch.Tensor | None = None,
        attention_mask: torch.Tensor | None = None,
        token_type_ids: torch.Tensor | None = None,
        position_ids: torch.Tensor | None = None,
        inputs_embeds: torch.Tensor | None = None,
        labels: torch.Tensor | None = None,
        **kwargs: Unpack[TransformersKwargs],
    ) -> tuple[torch.Tensor] | TokenClassifierOutput:
        r"""
        labels (`torch.LongTensor` of shape `(batch_size, sequence_length)`, *optional*):
            Labels for computing the token classification loss. Indices should be in `[0, ..., config.num_labels - 1]`.
        """
        discriminator_hidden_states = self.electra(
            input_ids,
            attention_mask=attention_mask,
            token_type_ids=token_type_ids,
            position_ids=position_ids,
            inputs_embeds=inputs_embeds,
            return_dict=True,
            **kwargs,
        )
        discriminator_sequence_output = discriminator_hidden_states[0]

        discriminator_sequence_output = self.dropout(discriminator_sequence_output)
        logits = self.classifier(discriminator_sequence_output)

        loss = None
        if labels is not None:
            loss_fct = CrossEntropyLoss()
            loss = loss_fct(logits.view(-1, self.num_labels), labels.view(-1))

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


@auto_docstring
class ElectraForQuestionAnswering(ElectraPreTrainedModel):
    config_class = ElectraConfig
    base_model_prefix = "electra"

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

        self.electra = ElectraModel(config)
        self.qa_outputs = nn.Linear(config.hidden_size, config.num_labels)

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

    @can_return_tuple
    @auto_docstring
    def forward(
        self,
        input_ids: torch.Tensor | None = None,
        attention_mask: torch.Tensor | None = None,
        token_type_ids: torch.Tensor | None = None,
        position_ids: torch.Tensor | None = None,
        inputs_embeds: torch.Tensor | None = None,
        start_positions: torch.Tensor | None = None,
        end_positions: torch.Tensor | None = None,
        **kwargs: Unpack[TransformersKwargs],
    ) -> tuple[torch.Tensor] | QuestionAnsweringModelOutput:
        discriminator_hidden_states = self.electra(
            input_ids,
            attention_mask=attention_mask,
            token_type_ids=token_type_ids,
            position_ids=position_ids,
            inputs_embeds=inputs_embeds,
            return_dict=True,
            **kwargs,
        )

        sequence_output = discriminator_hidden_states[0]

        logits = self.qa_outputs(sequence_output)
        start_logits, end_logits = logits.split(1, dim=-1)
        start_logits = start_logits.squeeze(-1).contiguous()
        end_logits = end_logits.squeeze(-1).contiguous()

        total_loss = None
        if start_positions is not None and end_positions is not None:
            # If we are on multi-GPU, split add a dimension
            if len(start_positions.size()) > 1:
                start_positions = start_positions.squeeze(-1)
            if len(end_positions.size()) > 1:
                end_positions = end_positions.squeeze(-1)
            # sometimes the start/end positions are outside our model inputs, we ignore these terms
            ignored_index = start_logits.size(1)
            start_positions = start_positions.clamp(0, ignored_index)
            end_positions = end_positions.clamp(0, ignored_index)

            loss_fct = CrossEntropyLoss(ignore_index=ignored_index)
            start_loss = loss_fct(start_logits, start_positions)
            end_loss = loss_fct(end_logits, end_positions)
            total_loss = (start_loss + end_loss) / 2

        return QuestionAnsweringModelOutput(
            loss=total_loss,
            start_logits=start_logits,
            end_logits=end_logits,
            hidden_states=discriminator_hidden_states.hidden_states,
            attentions=discriminator_hidden_states.attentions,
        )


@auto_docstring
class ElectraForMultipleChoice(ElectraPreTrainedModel):
    def __init__(self, config):
        super().__init__(config)

        self.electra = ElectraModel(config)
        self.sequence_summary = ElectraSequenceSummary(config)
        self.classifier = nn.Linear(config.hidden_size, 1)

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

    @can_return_tuple
    @auto_docstring
    def forward(
        self,
        input_ids: torch.Tensor | None = None,
        attention_mask: torch.Tensor | None = None,
        token_type_ids: torch.Tensor | None = None,
        position_ids: torch.Tensor | None = None,
        inputs_embeds: torch.Tensor | None = None,
        labels: torch.Tensor | None = None,
        **kwargs: Unpack[TransformersKwargs],
    ) -> tuple[torch.Tensor] | MultipleChoiceModelOutput:
        r"""
        input_ids (`torch.LongTensor` of shape `(batch_size, num_choices, sequence_length)`):
            Indices of input sequence tokens in the vocabulary.

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

            [What are input IDs?](../glossary#input-ids)
        token_type_ids (`torch.LongTensor` of shape `(batch_size, num_choices, sequence_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#token-type-ids)
        position_ids (`torch.LongTensor` of shape `(batch_size, num_choices, 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#position-ids)
        inputs_embeds (`torch.FloatTensor` of shape `(batch_size, num_choices, sequence_length, hidden_size)`, *optional*):
            Optionally, instead of passing `input_ids` you can choose to directly pass an embedded representation. This
            is useful if you want more control over how to convert `input_ids` indices into associated vectors than the
            model's internal embedding lookup matrix.
        labels (`torch.LongTensor` of shape `(batch_size,)`, *optional*):
            Labels for computing the multiple choice classification loss. Indices should be in `[0, ...,
            num_choices-1]` where `num_choices` is the size of the second dimension of the input tensors. (See
            `input_ids` above)
        """
        num_choices = input_ids.shape[1] if input_ids is not None else inputs_embeds.shape[1]

        input_ids = input_ids.view(-1, input_ids.size(-1)) if input_ids is not None else None
        attention_mask = attention_mask.view(-1, attention_mask.size(-1)) if attention_mask is not None else None
        token_type_ids = token_type_ids.view(-1, token_type_ids.size(-1)) if token_type_ids is not None else None
        position_ids = position_ids.view(-1, position_ids.size(-1)) if position_ids is not None else None
        inputs_embeds = (
            inputs_embeds.view(-1, inputs_embeds.size(-2), inputs_embeds.size(-1))
            if inputs_embeds is not None
            else None
        )

        discriminator_hidden_states = self.electra(
            input_ids,
            attention_mask=attention_mask,
            token_type_ids=token_type_ids,
            position_ids=position_ids,
            inputs_embeds=inputs_embeds,
            return_dict=True,
            **kwargs,
        )

        sequence_output = discriminator_hidden_states[0]

        pooled_output = self.sequence_summary(sequence_output)
        logits = self.classifier(pooled_output)
        reshaped_logits = logits.view(-1, num_choices)

        loss = None
        if labels is not None:
            loss_fct = CrossEntropyLoss()
            loss = loss_fct(reshaped_logits, labels)

        return MultipleChoiceModelOutput(
            loss=loss,
            logits=reshaped_logits,
            hidden_states=discriminator_hidden_states.hidden_states,
            attentions=discriminator_hidden_states.attentions,
        )


@auto_docstring(
    custom_intro="""
    ELECTRA Model with a `language modeling` head on top for CLM fine-tuning.
    """
)
class ElectraForCausalLM(ElectraPreTrainedModel, GenerationMixin):
    _tied_weights_keys = {"generator_lm_head.weight": "electra.embeddings.word_embeddings.weight"}

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

        if not config.is_decoder:
            logger.warning("If you want to use `ElectraForCausalLM` as a standalone, add `is_decoder=True.`")

        self.electra = ElectraModel(config)
        self.generator_predictions = ElectraGeneratorPredictions(config)
        self.generator_lm_head = nn.Linear(config.embedding_size, config.vocab_size)

        self.post_init()

    def get_output_embeddings(self):
        return self.generator_lm_head

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

    @can_return_tuple
    @auto_docstring
    def forward(
        self,
        input_ids: torch.Tensor | None = None,
        attention_mask: torch.Tensor | None = None,
        token_type_ids: torch.Tensor | None = None,
        position_ids: torch.Tensor | None = None,
        inputs_embeds: torch.Tensor | None = None,
        encoder_hidden_states: torch.Tensor | None = None,
        encoder_attention_mask: torch.Tensor | None = None,
        labels: torch.Tensor | None = None,
        past_key_values: Cache | None = None,
        use_cache: bool | None = None,
        cache_position: torch.Tensor | None = None,
        logits_to_keep: int | torch.Tensor = 0,
        **kwargs: Unpack[TransformersKwargs],
    ) -> tuple[torch.Tensor] | CausalLMOutputWithCrossAttentions:
        r"""
        labels (`torch.LongTensor` of shape `(batch_size, sequence_length)`, *optional*):
            Labels for computing the left-to-right language modeling loss (next word prediction). Indices should be in
            `[-100, 0, ..., config.vocab_size]` (see `input_ids` docstring) Tokens with indices set to `-100` are
            ignored (masked), the loss is only computed for the tokens with labels in `[0, ..., config.vocab_size]`

        Example:

        ```python
        >>> from transformers import AutoTokenizer, ElectraForCausalLM, ElectraConfig
        >>> import torch

        >>> tokenizer = AutoTokenizer.from_pretrained("google/electra-base-generator")
        >>> config = ElectraConfig.from_pretrained("google/electra-base-generator")
        >>> config.is_decoder = True
        >>> model = ElectraForCausalLM.from_pretrained("google/electra-base-generator", config=config)

        >>> inputs = tokenizer("Hello, my dog is cute", return_tensors="pt")
        >>> outputs = model(**inputs)

        >>> prediction_logits = outputs.logits
        ```"""
        if labels is not None:
            use_cache = False

        outputs: BaseModelOutputWithPastAndCrossAttentions = self.electra(
            input_ids,
            attention_mask=attention_mask,
            token_type_ids=token_type_ids,
            position_ids=position_ids,
            inputs_embeds=inputs_embeds,
            encoder_hidden_states=encoder_hidden_states,
            encoder_attention_mask=encoder_attention_mask,
            past_key_values=past_key_values,
            use_cache=use_cache,
            cache_position=cache_position,
            return_dict=True,
            **kwargs,
        )

        hidden_states = outputs.last_hidden_state
        # Only compute necessary logits, and do not upcast them to float if we are not computing the loss
        slice_indices = slice(-logits_to_keep, None) if isinstance(logits_to_keep, int) else logits_to_keep
        logits = self.generator_lm_head(self.generator_predictions(hidden_states[:, slice_indices, :]))

        loss = None
        if labels is not None:
            loss = self.loss_function(logits=logits, labels=labels, vocab_size=self.config.vocab_size, **kwargs)

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


__all__ = [
    "ElectraForCausalLM",
    "ElectraForMaskedLM",
    "ElectraForMultipleChoice",
    "ElectraForPreTraining",
    "ElectraForQuestionAnswering",
    "ElectraForSequenceClassification",
    "ElectraForTokenClassification",
    "ElectraModel",
    "ElectraPreTrainedModel",
]