image-match/python/py/Lib/site-packages/transformers/models/xmod/modeling_xmod.py

# Copyright 2023 Meta AI Team 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 X-MOD model."""

from collections.abc import Callable

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

from ... import initialization as init
from ...activations import ACT2FN, gelu
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 (
    BaseModelOutputWithPastAndCrossAttentions,
    BaseModelOutputWithPoolingAndCrossAttentions,
    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 TransformersKwargs, auto_docstring, logging
from ...utils.generic import can_return_tuple, merge_with_config_defaults
from ...utils.output_capturing import capture_outputs
from .configuration_xmod import XmodConfig


logger = logging.get_logger(__name__)


# Copied from transformers.models.roberta.modeling_roberta.RobertaEmbeddings with Roberta->Xmod
class XmodEmbeddings(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.hidden_size, padding_idx=config.pad_token_id)
        self.token_type_embeddings = nn.Embedding(config.type_vocab_size, config.hidden_size)

        self.LayerNorm = nn.LayerNorm(config.hidden_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
        )

        self.padding_idx = config.pad_token_id
        self.position_embeddings = nn.Embedding(
            config.max_position_embeddings, config.hidden_size, padding_idx=self.padding_idx
        )

    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 position_ids is None:
            if input_ids is not None:
                # Create the position ids from the input token ids. Any padded tokens remain padded.
                position_ids = self.create_position_ids_from_input_ids(
                    input_ids, self.padding_idx, past_key_values_length
                )
            else:
                position_ids = self.create_position_ids_from_inputs_embeds(inputs_embeds, self.padding_idx)

        if input_ids is not None:
            input_shape = input_ids.size()
        else:
            input_shape = inputs_embeds.size()[:-1]

        batch_size, seq_length = input_shape

        # 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

    @staticmethod
    def create_position_ids_from_inputs_embeds(inputs_embeds, padding_idx):
        """
        We are provided embeddings directly. We cannot infer which are padded so just generate sequential position ids.

        Args:
            inputs_embeds: torch.Tensor

        Returns: torch.Tensor
        """
        input_shape = inputs_embeds.size()[:-1]
        sequence_length = input_shape[1]

        position_ids = torch.arange(
            padding_idx + 1, sequence_length + padding_idx + 1, dtype=torch.long, device=inputs_embeds.device
        )
        return position_ids.unsqueeze(0).expand(input_shape)

    @staticmethod
    def create_position_ids_from_input_ids(input_ids, padding_idx, past_key_values_length=0):
        """
        Replace non-padding symbols with their position numbers. Position numbers begin at padding_idx+1. Padding symbols
        are ignored. This is modified from fairseq's `utils.make_positions`.

        Args:
            x: torch.Tensor x:

        Returns: torch.Tensor
        """
        # The series of casts and type-conversions here are carefully balanced to both work with ONNX export and XLA.
        mask = input_ids.ne(padding_idx).int()
        incremental_indices = (torch.cumsum(mask, dim=1).type_as(mask) + past_key_values_length) * mask
        return incremental_indices.long() + padding_idx


# 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.roberta.modeling_roberta.RobertaSelfAttention with Roberta->Xmod
class XmodSelfAttention(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,
        **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 roberta 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)

        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->Xmod
class XmodCrossAttention(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
        input_shape = hidden_states.shape[:-1]

        hidden_shape = (*input_shape, -1, self.attention_head_size)

        # get query proj
        query_layer = self.query(hidden_states).view(hidden_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:
            kv_shape = (*encoder_hidden_states.shape[:-1], -1, self.attention_head_size)
            key_layer = self.key(encoder_hidden_states).view(kv_shape).transpose(1, 2)
            value_layer = self.value(encoder_hidden_states).view(kv_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(*input_shape, -1).contiguous()
        return attn_output, attn_weights


class XmodSelfOutput(nn.Module):
    # Copied from transformers.models.roberta.modeling_roberta.RobertaSelfOutput.__init__
    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 = hidden_states + input_tensor
        return hidden_states


class XmodAttention(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 = XmodCrossAttention if is_cross_attention else XmodSelfAttention
        self.self = attention_class(config, is_causal=is_causal, layer_idx=layer_idx)
        self.output = XmodSelfOutput(config)

        self.pre_norm = config.pre_norm

    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: tuple[tuple[torch.FloatTensor]] | None = None,
        **kwargs: Unpack[TransformersKwargs],
    ) -> tuple[torch.Tensor]:
        residual = hidden_states
        if self.pre_norm:
            hidden_states = self.output.LayerNorm(hidden_states)

        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,
            **kwargs,
        )
        attention_output = self.output(attention_output, residual)

        if not self.pre_norm:
            attention_output = self.output.LayerNorm(attention_output)

        return attention_output, attn_weights


# Copied from transformers.models.roberta.modeling_roberta.RobertaIntermediate
class XmodIntermediate(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


class XmodAdapter(nn.Module):
    def __init__(self, config):
        super().__init__()
        self.bottleneck_size = config.hidden_size // config.adapter_reduction_factor
        self.dense1 = nn.Linear(config.hidden_size, self.bottleneck_size)
        self.dense2 = nn.Linear(self.bottleneck_size, config.hidden_size)
        if isinstance(config.hidden_act, str):
            self.adapter_act_fn = ACT2FN[config.hidden_act]
        else:
            self.adapter_act_fn = config.hidden_act

    def forward(self, hidden_states: torch.Tensor) -> torch.Tensor:
        hidden_states = self.dense1(hidden_states)
        hidden_states = self.adapter_act_fn(hidden_states)
        hidden_states = self.dense2(hidden_states)
        return hidden_states


class XmodOutput(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.ln_before_adapter = config.ln_before_adapter
        self.dropout = nn.Dropout(config.hidden_dropout_prob)
        if config.adapter_layer_norm:
            self.adapter_layer_norm = nn.LayerNorm(config.hidden_size, eps=config.layer_norm_eps)
        else:
            self.adapter_layer_norm = None
        self.adapter_reuse_layer_norm = config.adapter_reuse_layer_norm
        self.adapter_modules = nn.ModuleDict({})
        for language in config.languages:
            self.adapter_modules[str(language)] = XmodAdapter(config)

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

    def lang_adapter(self, lang_ids: torch.Tensor, hidden_states: torch.Tensor):
        if not self.ln_before_adapter:
            residual = hidden_states

        if self.adapter_layer_norm is not None:
            hidden_states = self.adapter_layer_norm(hidden_states)
        elif self.adapter_reuse_layer_norm:
            hidden_states = self.LayerNorm(hidden_states)

        if self.ln_before_adapter:
            residual = hidden_states

        new_hidden_states = torch.zeros_like(hidden_states)
        for adapter_idx, lang_key in enumerate(self.adapter_modules.keys()):
            lang_mask = lang_ids == adapter_idx
            lang_hidden_states = hidden_states[lang_mask]
            adapted_lang_hidden_states = self.adapter_modules[lang_key](lang_hidden_states)
            new_hidden_states[lang_mask] = adapted_lang_hidden_states

        hidden_states = self.dropout(new_hidden_states)
        hidden_states += residual
        return hidden_states


class XmodLayer(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 = XmodAttention(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 = XmodAttention(
                config,
                is_causal=False,
                layer_idx=layer_idx,
                is_cross_attention=True,
            )
        self.intermediate = XmodIntermediate(config)
        self.output = XmodOutput(config)
        self.pre_norm = config.pre_norm

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

        residual = attention_output
        if self.pre_norm:
            attention_output = self.output.LayerNorm(attention_output)
        intermediate_output = apply_chunking_to_forward(
            self.feed_forward_chunk,
            self.chunk_size_feed_forward,
            self.seq_len_dim,
            attention_output,
        )
        layer_output = self.output(intermediate_output, residual, lang_ids)
        if not self.pre_norm:
            layer_output = self.output.LayerNorm(layer_output)

        return layer_output

    def feed_forward_chunk(self, attention_output):
        return self.intermediate(attention_output)


class XmodEncoder(nn.Module):
    def __init__(self, config):
        super().__init__()
        self.config = config
        self.layer = nn.ModuleList([XmodLayer(config, layer_idx=i) for i in range(config.num_hidden_layers)])
        self.is_pre_norm = config.pre_norm
        if self.is_pre_norm:
            self.LayerNorm = nn.LayerNorm(config.hidden_size, eps=config.layer_norm_eps)

    def forward(
        self,
        hidden_states: torch.Tensor,
        lang_ids: 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: tuple[tuple[torch.FloatTensor]] | None = None,
        use_cache: bool | None = None,
        **kwargs: Unpack[TransformersKwargs],
    ) -> tuple[torch.Tensor] | BaseModelOutputWithPastAndCrossAttentions:
        for i, layer_module in enumerate(self.layer):
            hidden_states = layer_module(
                hidden_states,
                lang_ids,
                attention_mask,
                encoder_hidden_states,
                encoder_attention_mask,
                past_key_values,
                **kwargs,
            )

        if self.is_pre_norm:
            hidden_states = self.LayerNorm(hidden_states)

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


# Copied from transformers.models.roberta.modeling_roberta.RobertaPooler
class XmodPooler(nn.Module):
    def __init__(self, config):
        super().__init__()
        self.dense = nn.Linear(config.hidden_size, config.hidden_size)
        self.activation = nn.Tanh()

    def forward(self, hidden_states: torch.Tensor) -> torch.Tensor:
        # We "pool" the model by simply taking the hidden state corresponding
        # to the first token.
        first_token_tensor = hidden_states[:, 0]
        pooled_output = self.dense(first_token_tensor)
        pooled_output = self.activation(pooled_output)
        return pooled_output


@auto_docstring
class XmodPreTrainedModel(PreTrainedModel):
    config_class = XmodConfig
    base_model_prefix = "roberta"
    supports_gradient_checkpointing = True
    no_split_modules = ["XmodEmbeddings", "XmodSelfAttention", "XmodCrossAttention"]
    _supports_flash_attn = True
    _supports_sdpa = True
    _supports_flex_attn = True
    _supports_attention_backend = True
    _can_record_outputs = {
        "hidden_states": XmodLayer,
        "attentions": XmodSelfAttention,
        "cross_attentions": XmodCrossAttention,
    }

    @torch.no_grad()
    def _init_weights(self, module):
        """Initialize the weights"""
        super()._init_weights(module)
        if isinstance(module, XmodLMHead):
            init.zeros_(module.bias)
        elif isinstance(module, XmodEmbeddings):
            init.copy_(module.position_ids, torch.arange(module.position_ids.shape[-1]).expand((1, -1)))
            init.zeros_(module.token_type_ids)

    def set_default_language(self, language: str):
        """
        Set the default language code for the model. This is used when the language is not specified in the input.

        Args:
            language (`str`): The language code, such as `"en_XX"` or `"de_DE"`.
        """
        if language not in self.config.languages:
            raise ValueError(
                f"{self} does not have an adapter for {language}. Supported languages: {list(self.config.languages)}"
            )
        self.config.default_language = language

    def freeze_embeddings_and_language_adapters(self):
        """
        Freeze the embeddings and language adapters of the model. Usually, this is applied before the model is
        fine-tuned on a downstream task.
        """
        logger.info("Freezing embeddings")
        for parameter in self.roberta.embeddings.parameters():
            parameter.requires_grad = False
        logger.info("Freezing adapters")
        for layer in self.roberta.encoder.layer:
            if layer.output.adapter_layer_norm is not None:
                for parameter in layer.output.adapter_layer_norm.parameters():
                    parameter.requires_grad = False
            for parameter in layer.output.adapter_modules.parameters():
                parameter.requires_grad = False


@auto_docstring(
    custom_intro="""
    The model can behave as an encoder (with only self-attention) as well as a decoder, in which case a layer of
    cross-attention is added between the self-attention layers, following the architecture described in *Attention is
    all you need*_ by Ashish Vaswani, Noam Shazeer, Niki Parmar, Jakob Uszkoreit, Llion Jones, Aidan N. Gomez, Lukasz
    Kaiser and Illia Polosukhin.

    To behave as an decoder the model needs to be initialized with the `is_decoder` argument of the configuration set
    to `True`. To be used in a Seq2Seq model, the model needs to initialized with both `is_decoder` argument and
    `add_cross_attention` set to `True`; an `encoder_hidden_states` is then expected as an input to the forward pass.

    .. _*Attention is all you need*: https://huggingface.co/papers/1706.03762
    """
)
class XmodModel(XmodPreTrainedModel):
    def __init__(self, config, add_pooling_layer=True):
        r"""
        add_pooling_layer (bool, *optional*, defaults to `True`):
            Whether to add a pooling layer
        """
        super().__init__(config)
        self.config = config
        self.gradient_checkpointing = False

        self.embeddings = XmodEmbeddings(config)
        self.encoder = XmodEncoder(config)

        self.pooler = XmodPooler(config) if add_pooling_layer else None

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

    # Copied from transformers.models.roberta.modeling_roberta.RobertaModel.get_input_embeddings
    def get_input_embeddings(self):
        return self.embeddings.word_embeddings

    # Copied from transformers.models.roberta.modeling_roberta.RobertaModel.set_input_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,
        lang_ids: torch.LongTensor | 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,
        **kwargs: Unpack[TransformersKwargs],
    ) -> tuple[torch.Tensor] | BaseModelOutputWithPoolingAndCrossAttentions:
        r"""
        lang_ids (`torch.LongTensor` of shape `(batch_size, sequence_length)`, *optional*):
            Indices of the language adapters that should be activated for each sample, respectively. Default: the index
            that corresponds to `self.config.default_language`.
        """
        if (input_ids is None) ^ (inputs_embeds is not None):
            raise ValueError("You must specify exactly one of input_ids or inputs_embeds")

        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)
            )

        batch_size = input_ids.shape[0] if input_ids is not None else inputs_embeds.shape[0]
        device = input_ids.device if input_ids is not None else inputs_embeds.device
        past_key_values_length = past_key_values.get_seq_length() if past_key_values is not None else 0

        if lang_ids is None:
            if self.config.default_language is None:
                raise ValueError("Input language unknown. Please call `XmodPreTrainedModel.set_default_language()`")
            adapter_languages = list(self.encoder.layer[0].output.adapter_modules.keys())
            default_lang_id = adapter_languages.index(self.config.default_language)
            lang_ids = default_lang_id * torch.ones(batch_size, 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,
        )

        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,
            past_key_values=past_key_values,
        )

        encoder_outputs = self.encoder(
            embedding_output,
            lang_ids=lang_ids,
            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,
        )
        sequence_output = encoder_outputs[0]
        pooled_output = self.pooler(sequence_output) if self.pooler is not None else None

        return BaseModelOutputWithPoolingAndCrossAttentions(
            last_hidden_state=sequence_output,
            pooler_output=pooled_output,
            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,
        past_key_values,
    ):
        if self.config.is_decoder:
            attention_mask = create_causal_mask(
                config=self.config,
                inputs_embeds=embedding_output,
                attention_mask=attention_mask,
                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


@auto_docstring(
    custom_intro="""
    X-MOD Model with a `language modeling` head on top for CLM fine-tuning.
    """
)
class XmodForCausalLM(XmodPreTrainedModel, GenerationMixin):
    _tied_weights_keys = {
        "lm_head.decoder.weight": "roberta.embeddings.word_embeddings.weight",
        "lm_head.decoder.bias": "lm_head.bias",
    }

    # Copied from transformers.models.roberta.modeling_roberta.RobertaForCausalLM.__init__ with Roberta->Xmod
    def __init__(self, config):
        super().__init__(config)

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

        self.roberta = XmodModel(config, add_pooling_layer=False)
        self.lm_head = XmodLMHead(config)

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

    # Copied from transformers.models.roberta.modeling_roberta.RobertaForCausalLM.get_output_embeddings
    def get_output_embeddings(self):
        return self.lm_head.decoder

    # Copied from transformers.models.roberta.modeling_roberta.RobertaForCausalLM.set_output_embeddings
    def set_output_embeddings(self, new_embeddings):
        self.lm_head.decoder = new_embeddings

    @can_return_tuple
    @auto_docstring
    def forward(
        self,
        input_ids: torch.LongTensor | None = None,
        lang_ids: torch.LongTensor | None = None,
        attention_mask: torch.FloatTensor | None = None,
        token_type_ids: torch.LongTensor | None = None,
        position_ids: torch.LongTensor | None = None,
        inputs_embeds: torch.FloatTensor | None = None,
        encoder_hidden_states: torch.FloatTensor | None = None,
        encoder_attention_mask: torch.FloatTensor | None = None,
        labels: torch.LongTensor | None = None,
        past_key_values: tuple[tuple[torch.FloatTensor]] | None = None,
        use_cache: bool | None = None,
        logits_to_keep: int | torch.Tensor = 0,
        **kwargs: Unpack[TransformersKwargs],
    ) -> tuple[torch.Tensor] | CausalLMOutputWithCrossAttentions:
        r"""
        lang_ids (`torch.LongTensor` of shape `(batch_size, sequence_length)`, *optional*):
            Indices of the language adapters that should be activated for each sample, respectively. Default: the index
            that corresponds to `self.config.default_language`.
        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, XmodForCausalLM, AutoConfig
        >>> import torch

        >>> tokenizer = AutoTokenizer.from_pretrained("FacebookAI/xlm-roberta-base")
        >>> config = AutoConfig.from_pretrained("facebook/xmod-base")
        >>> config.is_decoder = True
        >>> model = XmodForCausalLM.from_pretrained("facebook/xmod-base", config=config)
        >>> model.set_default_language("en_XX")

        >>> 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: BaseModelOutputWithPoolingAndCrossAttentions = self.roberta(
            input_ids,
            lang_ids=lang_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,
            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.lm_head(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,
        )


@auto_docstring
class XmodForMaskedLM(XmodPreTrainedModel):
    _tied_weights_keys = {
        "lm_head.decoder.weight": "roberta.embeddings.word_embeddings.weight",
        "lm_head.decoder.bias": "lm_head.bias",
    }

    # Copied from transformers.models.roberta.modeling_roberta.RobertaForMaskedLM.__init__ with Roberta->Xmod
    def __init__(self, config):
        super().__init__(config)

        if config.is_decoder:
            logger.warning(
                "If you want to use `XmodForMaskedLM` make sure `config.is_decoder=False` for "
                "bi-directional self-attention."
            )

        self.roberta = XmodModel(config, add_pooling_layer=False)
        self.lm_head = XmodLMHead(config)

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

    # Copied from transformers.models.roberta.modeling_roberta.RobertaForMaskedLM.get_output_embeddings
    def get_output_embeddings(self):
        return self.lm_head.decoder

    # Copied from transformers.models.roberta.modeling_roberta.RobertaForMaskedLM.set_output_embeddings
    def set_output_embeddings(self, new_embeddings):
        self.lm_head.decoder = new_embeddings

    @can_return_tuple
    @auto_docstring
    def forward(
        self,
        input_ids: torch.LongTensor | None = None,
        lang_ids: torch.LongTensor | None = None,
        attention_mask: torch.FloatTensor | None = None,
        token_type_ids: torch.LongTensor | None = None,
        position_ids: torch.LongTensor | None = None,
        inputs_embeds: torch.FloatTensor | None = None,
        encoder_hidden_states: torch.FloatTensor | None = None,
        encoder_attention_mask: torch.FloatTensor | None = None,
        labels: torch.LongTensor | None = None,
        **kwargs: Unpack[TransformersKwargs],
    ) -> tuple[torch.Tensor] | MaskedLMOutput:
        r"""
        lang_ids (`torch.LongTensor` of shape `(batch_size, sequence_length)`, *optional*):
            Indices of the language adapters that should be activated for each sample, respectively. Default: the index
            that corresponds to `self.config.default_language`.
        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]`
        """
        outputs = self.roberta(
            input_ids,
            lang_ids=lang_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,
            return_dict=True,
            **kwargs,
        )
        sequence_output = outputs[0]
        prediction_scores = self.lm_head(sequence_output)

        masked_lm_loss = None
        if labels is not None:
            loss_fct = CrossEntropyLoss()
            masked_lm_loss = loss_fct(prediction_scores.view(-1, self.config.vocab_size), labels.view(-1))

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


# Copied from transformers.models.roberta.modeling_roberta.RobertaLMHead
class XmodLMHead(nn.Module):
    """Roberta Head for masked language modeling."""

    def __init__(self, config):
        super().__init__()
        self.dense = nn.Linear(config.hidden_size, config.hidden_size)
        self.layer_norm = nn.LayerNorm(config.hidden_size, eps=config.layer_norm_eps)

        self.decoder = nn.Linear(config.hidden_size, config.vocab_size)
        self.bias = nn.Parameter(torch.zeros(config.vocab_size))

    def forward(self, features, **kwargs):
        x = self.dense(features)
        x = gelu(x)
        x = self.layer_norm(x)

        # project back to size of vocabulary with bias
        x = self.decoder(x)

        return x


@auto_docstring(
    custom_intro="""
    X-MOD 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 XmodForSequenceClassification(XmodPreTrainedModel):
    # Copied from transformers.models.roberta.modeling_roberta.RobertaForSequenceClassification.__init__ with Roberta->Xmod
    def __init__(self, config):
        super().__init__(config)
        self.num_labels = config.num_labels
        self.config = config

        self.roberta = XmodModel(config, add_pooling_layer=False)
        self.classifier = XmodClassificationHead(config)

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

    @can_return_tuple
    @auto_docstring
    def forward(
        self,
        input_ids: torch.LongTensor | None = None,
        lang_ids: torch.LongTensor | None = None,
        attention_mask: torch.FloatTensor | None = None,
        token_type_ids: torch.LongTensor | None = None,
        position_ids: torch.LongTensor | None = None,
        inputs_embeds: torch.FloatTensor | None = None,
        labels: torch.LongTensor | None = None,
        **kwargs: Unpack[TransformersKwargs],
    ) -> tuple[torch.Tensor] | SequenceClassifierOutput:
        r"""
        lang_ids (`torch.LongTensor` of shape `(batch_size, sequence_length)`, *optional*):
            Indices of the language adapters that should be activated for each sample, respectively. Default: the index
            that corresponds to `self.config.default_language`.
        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).
        """
        outputs = self.roberta(
            input_ids,
            lang_ids=lang_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 = outputs[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=outputs.hidden_states,
            attentions=outputs.attentions,
        )


@auto_docstring
class XmodForMultipleChoice(XmodPreTrainedModel):
    # Copied from transformers.models.roberta.modeling_roberta.RobertaForMultipleChoice.__init__ with Roberta->Xmod
    def __init__(self, config):
        super().__init__(config)

        self.roberta = XmodModel(config)
        self.dropout = nn.Dropout(config.hidden_dropout_prob)
        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.LongTensor | None = None,
        lang_ids: torch.LongTensor | None = None,
        token_type_ids: torch.LongTensor | None = None,
        attention_mask: torch.FloatTensor | None = None,
        labels: torch.LongTensor | None = None,
        position_ids: torch.LongTensor | None = None,
        inputs_embeds: torch.FloatTensor | 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)
        lang_ids (`torch.LongTensor` of shape `(batch_size, num_choices, sequence_length)`, *optional*):
            Indices of the language adapters that should be activated for each sample, respectively. Default: the index
            that corresponds to `self.config.default_language`.
        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)
        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)
        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.
        """
        num_choices = input_ids.shape[1] if input_ids is not None else inputs_embeds.shape[1]

        flat_input_ids = input_ids.view(-1, input_ids.size(-1)) if input_ids is not None else None
        flat_lang_ids = lang_ids.repeat(input_ids.size(0) * input_ids.size(1)) if lang_ids is not None else None
        flat_position_ids = position_ids.view(-1, position_ids.size(-1)) if position_ids is not None else None
        flat_token_type_ids = token_type_ids.view(-1, token_type_ids.size(-1)) if token_type_ids is not None else None
        flat_attention_mask = attention_mask.view(-1, attention_mask.size(-1)) if attention_mask is not None else None
        flat_inputs_embeds = (
            inputs_embeds.view(-1, inputs_embeds.size(-2), inputs_embeds.size(-1))
            if inputs_embeds is not None
            else None
        )

        outputs = self.roberta(
            flat_input_ids,
            lang_ids=flat_lang_ids,
            position_ids=flat_position_ids,
            token_type_ids=flat_token_type_ids,
            attention_mask=flat_attention_mask,
            inputs_embeds=flat_inputs_embeds,
            return_dict=True,
            **kwargs,
        )
        pooled_output = outputs[1]

        pooled_output = self.dropout(pooled_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=outputs.hidden_states,
            attentions=outputs.attentions,
        )


@auto_docstring
class XmodForTokenClassification(XmodPreTrainedModel):
    # Copied from transformers.models.roberta.modeling_roberta.RobertaForTokenClassification.__init__ with Roberta->Xmod
    def __init__(self, config):
        super().__init__(config)
        self.num_labels = config.num_labels

        self.roberta = XmodModel(config, add_pooling_layer=False)
        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.LongTensor | None = None,
        lang_ids: torch.LongTensor | None = None,
        attention_mask: torch.FloatTensor | None = None,
        token_type_ids: torch.LongTensor | None = None,
        position_ids: torch.LongTensor | None = None,
        inputs_embeds: torch.FloatTensor | None = None,
        labels: torch.LongTensor | None = None,
        **kwargs: Unpack[TransformersKwargs],
    ) -> tuple[torch.Tensor] | TokenClassifierOutput:
        r"""
        lang_ids (`torch.LongTensor` of shape `(batch_size, sequence_length)`, *optional*):
            Indices of the language adapters that should be activated for each sample, respectively. Default: the index
            that corresponds to `self.config.default_language`.
        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]`.
        """
        outputs = self.roberta(
            input_ids,
            lang_ids=lang_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 = outputs[0]

        sequence_output = self.dropout(sequence_output)
        logits = self.classifier(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=outputs.hidden_states,
            attentions=outputs.attentions,
        )


# Copied from transformers.models.roberta.modeling_roberta.RobertaClassificationHead
class XmodClassificationHead(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.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 = torch.tanh(x)
        x = self.dropout(x)
        x = self.out_proj(x)
        return x


@auto_docstring
class XmodForQuestionAnswering(XmodPreTrainedModel):
    # Copied from transformers.models.roberta.modeling_roberta.RobertaForQuestionAnswering.__init__ with Roberta->Xmod
    def __init__(self, config):
        super().__init__(config)
        self.num_labels = config.num_labels

        self.roberta = XmodModel(config, add_pooling_layer=False)
        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.LongTensor | None = None,
        lang_ids: torch.LongTensor | None = None,
        attention_mask: torch.FloatTensor | None = None,
        token_type_ids: torch.LongTensor | None = None,
        position_ids: torch.LongTensor | None = None,
        inputs_embeds: torch.FloatTensor | None = None,
        start_positions: torch.LongTensor | None = None,
        end_positions: torch.LongTensor | None = None,
        **kwargs: Unpack[TransformersKwargs],
    ) -> tuple[torch.Tensor] | QuestionAnsweringModelOutput:
        r"""
        lang_ids (`torch.LongTensor` of shape `(batch_size, sequence_length)`, *optional*):
            Indices of the language adapters that should be activated for each sample, respectively. Default: the index
            that corresponds to `self.config.default_language`.
        """
        outputs = self.roberta(
            input_ids,
            lang_ids=lang_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 = outputs[0]

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

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

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

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


__all__ = [
    "XmodForCausalLM",
    "XmodForMaskedLM",
    "XmodForMultipleChoice",
    "XmodForQuestionAnswering",
    "XmodForSequenceClassification",
    "XmodForTokenClassification",
    "XmodModel",
    "XmodPreTrainedModel",
]