image-match/python/py/Lib/site-packages/transformers/models/mbart/modeling_mbart.py

# Copyright 2021, The Facebook AI Research Team and The HuggingFace Inc. team. All rights reserved.
#
# Licensed under the Apache License, Version 2.0 (the "License");
# you may not use this file except in compliance with the License.
# You may obtain a copy of the License at
#
#     http://www.apache.org/licenses/LICENSE-2.0
#
# Unless required by applicable law or agreed to in writing, software
# distributed under the License is distributed on an "AS IS" BASIS,
# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
# See the License for the specific language governing permissions and
# limitations under the License.
"""PyTorch MBART model."""

import math
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
from ...cache_utils import Cache, DynamicCache, EncoderDecoderCache
from ...generation import GenerationMixin
from ...masking_utils import create_bidirectional_mask, create_causal_mask
from ...modeling_flash_attention_utils import (
    FlashAttentionKwargs,
)
from ...modeling_layers import GradientCheckpointingLayer
from ...modeling_outputs import (
    BaseModelOutput,
    BaseModelOutputWithPastAndCrossAttentions,
    CausalLMOutputWithCrossAttentions,
    Seq2SeqLMOutput,
    Seq2SeqModelOutput,
    Seq2SeqQuestionAnsweringModelOutput,
    Seq2SeqSequenceClassifierOutput,
)
from ...modeling_utils import ALL_ATTENTION_FUNCTIONS, PreTrainedModel
from ...processing_utils import Unpack
from ...utils import (
    TransformersKwargs,
    auto_docstring,
    can_return_tuple,
    is_torch_flex_attn_available,
    is_torchdynamo_compiling,
    logging,
    torch_compilable_check,
)
from ...utils.generic import merge_with_config_defaults
from ...utils.output_capturing import OutputRecorder, capture_outputs
from .configuration_mbart import MBartConfig


if is_torch_flex_attn_available():
    pass


logger = logging.get_logger(__name__)


def shift_tokens_right(input_ids: torch.Tensor, pad_token_id: int):
    """
    Shift input ids one token to the right, and wrap the last non pad token (the <LID> token) Note that MBart does not
    have a single `decoder_start_token_id` in contrast to other Bart-like models.
    """
    prev_output_tokens = input_ids.clone()

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

    index_of_eos = (prev_output_tokens.ne(pad_token_id).sum(dim=1) - 1).unsqueeze(-1)
    decoder_start_tokens = prev_output_tokens.gather(1, index_of_eos).squeeze()
    prev_output_tokens[:, 1:] = prev_output_tokens[:, :-1].clone()
    prev_output_tokens[:, 0] = decoder_start_tokens

    return prev_output_tokens


# Copied from transformers.models.bart.modeling_bart.BartLearnedPositionalEmbedding with Bart->MBart
class MBartLearnedPositionalEmbedding(nn.Embedding):
    """
    This module learns positional embeddings up to a fixed maximum size.
    """

    def __init__(self, num_embeddings: int, embedding_dim: int):
        # MBart is set up so that if padding_idx is specified then offset the embedding ids by 2
        # and adjust num_embeddings appropriately. Other models don't have this hack
        self.offset = 2
        super().__init__(num_embeddings + self.offset, embedding_dim)

    def forward(
        self, input_ids: torch.Tensor, past_key_values_length: int = 0, position_ids: torch.Tensor | None = None
    ):
        """`input_ids' shape is expected to be [bsz x seqlen]."""

        if position_ids is None:
            bsz, seq_len = input_ids.shape[:2]
            position_ids = torch.arange(
                past_key_values_length, past_key_values_length + seq_len, dtype=torch.long, device=self.weight.device
            ).expand(bsz, -1)
        else:
            position_ids = position_ids.unsqueeze(0)

        return super().forward(position_ids + self.offset)


# Copied from transformers.models.bart.modeling_bart.BartScaledWordEmbedding with Bart->MBart
class MBartScaledWordEmbedding(nn.Embedding):
    """
    This module overrides nn.Embeddings' forward by multiplying with embeddings scale.
    """

    def __init__(self, num_embeddings: int, embedding_dim: int, padding_idx: int, embed_scale: float | None = 1.0):
        super().__init__(num_embeddings, embedding_dim, padding_idx)
        self.embed_scale = embed_scale

    def forward(self, input_ids: torch.Tensor):
        return super().forward(input_ids) * self.embed_scale


# 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.bart.modeling_bart.BartAttention with Bart->MBart
class MBartAttention(nn.Module):
    """Multi-headed attention from 'Attention Is All You Need' paper"""

    def __init__(
        self,
        embed_dim: int,
        num_heads: int,
        dropout: float = 0.0,
        is_decoder: bool = False,
        bias: bool = True,
        is_causal: bool = False,
        config: MBartConfig | None = None,
        layer_idx: int | None = None,
    ):
        super().__init__()
        self.embed_dim = embed_dim
        self.num_heads = num_heads
        self.dropout = dropout
        self.head_dim = embed_dim // num_heads
        self.config = config

        if (self.head_dim * num_heads) != self.embed_dim:
            raise ValueError(
                f"embed_dim must be divisible by num_heads (got `embed_dim`: {self.embed_dim}"
                f" and `num_heads`: {num_heads})."
            )
        self.scaling = self.head_dim**-0.5
        self.is_decoder = is_decoder
        self.is_causal = is_causal
        self.layer_idx = layer_idx
        if layer_idx is None and self.is_decoder:
            logger.warning_once(
                f"Instantiating a decoder {self.__class__.__name__} without passing `layer_idx` is not recommended and "
                "will lead to errors during the forward call, if caching is used. Please make sure to provide a `layer_idx` "
                "when creating this class."
            )

        self.k_proj = nn.Linear(embed_dim, embed_dim, bias=bias)
        self.v_proj = nn.Linear(embed_dim, embed_dim, bias=bias)
        self.q_proj = nn.Linear(embed_dim, embed_dim, bias=bias)
        self.out_proj = nn.Linear(embed_dim, embed_dim, bias=bias)

    def forward(
        self,
        hidden_states: torch.Tensor,
        key_value_states: torch.Tensor | None = None,
        past_key_values: Cache | None = None,
        attention_mask: torch.Tensor | None = None,
        # TODO: we need a refactor so that the different attention modules can get their specific kwargs
        # ATM, we have mixed things encoder, decoder, and encoder-decoder attn
        **kwargs: Unpack[FlashAttentionKwargs],
    ) -> tuple[torch.Tensor, torch.Tensor | None]:
        """Input shape: Batch x Time x Channel"""

        # if key_value_states are provided this layer is used as a cross-attention layer
        # for the decoder
        is_cross_attention = key_value_states is not None

        # determine input shapes
        input_shape = hidden_states.shape[:-1]

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

        # get query proj
        query_states = self.q_proj(hidden_states).view(hidden_shape).transpose(1, 2)

        is_updated = False
        if past_key_values is not None:
            if isinstance(past_key_values, EncoderDecoderCache):
                is_updated = past_key_values.is_updated.get(self.layer_idx)
                if is_cross_attention:
                    # after the first generated id, we can subsequently re-use all key/value_states from cache
                    curr_past_key_values = past_key_values.cross_attention_cache
                else:
                    curr_past_key_values = past_key_values.self_attention_cache
            else:
                curr_past_key_values = past_key_values

        current_states = key_value_states if is_cross_attention else hidden_states
        if is_cross_attention and past_key_values is not None and is_updated:
            # reuse k,v, cross_attentions
            key_states = curr_past_key_values.layers[self.layer_idx].keys
            value_states = curr_past_key_values.layers[self.layer_idx].values
        else:
            key_states = self.k_proj(current_states)
            value_states = self.v_proj(current_states)
            kv_shape = (*current_states.shape[:-1], -1, self.head_dim)
            key_states = key_states.view(kv_shape).transpose(1, 2)
            value_states = value_states.view(kv_shape).transpose(1, 2)

            if past_key_values is not None:
                key_states, value_states = curr_past_key_values.update(key_states, value_states, self.layer_idx)
                # set flag that curr layer for cross-attn is already updated so we can re-use in subsequent calls
                if is_cross_attention and isinstance(past_key_values, EncoderDecoderCache):
                    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_states,
            key_states,
            value_states,
            attention_mask,
            dropout=0.0 if not self.training else self.dropout,
            scaling=self.scaling,
            **kwargs,
        )

        attn_output = attn_output.reshape(*input_shape, -1).contiguous()
        attn_output = self.out_proj(attn_output)

        return attn_output, attn_weights


class MBartEncoderLayer(GradientCheckpointingLayer):
    def __init__(self, config: MBartConfig):
        super().__init__()
        self.embed_dim = config.d_model

        self.self_attn = MBartAttention(
            embed_dim=self.embed_dim,
            num_heads=config.encoder_attention_heads,
            dropout=config.attention_dropout,
            config=config,
        )
        self.self_attn_layer_norm = nn.LayerNorm(self.embed_dim)
        self.dropout = config.dropout
        self.activation_fn = ACT2FN[config.activation_function]
        self.activation_dropout = config.activation_dropout
        self.fc1 = nn.Linear(self.embed_dim, config.encoder_ffn_dim)
        self.fc2 = nn.Linear(config.encoder_ffn_dim, self.embed_dim)
        self.final_layer_norm = nn.LayerNorm(self.embed_dim)

    def forward(
        self,
        hidden_states: torch.Tensor,
        attention_mask: torch.Tensor,
        **kwargs: Unpack[TransformersKwargs],
    ) -> torch.Tensor:
        """
        Args:
            hidden_states (`torch.FloatTensor`): input to the layer of shape `(batch, seq_len, embed_dim)`
            attention_mask (`torch.FloatTensor`): attention mask of size
                `(batch, 1, tgt_len, src_len)` where padding elements are indicated by very large negative values.
        """
        residual = hidden_states
        hidden_states = self.self_attn_layer_norm(hidden_states)
        hidden_states, _ = self.self_attn(
            hidden_states=hidden_states,
            attention_mask=attention_mask,
            **kwargs,
        )
        hidden_states = nn.functional.dropout(hidden_states, p=self.dropout, training=self.training)
        hidden_states = residual + hidden_states

        residual = hidden_states
        hidden_states = self.final_layer_norm(hidden_states)
        hidden_states = self.activation_fn(self.fc1(hidden_states))
        hidden_states = nn.functional.dropout(hidden_states, p=self.activation_dropout, training=self.training)
        hidden_states = self.fc2(hidden_states)
        hidden_states = nn.functional.dropout(hidden_states, p=self.dropout, training=self.training)
        hidden_states = residual + hidden_states

        if hidden_states.dtype == torch.float16:
            clamp_value = torch.finfo(hidden_states.dtype).max - 1000
            hidden_states = torch.clamp(hidden_states, min=-clamp_value, max=clamp_value)

        return hidden_states


class MBartDecoderLayer(GradientCheckpointingLayer):
    def __init__(self, config: MBartConfig, layer_idx: int | None = None):
        super().__init__()
        self.embed_dim = config.d_model

        self.self_attn = MBartAttention(
            embed_dim=self.embed_dim,
            num_heads=config.decoder_attention_heads,
            dropout=config.attention_dropout,
            is_decoder=True,
            is_causal=True,
            config=config,
            layer_idx=layer_idx,
        )
        self.dropout = config.dropout
        self.activation_fn = ACT2FN[config.activation_function]
        self.activation_dropout = config.activation_dropout

        self.self_attn_layer_norm = nn.LayerNorm(self.embed_dim)
        self.encoder_attn = MBartAttention(
            self.embed_dim,
            config.decoder_attention_heads,
            dropout=config.attention_dropout,
            is_decoder=True,
            config=config,
            layer_idx=layer_idx,
        )
        self.encoder_attn_layer_norm = nn.LayerNorm(self.embed_dim)
        self.fc1 = nn.Linear(self.embed_dim, config.decoder_ffn_dim)
        self.fc2 = nn.Linear(config.decoder_ffn_dim, self.embed_dim)
        self.final_layer_norm = nn.LayerNorm(self.embed_dim)

    def forward(
        self,
        hidden_states: torch.Tensor,
        attention_mask: torch.Tensor | None = None,
        encoder_hidden_states: torch.Tensor | None = None,
        encoder_attention_mask: torch.Tensor | None = None,
        past_key_values: Cache | None = None,
        use_cache: bool | None = True,
        **kwargs: Unpack[TransformersKwargs],
    ) -> torch.Tensor:
        """
        Args:
            hidden_states (`torch.FloatTensor`): input to the layer of shape `(batch, seq_len, embed_dim)`
            attention_mask (`torch.FloatTensor`): attention mask of size
                `(batch, 1, tgt_len, src_len)` where padding elements are indicated by very large negative values.
            encoder_hidden_states (`torch.FloatTensor`):
                cross attention input to the layer of shape `(batch, seq_len, embed_dim)`
            encoder_attention_mask (`torch.FloatTensor`): encoder attention mask of size
                `(batch, 1, tgt_len, src_len)` where padding elements are indicated by very large negative values.
            past_key_values (`Cache`): cached past key and value projection states
        """
        residual = hidden_states
        hidden_states = self.self_attn_layer_norm(hidden_states)

        # Self Attention
        hidden_states, _ = self.self_attn(
            hidden_states=hidden_states,
            past_key_values=past_key_values,
            attention_mask=attention_mask,
            **kwargs,
        )
        hidden_states = nn.functional.dropout(hidden_states, p=self.dropout, training=self.training)
        hidden_states = residual + hidden_states

        # Cross-Attention Block
        if encoder_hidden_states is not None:
            residual = hidden_states
            hidden_states = self.encoder_attn_layer_norm(hidden_states)

            hidden_states, _ = self.encoder_attn(
                hidden_states=hidden_states,
                key_value_states=encoder_hidden_states,
                attention_mask=encoder_attention_mask,
                past_key_values=past_key_values,
                **kwargs,
            )
            hidden_states = nn.functional.dropout(hidden_states, p=self.dropout, training=self.training)
            hidden_states = residual + hidden_states

        # Fully Connected
        residual = hidden_states
        hidden_states = self.final_layer_norm(hidden_states)
        hidden_states = self.activation_fn(self.fc1(hidden_states))
        hidden_states = nn.functional.dropout(hidden_states, p=self.activation_dropout, training=self.training)
        hidden_states = self.fc2(hidden_states)
        hidden_states = nn.functional.dropout(hidden_states, p=self.dropout, training=self.training)
        hidden_states = residual + hidden_states

        return hidden_states


# Copied from transformers.models.bart.modeling_bart.BartClassificationHead with Bart->MBart
class MBartClassificationHead(nn.Module):
    """Head for sentence-level classification tasks."""

    def __init__(
        self,
        input_dim: int,
        inner_dim: int,
        num_classes: int,
        pooler_dropout: float,
    ):
        super().__init__()
        self.dense = nn.Linear(input_dim, inner_dim)
        self.dropout = nn.Dropout(p=pooler_dropout)
        self.out_proj = nn.Linear(inner_dim, num_classes)

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


@auto_docstring
class MBartPreTrainedModel(PreTrainedModel):
    config: MBartConfig
    base_model_prefix = "model"
    supports_gradient_checkpointing = True
    _no_split_modules = ["MBartDecoderLayer", "MBartEncoderLayer", "MBartAttention"]
    _supports_flash_attn = True
    _supports_sdpa = True
    _supports_flex_attn = True
    _can_compile_fullgraph = True

    def _init_weights(self, module):
        super()._init_weights(module)
        if isinstance(module, MBartForConditionalGeneration):
            init.zeros_(module.final_logits_bias)

    @property
    def dummy_inputs(self):
        pad_token = self.config.pad_token_id
        input_ids = torch.tensor([[0, 6, 10, 4, 2], [0, 8, 12, 2, pad_token]], device=self.device)
        dummy_inputs = {
            "attention_mask": input_ids.ne(pad_token),
            "input_ids": input_ids,
        }
        return dummy_inputs


class MBartEncoder(MBartPreTrainedModel):
    """
    Transformer encoder consisting of *config.encoder_layers* self attention layers. Each layer is a
    [`MBartEncoderLayer`].

    Args:
        config: MBartConfig
        embed_tokens (nn.Embedding): output embedding
    """

    _can_record_outputs = {
        "hidden_states": MBartEncoderLayer,
        "attentions": OutputRecorder(MBartAttention, index=1, layer_name="self_attn"),
    }

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

        self.dropout = config.dropout
        self.layerdrop = config.encoder_layerdrop

        embed_dim = config.d_model
        self.padding_idx = config.pad_token_id
        self.max_source_positions = config.max_position_embeddings
        embed_scale = math.sqrt(embed_dim) if config.scale_embedding else 1.0

        self.embed_tokens = MBartScaledWordEmbedding(
            config.vocab_size, embed_dim, self.padding_idx, embed_scale=embed_scale
        )

        self.embed_positions = MBartLearnedPositionalEmbedding(
            config.max_position_embeddings,
            embed_dim,
        )
        self.layers = nn.ModuleList([MBartEncoderLayer(config) for _ in range(config.encoder_layers)])
        self.config = config
        self.layernorm_embedding = nn.LayerNorm(embed_dim)
        self.layer_norm = nn.LayerNorm(config.d_model)

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

    def _backward_compatibility_gradient_checkpointing(self):
        # Override to not delete the attribute from the config
        if self.supports_gradient_checkpointing and getattr(self.config, "gradient_checkpointing", False):
            self.gradient_checkpointing_enable()

    @merge_with_config_defaults
    @capture_outputs
    def forward(
        self,
        input_ids: torch.LongTensor | None = None,
        attention_mask: torch.Tensor | None = None,
        inputs_embeds: torch.FloatTensor | None = None,
        **kwargs: Unpack[TransformersKwargs],
    ) -> tuple | BaseModelOutput:
        r"""
        Args:
            input_ids (`torch.LongTensor` of shape `(batch_size, sequence_length)`):
                Indices of input sequence tokens in the vocabulary. Padding will be ignored by default should you
                provide it.

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

                [What are input IDs?](../glossary#input-ids)
            attention_mask (`torch.Tensor` of shape `(batch_size, sequence_length)`, *optional*):
                Mask to avoid performing attention on padding token indices. Mask values selected in `[0, 1]`:

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

                [What are attention masks?](../glossary#attention-mask)
            inputs_embeds (`torch.FloatTensor` of shape `(batch_size, sequence_length, hidden_size)`, *optional*):
                Optionally, instead of passing `input_ids` you can choose to directly pass an embedded representation.
                This is useful if you want more control over how to convert `input_ids` indices into associated vectors
                than the model's internal embedding lookup matrix.
        """
        if (input_ids is None) ^ (inputs_embeds is not None):
            raise ValueError("You must specify exactly one of input_ids or inputs_embeds")

        if inputs_embeds is None:
            inputs_embeds = self.embed_tokens(input_ids)

        embed_pos = self.embed_positions(inputs_embeds[..., -1])  # just for the shape

        hidden_states = inputs_embeds + embed_pos.to(inputs_embeds.device)
        hidden_states = self.layernorm_embedding(hidden_states)
        hidden_states = nn.functional.dropout(hidden_states, p=self.dropout, training=self.training)

        attention_mask = create_bidirectional_mask(
            config=self.config,
            inputs_embeds=inputs_embeds,
            attention_mask=attention_mask,
        )

        for idx, encoder_layer in enumerate(self.layers):
            # add LayerDrop (see https://huggingface.co/papers/1909.11556 for description)
            to_drop = False
            if self.training:
                dropout_probability = torch.rand([])
                if dropout_probability < self.layerdrop:  # skip the layer
                    to_drop = True

            if not to_drop:
                hidden_states = encoder_layer(
                    hidden_states,
                    attention_mask,
                    **kwargs,
                )

        hidden_states = self.layer_norm(hidden_states)

        return BaseModelOutput(last_hidden_state=hidden_states)


class MBartDecoder(MBartPreTrainedModel):
    """
    Transformer decoder consisting of *config.decoder_layers* layers. Each layer is a [`MBartDecoderLayer`]

    Args:
        config: MBartConfig
        embed_tokens (nn.Embedding): output embedding
    """

    _can_record_outputs = {
        "hidden_states": MBartDecoderLayer,
        "attentions": OutputRecorder(MBartAttention, index=1, layer_name="self_attn"),
        "cross_attentions": OutputRecorder(MBartAttention, index=1, layer_name="encoder_attn"),
    }

    def __init__(self, config: MBartConfig):
        super().__init__(config)
        self.dropout = config.dropout
        self.layerdrop = config.decoder_layerdrop
        self.padding_idx = config.pad_token_id
        self.max_target_positions = config.max_position_embeddings
        embed_scale = math.sqrt(config.d_model) if config.scale_embedding else 1.0

        self.embed_tokens = MBartScaledWordEmbedding(
            config.vocab_size, config.d_model, self.padding_idx, embed_scale=embed_scale
        )

        self.embed_positions = MBartLearnedPositionalEmbedding(
            config.max_position_embeddings,
            config.d_model,
        )
        self.layers = nn.ModuleList([MBartDecoderLayer(config, layer_idx=i) for i in range(config.decoder_layers)])
        self.config = config

        self.layernorm_embedding = nn.LayerNorm(config.d_model)
        self.layer_norm = nn.LayerNorm(config.d_model)

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

    @merge_with_config_defaults
    @capture_outputs
    def forward(
        self,
        input_ids: torch.LongTensor | None = None,
        attention_mask: torch.Tensor | None = None,
        encoder_hidden_states: torch.FloatTensor | None = None,
        encoder_attention_mask: torch.LongTensor | None = None,
        past_key_values: Cache | None = None,
        inputs_embeds: torch.FloatTensor | None = None,
        use_cache: bool | None = None,
        **kwargs: Unpack[TransformersKwargs],
    ) -> tuple | BaseModelOutputWithPastAndCrossAttentions:
        r"""
        Args:
            input_ids (`torch.LongTensor` of shape `(batch_size, sequence_length)`):
                Indices of input sequence tokens in the vocabulary. Padding will be ignored by default should you
                provide it.

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

                [What are input IDs?](../glossary#input-ids)
            attention_mask (`torch.Tensor` of shape `(batch_size, sequence_length)`, *optional*):
                Mask to avoid performing attention on padding token indices. Mask values selected in `[0, 1]`:

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

                [What are attention masks?](../glossary#attention-mask)
            encoder_hidden_states (`torch.FloatTensor` of shape `(batch_size, encoder_sequence_length, hidden_size)`, *optional*):
                Sequence of hidden-states at the output of the last layer of the encoder. Used in the cross-attention
                of the decoder.
            encoder_attention_mask (`torch.LongTensor` of shape `(batch_size, encoder_sequence_length)`, *optional*):
                Mask to avoid performing cross-attention on padding tokens indices of encoder input_ids. Mask values
                selected in `[0, 1]`:

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

                [What are attention masks?](../glossary#attention-mask)
            past_key_values (`Cache`, *optional*, returned when `use_cache=True` is passed or when `config.use_cache=True`):
                It is a [`~cache_utils.Cache`] instance. For more details, see our [kv cache guide](https://huggingface.co/docs/transformers/en/kv_cache).

                Contains pre-computed hidden-states (key and values in the self-attention blocks and in the
                cross-attention blocks) that can be used (see `past_key_values` input) to speed up sequential decoding.

                If `past_key_values` are used, the user can optionally input only the last `decoder_input_ids` (those
                that don't have their past key value states given to this model) of shape `(batch_size, 1)` instead of
                all `decoder_input_ids` of shape `(batch_size, sequence_length)`.
            inputs_embeds (`torch.FloatTensor` of shape `(batch_size, sequence_length, hidden_size)`, *optional*):
                Optionally, instead of passing `input_ids` you can choose to directly pass an embedded representation.
                This is useful if you want more control over how to convert `input_ids` indices into associated vectors
                than the model's internal embedding lookup matrix.
        """
        if (input_ids is None) ^ (inputs_embeds is not None):
            raise ValueError("You cannot specify both decoder_input_ids and decoder_inputs_embeds at the same time")

        if inputs_embeds is None:
            inputs_embeds = self.embed_tokens(input_ids)

        # initialize `past_key_values`
        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, seq_length = inputs_embeds.size()[:-1]
        past_key_values_length = past_key_values.get_seq_length() if past_key_values is not None else 0
        position_ids = torch.arange(seq_length, device=inputs_embeds.device) + past_key_values_length

        if attention_mask is None and not is_torchdynamo_compiling():
            # required mask seq length can be calculated via length of past cache
            mask_seq_length = past_key_values_length + seq_length
            attention_mask = torch.ones(batch_size, mask_seq_length, device=inputs_embeds.device)

        self_attn_cache = (
            past_key_values.self_attention_cache
            if isinstance(past_key_values, EncoderDecoderCache)
            else past_key_values
        )

        causal_mask = create_causal_mask(
            config=self.config,
            inputs_embeds=inputs_embeds,
            attention_mask=attention_mask,
            past_key_values=self_attn_cache,
        )
        encoder_attention_mask = create_bidirectional_mask(
            config=self.config,
            inputs_embeds=inputs_embeds,
            attention_mask=encoder_attention_mask,
            encoder_hidden_states=encoder_hidden_states,
        )

        # embed positions
        position_ids = self.embed_positions(input, past_key_values_length, position_ids=position_ids)

        hidden_states = inputs_embeds + position_ids.to(inputs_embeds.device)
        hidden_states = self.layernorm_embedding(hidden_states)

        hidden_states = nn.functional.dropout(hidden_states, p=self.dropout, training=self.training)

        for idx, decoder_layer in enumerate(self.layers):
            # add LayerDrop (see https://huggingface.co/papers/1909.11556 for description)
            if self.training:
                dropout_probability = torch.rand([])
                if dropout_probability < self.layerdrop:
                    continue

            hidden_states = decoder_layer(
                hidden_states,
                causal_mask,
                encoder_hidden_states,  # as a positional argument for gradient checkpointing
                encoder_attention_mask=encoder_attention_mask,
                past_key_values=past_key_values,
                use_cache=use_cache,
                **kwargs,
            )

        hidden_states = self.layer_norm(hidden_states)

        return BaseModelOutputWithPastAndCrossAttentions(
            last_hidden_state=hidden_states,
            past_key_values=past_key_values,
        )


@auto_docstring
class MBartModel(MBartPreTrainedModel):
    _tied_weights_keys = {
        "decoder.embed_tokens.weight": "shared.weight",
        "encoder.embed_tokens.weight": "shared.weight",
    }

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

        padding_idx, vocab_size = config.pad_token_id, config.vocab_size
        embed_scale = math.sqrt(config.d_model) if config.scale_embedding else 1.0
        self.shared = MBartScaledWordEmbedding(vocab_size, config.d_model, padding_idx, embed_scale=embed_scale)

        self.encoder = MBartEncoder(config)
        self.decoder = MBartDecoder(config)

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

    def get_input_embeddings(self):
        return self.shared

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

    @can_return_tuple
    @auto_docstring
    def forward(
        self,
        input_ids: torch.LongTensor | None = None,
        attention_mask: torch.Tensor | None = None,
        decoder_input_ids: torch.LongTensor | None = None,
        decoder_attention_mask: torch.LongTensor | None = None,
        encoder_outputs: tuple[tuple[torch.FloatTensor]] | None = None,
        past_key_values: Cache | None = None,
        inputs_embeds: torch.FloatTensor | None = None,
        decoder_inputs_embeds: torch.FloatTensor | None = None,
        use_cache: bool | None = None,
        return_dict: bool | None = None,
        **kwargs: Unpack[TransformersKwargs],
    ) -> Seq2SeqModelOutput | tuple[torch.FloatTensor]:
        r"""
        decoder_input_ids (`torch.LongTensor` of shape `(batch_size, target_sequence_length)`, *optional*):
            Indices of decoder input sequence tokens in the vocabulary.

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

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

            MBart uses a specific language id token as the starting token for `decoder_input_ids` generation that
            varies according to source and target language, *e.g.* 25004 for *en_XX*, and 25003 for *de_DE*. If
            `past_key_values` is used, optionally only the last `decoder_input_ids` have to be input (see
            `past_key_values`).

            For translation and summarization training, `decoder_input_ids` should be provided. If no
            `decoder_input_ids` is provided, the model will create this tensor by shifting the `input_ids` to the right
            for denoising pre-training following the paper.
        decoder_attention_mask (`torch.LongTensor` of shape `(batch_size, target_sequence_length)`, *optional*):
            Default behavior: generate a tensor that ignores pad tokens in `decoder_input_ids`. Causal mask will also
            be used by default.
        """
        return_dict = return_dict if return_dict is not None else self.config.use_return_dict

        # different to other models, MBart automatically creates decoder_input_ids from
        # input_ids if no decoder_input_ids are provided
        if decoder_input_ids is None and decoder_inputs_embeds is None:
            decoder_input_ids = shift_tokens_right(input_ids, self.config.pad_token_id)

        if encoder_outputs is None:
            encoder_outputs = self.encoder(
                input_ids=input_ids,
                attention_mask=attention_mask,
                inputs_embeds=inputs_embeds,
                return_dict=return_dict,
                **kwargs,
            )
        # If the user passed a tuple for encoder_outputs, we wrap it in a BaseModelOutput when return_dict=True
        elif return_dict and not isinstance(encoder_outputs, BaseModelOutput):
            encoder_outputs = BaseModelOutput(
                last_hidden_state=encoder_outputs[0],
                hidden_states=encoder_outputs[1] if len(encoder_outputs) > 1 else None,
                attentions=encoder_outputs[2] if len(encoder_outputs) > 2 else None,
            )

        # decoder outputs consists of (dec_features, past_key_values, dec_hidden, dec_attn)
        decoder_outputs = self.decoder(
            input_ids=decoder_input_ids,
            attention_mask=decoder_attention_mask,
            encoder_hidden_states=encoder_outputs[0],
            encoder_attention_mask=attention_mask,
            past_key_values=past_key_values,
            inputs_embeds=decoder_inputs_embeds,
            use_cache=use_cache,
            return_dict=return_dict,
            **kwargs,
        )

        if not return_dict:
            return decoder_outputs + encoder_outputs

        return Seq2SeqModelOutput(
            last_hidden_state=decoder_outputs.last_hidden_state,
            past_key_values=decoder_outputs.past_key_values,
            decoder_hidden_states=decoder_outputs.hidden_states,
            decoder_attentions=decoder_outputs.attentions,
            cross_attentions=decoder_outputs.cross_attentions,
            encoder_last_hidden_state=encoder_outputs.last_hidden_state,
            encoder_hidden_states=encoder_outputs.hidden_states,
            encoder_attentions=encoder_outputs.attentions,
        )


@auto_docstring(
    custom_intro="""
    The MBART Model with a language modeling head. Can be used for summarization, after fine-tuning the pretrained models.
    """
)
class MBartForConditionalGeneration(MBartPreTrainedModel, GenerationMixin):
    base_model_prefix = "model"
    _keys_to_ignore_on_load_missing = ["final_logits_bias"]
    _tied_weights_keys = {"lm_head.weight": "model.shared.weight"}

    def __init__(self, config: MBartConfig):
        super().__init__(config)
        self.model = MBartModel(config)
        self.register_buffer("final_logits_bias", torch.zeros((1, self.model.shared.num_embeddings)))
        self.lm_head = nn.Linear(config.d_model, self.model.shared.num_embeddings, bias=False)

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

    def resize_token_embeddings(
        self, new_num_tokens: int, pad_to_multiple_of: int | None = None, mean_resizing: bool = True
    ) -> nn.Embedding:
        new_embeddings = super().resize_token_embeddings(new_num_tokens, pad_to_multiple_of, mean_resizing)
        self._resize_final_logits_bias(new_embeddings.weight.shape[0])
        return new_embeddings

    def _resize_final_logits_bias(self, new_num_tokens: int) -> None:
        old_num_tokens = self.final_logits_bias.shape[-1]
        if new_num_tokens <= old_num_tokens:
            new_bias = self.final_logits_bias[:, :new_num_tokens]
        else:
            extra_bias = torch.zeros((1, new_num_tokens - old_num_tokens), device=self.final_logits_bias.device)
            new_bias = torch.cat([self.final_logits_bias, extra_bias], dim=1)
        self.register_buffer("final_logits_bias", new_bias)

    @auto_docstring
    def forward(
        self,
        input_ids: torch.LongTensor | None = None,
        attention_mask: torch.Tensor | None = None,
        decoder_input_ids: torch.LongTensor | None = None,
        decoder_attention_mask: torch.LongTensor | None = None,
        encoder_outputs: tuple[tuple[torch.FloatTensor]] | None = None,
        past_key_values: Cache | None = None,
        inputs_embeds: torch.FloatTensor | None = None,
        decoder_inputs_embeds: torch.FloatTensor | None = None,
        labels: torch.LongTensor | None = None,
        use_cache: bool | None = None,
        return_dict: bool | None = None,
        **kwargs: Unpack[TransformersKwargs],
    ) -> Seq2SeqLMOutput | tuple[torch.FloatTensor]:
        r"""
        decoder_input_ids (`torch.LongTensor` of shape `(batch_size, target_sequence_length)`, *optional*):
            Indices of decoder input sequence tokens in the vocabulary.

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

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

            MBart uses a specific language id token as the starting token for `decoder_input_ids` generation that
            varies according to source and target language, *e.g.* 25004 for *en_XX*, and 25003 for *de_DE*. If
            `past_key_values` is used, optionally only the last `decoder_input_ids` have to be input (see
            `past_key_values`).

            For translation and summarization training, `decoder_input_ids` should be provided. If no
            `decoder_input_ids` is provided, the model will create this tensor by shifting the `input_ids` to the right
            for denoising pre-training following the paper.
        decoder_attention_mask (`torch.LongTensor` of shape `(batch_size, target_sequence_length)`, *optional*):
            Default behavior: generate a tensor that ignores pad tokens in `decoder_input_ids`. Causal mask will also
            be used by default.
        labels (`torch.LongTensor` of shape `(batch_size, sequence_length)`, *optional*):
            Labels for computing the masked language modeling loss. Indices should either be in `[0, ...,
            config.vocab_size]` or -100 (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 Translation:

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

        >>> model = MBartForConditionalGeneration.from_pretrained("facebook/mbart-large-en-ro")
        >>> tokenizer = AutoTokenizer.from_pretrained("facebook/mbart-large-en-ro")

        >>> example_english_phrase = "42 is the answer"
        >>> inputs = tokenizer(example_english_phrase, return_tensors="pt")

        >>> # Translate
        >>> generated_ids = model.generate(**inputs, num_beams=4, max_length=5)
        >>> tokenizer.batch_decode(generated_ids, skip_special_tokens=True, clean_up_tokenization_spaces=False)[0]
        '42 este răspuns'
        ```

        Mask filling example:

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

        >>> model = MBartForConditionalGeneration.from_pretrained("facebook/mbart-large-cc25")
        >>> tokenizer = AutoTokenizer.from_pretrained("facebook/mbart-large-cc25")

        >>> # de_DE is the language symbol id <LID> for German
        >>> TXT = "</s> Meine Freunde sind <mask> nett aber sie essen zu viel Kuchen. </s> de_DE"

        >>> input_ids = tokenizer([TXT], add_special_tokens=False, return_tensors="pt")["input_ids"]
        >>> logits = model(input_ids).logits

        >>> masked_index = (input_ids[0] == tokenizer.mask_token_id).nonzero().item()
        >>> probs = logits[0, masked_index].softmax(dim=0)
        >>> values, predictions = probs.topk(5)

        >>> tokenizer.decode(predictions).split()
        ['nett', 'sehr', 'ganz', 'nicht', 'so']
        ```
        """
        return_dict = return_dict if return_dict is not None else self.config.return_dict

        if labels is not None:
            if use_cache:
                logger.warning("The `use_cache` argument is changed to `False` since `labels` is provided.")
            use_cache = False
            if decoder_input_ids is None and decoder_inputs_embeds is None:
                decoder_input_ids = shift_tokens_right(labels, self.config.pad_token_id)

        outputs = self.model(
            input_ids,
            attention_mask=attention_mask,
            decoder_input_ids=decoder_input_ids,
            encoder_outputs=encoder_outputs,
            decoder_attention_mask=decoder_attention_mask,
            past_key_values=past_key_values,
            inputs_embeds=inputs_embeds,
            decoder_inputs_embeds=decoder_inputs_embeds,
            use_cache=use_cache,
            return_dict=return_dict,
            **kwargs,
        )
        lm_logits = self.lm_head(outputs[0]) + self.final_logits_bias

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

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

        return Seq2SeqLMOutput(
            loss=masked_lm_loss,
            logits=lm_logits,
            past_key_values=outputs.past_key_values,
            decoder_hidden_states=outputs.decoder_hidden_states,
            decoder_attentions=outputs.decoder_attentions,
            cross_attentions=outputs.cross_attentions,
            encoder_last_hidden_state=outputs.encoder_last_hidden_state,
            encoder_hidden_states=outputs.encoder_hidden_states,
            encoder_attentions=outputs.encoder_attentions,
        )

    def prepare_decoder_input_ids_from_labels(self, labels: torch.Tensor):
        return shift_tokens_right(labels, self.config.pad_token_id)


@auto_docstring(
    custom_intro="""
    MBart model with a sequence classification/head on top (a linear layer on top of the pooled output) e.g. for GLUE
    tasks.
    """
)
class MBartForSequenceClassification(MBartPreTrainedModel):
    def __init__(self, config: MBartConfig, **kwargs):
        super().__init__(config, **kwargs)
        self.model = MBartModel(config)
        self.classification_head = MBartClassificationHead(
            config.d_model,
            config.d_model,
            config.num_labels,
            config.classifier_dropout,
        )

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

    @can_return_tuple
    @auto_docstring
    # Copied from transformers.models.bart.modeling_bart.BartForSequenceClassification.forward
    def forward(
        self,
        input_ids: torch.LongTensor | None = None,
        attention_mask: torch.Tensor | None = None,
        decoder_input_ids: torch.LongTensor | None = None,
        decoder_attention_mask: torch.LongTensor | None = None,
        encoder_outputs: list[torch.FloatTensor] | None = None,
        inputs_embeds: torch.FloatTensor | None = None,
        decoder_inputs_embeds: torch.FloatTensor | None = None,
        labels: torch.LongTensor | None = None,
        use_cache: bool | None = None,
        **kwargs: Unpack[TransformersKwargs],
    ) -> tuple | Seq2SeqSequenceClassifierOutput:
        r"""
        decoder_input_ids (`torch.LongTensor` of shape `(batch_size, target_sequence_length)`, *optional*):
            Indices of decoder input sequence tokens in the vocabulary.

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

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

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

            For translation and summarization training, `decoder_input_ids` should be provided. If no
            `decoder_input_ids` is provided, the model will create this tensor by shifting the `input_ids` to the right
            for denoising pre-training following the paper.
        decoder_attention_mask (`torch.LongTensor` of shape `(batch_size, target_sequence_length)`, *optional*):
            Default behavior: generate a tensor that ignores pad tokens in `decoder_input_ids`. Causal mask will also
            be used by default.

            If you want to change padding behavior, you should read [`modeling_bart._prepare_decoder_attention_mask`]
            and modify to your needs. See diagram 1 in [the paper](https://huggingface.co/papers/1910.13461) for more
            information on the default strategy.
        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 classification loss is computed (Cross-Entropy).
        """
        if labels is not None:
            use_cache = False

        if input_ids is None and inputs_embeds is not None:
            raise NotImplementedError(
                f"Passing input embeddings is currently not supported for {self.__class__.__name__}"
            )

        outputs: Seq2SeqModelOutput = self.model(
            input_ids,
            attention_mask=attention_mask,
            decoder_input_ids=decoder_input_ids,
            decoder_attention_mask=decoder_attention_mask,
            encoder_outputs=encoder_outputs,
            inputs_embeds=inputs_embeds,
            decoder_inputs_embeds=decoder_inputs_embeds,
            use_cache=use_cache,
            **kwargs,
        )
        hidden_states = outputs[0]  # last hidden state

        eos_mask = input_ids.eq(self.config.eos_token_id).to(hidden_states.device)

        torch_compilable_check(
            torch.unique_consecutive(eos_mask.sum(1)).numel() == 1,
            "All examples must have the same number of <eos> tokens.",
        )
        sentence_representation = hidden_states[eos_mask, :].view(hidden_states.size(0), -1, hidden_states.size(-1))[
            :, -1, :
        ]
        logits = self.classification_head(sentence_representation)

        loss = None
        if labels is not None:
            labels = labels.to(logits.device)
            if self.config.problem_type is None:
                if self.config.num_labels == 1:
                    self.config.problem_type = "regression"
                elif self.config.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.config.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.config.num_labels), labels.view(-1))
            elif self.config.problem_type == "multi_label_classification":
                loss_fct = BCEWithLogitsLoss()
                loss = loss_fct(logits, labels)

        return Seq2SeqSequenceClassifierOutput(
            loss=loss,
            logits=logits,
            past_key_values=outputs.past_key_values,
            decoder_hidden_states=outputs.decoder_hidden_states,
            decoder_attentions=outputs.decoder_attentions,
            cross_attentions=outputs.cross_attentions,
            encoder_last_hidden_state=outputs.encoder_last_hidden_state,
            encoder_hidden_states=outputs.encoder_hidden_states,
            encoder_attentions=outputs.encoder_attentions,
        )


@auto_docstring
class MBartForQuestionAnswering(MBartPreTrainedModel):
    def __init__(self, config):
        super().__init__(config)

        config.num_labels = 2
        self.num_labels = config.num_labels

        self.model = MBartModel(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
    # Copied from transformers.models.bart.modeling_bart.BartForQuestionAnswering.forward
    def forward(
        self,
        input_ids: torch.Tensor | None = None,
        attention_mask: torch.Tensor | None = None,
        decoder_input_ids: torch.LongTensor | None = None,
        decoder_attention_mask: torch.LongTensor | None = None,
        encoder_outputs: list[torch.FloatTensor] | None = None,
        start_positions: torch.LongTensor | None = None,
        end_positions: torch.LongTensor | None = None,
        inputs_embeds: torch.FloatTensor | None = None,
        decoder_inputs_embeds: torch.FloatTensor | None = None,
        use_cache: bool | None = None,
        **kwargs: Unpack[TransformersKwargs],
    ) -> tuple | Seq2SeqQuestionAnsweringModelOutput:
        r"""
        decoder_input_ids (`torch.LongTensor` of shape `(batch_size, target_sequence_length)`, *optional*):
            Indices of decoder input sequence tokens in the vocabulary.

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

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

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

            For translation and summarization training, `decoder_input_ids` should be provided. If no
            `decoder_input_ids` is provided, the model will create this tensor by shifting the `input_ids` to the right
            for denoising pre-training following the paper.
        decoder_attention_mask (`torch.LongTensor` of shape `(batch_size, target_sequence_length)`, *optional*):
            Default behavior: generate a tensor that ignores pad tokens in `decoder_input_ids`. Causal mask will also
            be used by default.

            If you want to change padding behavior, you should read [`modeling_bart._prepare_decoder_attention_mask`]
            and modify to your needs. See diagram 1 in [the paper](https://huggingface.co/papers/1910.13461) for more
            information on the default strategy.
        """
        if start_positions is not None and end_positions is not None:
            use_cache = False

        outputs: Seq2SeqModelOutput = self.model(
            input_ids,
            attention_mask=attention_mask,
            decoder_input_ids=decoder_input_ids,
            decoder_attention_mask=decoder_attention_mask,
            encoder_outputs=encoder_outputs,
            inputs_embeds=inputs_embeds,
            decoder_inputs_embeds=decoder_inputs_embeds,
            use_cache=use_cache,
            **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 Seq2SeqQuestionAnsweringModelOutput(
            loss=total_loss,
            start_logits=start_logits,
            end_logits=end_logits,
            past_key_values=outputs.past_key_values,
            decoder_hidden_states=outputs.decoder_hidden_states,
            decoder_attentions=outputs.decoder_attentions,
            cross_attentions=outputs.cross_attentions,
            encoder_last_hidden_state=outputs.encoder_last_hidden_state,
            encoder_hidden_states=outputs.encoder_hidden_states,
            encoder_attentions=outputs.encoder_attentions,
        )


# Copied from transformers.models.bart.modeling_bart.BartDecoderWrapper with Bart->MBart
class MBartDecoderWrapper(MBartPreTrainedModel):
    """
    This wrapper class is a helper class to correctly load pretrained checkpoints when the causal language model is
    used in combination with the [`EncoderDecoderModel`] framework.
    """

    def __init__(self, config):
        super().__init__(config)
        self.decoder = MBartDecoder(config)
        self.post_init()

    def forward(self, *args, **kwargs):
        return self.decoder(*args, **kwargs)


# Copied from transformers.models.bart.modeling_bart.BartForCausalLM with Bart->MBart, facebook/bart-base->facebook/mbart-large-cc25
class MBartForCausalLM(MBartPreTrainedModel, GenerationMixin):
    _tied_weights_keys = {
        "lm_head.weight": "model.decoder.embed_tokens.weight",
    }

    def __init__(self, config):
        config.is_decoder = True
        config.is_encoder_decoder = False
        super().__init__(config)
        self.model = MBartDecoderWrapper(config)

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

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

    def get_input_embeddings(self):
        return self.model.decoder.embed_tokens

    def set_input_embeddings(self, value):
        self.model.decoder.embed_tokens = value

    @can_return_tuple
    @auto_docstring
    def forward(
        self,
        input_ids: torch.LongTensor | None = None,
        attention_mask: torch.Tensor | None = None,
        encoder_hidden_states: torch.FloatTensor | None = None,
        encoder_attention_mask: torch.FloatTensor | None = None,
        past_key_values: Cache | None = None,
        inputs_embeds: torch.FloatTensor | None = None,
        labels: torch.LongTensor | None = None,
        use_cache: bool | None = None,
        logits_to_keep: int | torch.Tensor = 0,
        **kwargs: Unpack[TransformersKwargs],
    ) -> tuple | CausalLMOutputWithCrossAttentions:
        r"""
        labels (`torch.LongTensor` of shape `(batch_size, sequence_length)`, *optional*):
            Labels for computing the masked language modeling loss. Indices should either be in `[0, ...,
            config.vocab_size]` or -100 (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, MBartForCausalLM

        >>> tokenizer = AutoTokenizer.from_pretrained("facebook/mbart-large-cc25")
        >>> model = MBartForCausalLM.from_pretrained("facebook/mbart-large-cc25")
        >>> assert model.config.is_decoder, f"{model.__class__} has to be configured as a decoder."
        >>> inputs = tokenizer("Hello, my dog is cute", return_tensors="pt")
        >>> outputs = model(**inputs)

        >>> logits = outputs.logits
        >>> expected_shape = [1, inputs.input_ids.shape[-1], model.config.vocab_size]
        >>> list(logits.shape) == expected_shape
        True
        ```"""

        outputs: BaseModelOutputWithPastAndCrossAttentions = self.model.decoder(
            input_ids=input_ids,
            attention_mask=attention_mask,
            encoder_hidden_states=encoder_hidden_states,
            encoder_attention_mask=encoder_attention_mask,
            past_key_values=past_key_values,
            inputs_embeds=inputs_embeds,
            use_cache=use_cache,
            **kwargs,
        )

        hidden_states = outputs[0]
        # Only compute necessary logits
        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:
            labels = labels.to(logits.device)
            loss_fct = CrossEntropyLoss()
            loss = loss_fct(logits.view(-1, self.config.vocab_size), labels.view(-1))

        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__ = [
    "MBartForCausalLM",
    "MBartForConditionalGeneration",
    "MBartForQuestionAnswering",
    "MBartForSequenceClassification",
    "MBartModel",
    "MBartPreTrainedModel",
]