# Copyright 2021 The Marian Team Authors 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 MarianMTModel model, ported from the Marian C++ repo."""

import math
from collections.abc import Callable

import numpy as np
import torch
from torch import nn
from torch.nn import CrossEntropyLoss

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,
)
from ...modeling_utils import ALL_ATTENTION_FUNCTIONS, PreTrainedModel
from ...processing_utils import Unpack
from ...utils import (
    TransformersKwargs,
    auto_docstring,
    can_return_tuple,
    is_torchdynamo_compiling,
    logging,
)
from ...utils.generic import merge_with_config_defaults
from ...utils.output_capturing import OutputRecorder, capture_outputs
from .configuration_marian import MarianConfig


logger = logging.get_logger(__name__)


# Copied from transformers.models.bart.modeling_bart.shift_tokens_right
def shift_tokens_right(input_ids: torch.Tensor, pad_token_id: int, decoder_start_token_id: int):
    """
    Shift input ids one token to the right.
    """
    shifted_input_ids = input_ids.new_zeros(input_ids.shape)
    shifted_input_ids[:, 1:] = input_ids[:, :-1].clone()
    shifted_input_ids[:, 0] = decoder_start_token_id

    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`
    shifted_input_ids.masked_fill_(shifted_input_ids == -100, pad_token_id)

    return shifted_input_ids


class MarianSinusoidalPositionalEmbedding(nn.Embedding):
    """This module produces sinusoidal positional embeddings of any length."""

    def __init__(self, num_positions: int, embedding_dim: int, padding_idx: int | None = None) -> None:
        super().__init__(num_positions, embedding_dim, _freeze=True)

    def create_weight(self):
        """
        Identical to the XLM create_sinusoidal_embeddings except features are not interleaved. The cos features are in
        the 2nd half of the vector. [dim // 2:]
        """
        n_pos, dim = self.weight.shape
        position_enc = np.array(
            [[pos / np.power(10000, 2 * (j // 2) / dim) for j in range(dim)] for pos in range(n_pos)]
        )
        out = torch.empty(n_pos, dim, dtype=self.weight.dtype, requires_grad=False)
        sentinel = dim // 2 if dim % 2 == 0 else (dim // 2) + 1
        out[:, 0:sentinel] = torch.FloatTensor(np.sin(position_enc[:, 0::2]))
        out[:, sentinel:] = torch.FloatTensor(np.cos(position_enc[:, 1::2]))
        return out

    @torch.no_grad()
    def forward(
        self, input_ids_shape: torch.Size, past_key_values_length: int = 0, position_ids: torch.Tensor | None = None
    ) -> torch.Tensor:
        """`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
            )
        return super().forward(position_ids)


# 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->Marian
class MarianAttention(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: MarianConfig | 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


# Copied from transformers.models.bart.modeling_bart.BartEncoderLayer with Bart->Marian, BART->MARIAN
class MarianEncoderLayer(GradientCheckpointingLayer):
    def __init__(self, config: MarianConfig, layer_idx: int | None = None):
        super().__init__()
        self.embed_dim = config.d_model

        self.self_attn = MarianAttention(
            embed_dim=self.embed_dim,
            num_heads=config.encoder_attention_heads,
            dropout=config.attention_dropout,
            config=config,
            layer_idx=layer_idx,
        )
        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.FloatTensor,
        attention_mask: torch.FloatTensor,
        **kwargs: Unpack[TransformersKwargs],
    ) -> torch.Tensor:
        residual = hidden_states
        hidden_states, _ = self.self_attn(
            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
        hidden_states = self.self_attn_layer_norm(hidden_states)

        residual = 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
        hidden_states = self.final_layer_norm(hidden_states)

        if hidden_states.dtype == torch.float16 and not torch.isfinite(hidden_states).all():
            clamp_value = torch.finfo(hidden_states.dtype).max - 1000
            hidden_states = torch.clamp(hidden_states, min=-clamp_value, max=clamp_value)

        return hidden_states


# Copied from transformers.models.bart.modeling_bart.BartDecoderLayer with Bart->Marian, BART->MARIAN
class MarianDecoderLayer(GradientCheckpointingLayer):
    def __init__(self, config: MarianConfig, layer_idx: int | None = None):
        super().__init__()
        self.embed_dim = config.d_model

        self.self_attn = MarianAttention(
            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 = MarianAttention(
            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:
        residual = hidden_states

        # Self Attention
        hidden_states, _ = self.self_attn(
            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
        hidden_states = self.self_attn_layer_norm(hidden_states)

        # Cross-Attention Block
        if encoder_hidden_states is not None:
            residual = hidden_states

            hidden_states, _ = self.encoder_attn(
                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
            hidden_states = self.encoder_attn_layer_norm(hidden_states)

        # Fully Connected
        residual = 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
        hidden_states = self.final_layer_norm(hidden_states)

        return hidden_states


@auto_docstring
class MarianPreTrainedModel(PreTrainedModel):
    config: MarianConfig
    base_model_prefix = "model"
    supports_gradient_checkpointing = True
    _supports_flash_attn = True
    _supports_sdpa = True
    _supports_flex_attn = True

    _can_compile_fullgraph = True

    @torch.no_grad()
    def _init_weights(self, module):
        super()._init_weights(module)
        if isinstance(module, MarianSinusoidalPositionalEmbedding):
            init.copy_(module.weight, module.create_weight())
        elif isinstance(module, MarianMTModel):
            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,
            "decoder_input_ids": input_ids,
        }
        return dummy_inputs


class MarianEncoder(MarianPreTrainedModel):
    """
    Transformer encoder consisting of *config.encoder_layers* self attention layers. Each layer is a
    [`MarianEncoderLayer`].

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

    _can_record_outputs = {
        "hidden_states": MarianEncoderLayer,
        "attentions": MarianAttention,
    }

    def __init__(self, config: MarianConfig):
        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
        self.embed_scale = math.sqrt(embed_dim) if config.scale_embedding else 1.0

        self.embed_tokens = nn.Embedding(config.vocab_size, embed_dim, self.padding_idx)

        self.embed_positions = MarianSinusoidalPositionalEmbedding(
            config.max_position_embeddings, embed_dim, self.padding_idx
        )
        self.layers = nn.ModuleList([MarianEncoderLayer(config) for _ in range(config.encoder_layers)])

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

    @merge_with_config_defaults
    @capture_outputs
    @auto_docstring
    def forward(
        self,
        input_ids: torch.LongTensor | None = None,
        attention_mask: torch.LongTensor | None = None,
        inputs_embeds: torch.FloatTensor | None = None,
        **kwargs: Unpack[TransformersKwargs],
    ) -> BaseModelOutput:
        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) * self.embed_scale

        embed_pos = self.embed_positions(inputs_embeds.shape[:-1])

        hidden_states = inputs_embeds + embed_pos
        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,
                )
        return BaseModelOutput(
            last_hidden_state=hidden_states,
        )


class MarianDecoder(MarianPreTrainedModel):
    """
    Transformer decoder consisting of *config.decoder_layers* layers. Each layer is a [`MarianDecoderLayer`]

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

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

    def __init__(self, config: MarianConfig):
        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
        self.embed_scale = math.sqrt(config.d_model) if config.scale_embedding else 1.0

        self.embed_tokens = nn.Embedding(config.decoder_vocab_size, config.d_model, self.padding_idx)

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

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

    @merge_with_config_defaults
    @capture_outputs
    @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.LongTensor | None = None,
        past_key_values: Cache | None = None,
        inputs_embeds: torch.FloatTensor | None = None,
        use_cache: bool | None = None,
        **kwargs: Unpack[TransformersKwargs],
    ) -> BaseModelOutputWithPastAndCrossAttentions:
        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)

        # Important to apply outside of the above `if`, in case user passes `embeds`
        inputs_embeds = inputs_embeds * self.embed_scale

        # 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(
            (batch_size, seq_length), past_key_values_length, position_ids=position_ids
        )
        hidden_states = inputs_embeds + position_ids
        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,
            )
        return BaseModelOutputWithPastAndCrossAttentions(
            last_hidden_state=hidden_states,
            past_key_values=past_key_values,
        )


@auto_docstring
class MarianModel(MarianPreTrainedModel):
    _keys_to_ignore_on_load_missing = [
        "model.encoder.embed_positions.weight",
        "model.decoder.embed_positions.weight",
    ]

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

        padding_idx, vocab_size = config.pad_token_id, config.vocab_size

        # We always use self.shared for token embeddings to ensure compatibility with all marian models
        if self.config.share_encoder_decoder_embeddings:
            self.shared = nn.Embedding(vocab_size, config.d_model, padding_idx)
            self._tied_weights_keys = {
                "decoder.embed_tokens.weight": "shared.weight",
                "encoder.embed_tokens.weight": "shared.weight",
            }
        else:
            self._tied_weights_keys = None

        self.encoder = MarianEncoder(config)
        self.decoder = MarianDecoder(config)

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

    def get_input_embeddings(self):
        # This will return shared embeddings if they are shared else specific to encoder.
        return self.get_encoder().get_input_embeddings()

    def set_input_embeddings(self, value):
        if self.config.share_encoder_decoder_embeddings:
            self.shared = value
            self.encoder.embed_tokens = self.shared
            self.decoder.embed_tokens = self.shared
        else:  # if not shared only set encoder embeedings
            self.encoder.embed_tokens = value

    def get_decoder_input_embeddings(self):
        if self.config.share_encoder_decoder_embeddings:
            raise ValueError(
                "`get_decoder_input_embeddings` should not be called if `config.share_encoder_decoder_embeddings` "
                "is `True`. Please use `get_input_embeddings` instead."
            )
        return self.get_decoder().get_input_embeddings()

    def set_decoder_input_embeddings(self, value):
        if self.config.share_encoder_decoder_embeddings:
            raise ValueError(
                "`config.share_encoder_decoder_embeddings` is set to `True` meaning the decoder input embeddings "
                "are shared with the encoder. In order to set the decoder input embeddings, you should simply set "
                "the encoder input embeddings by calling `set_input_embeddings` with the appropriate embeddings."
            )
        self.decoder.embed_tokens = value

    def resize_decoder_token_embeddings(self, new_num_tokens: int) -> nn.Embedding:
        if self.config.share_encoder_decoder_embeddings:
            raise ValueError(
                "`resize_decoder_token_embeddings` should not be called if `config.share_encoder_decoder_embeddings` "
                "is `True`. Please use `resize_token_embeddings` instead."
            )

        old_embeddings = self.get_decoder_input_embeddings()
        new_embeddings = self._get_resized_embeddings(old_embeddings, new_num_tokens)
        self.set_decoder_input_embeddings(new_embeddings)

        model_embeds = self.get_decoder_input_embeddings()

        if new_num_tokens is None:
            return model_embeds

        # Update base model and current model config
        self.config.decoder_vocab_size = new_num_tokens

        # Tie weights again if needed
        self.tie_weights()

        return model_embeds

    @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.Tensor | None = None,
        encoder_outputs: tuple[torch.Tensor] | BaseModelOutput | 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,
        **kwargs: Unpack[TransformersKwargs],
    ) -> Seq2SeqModelOutput:
        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)

            Marian uses the `pad_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`).
        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.

        Example:

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

        >>> tokenizer = AutoTokenizer.from_pretrained("Helsinki-NLP/opus-mt-en-de")
        >>> model = MarianModel.from_pretrained("Helsinki-NLP/opus-mt-en-de")

        >>> inputs = tokenizer("Studies have been shown that owning a dog is good for you", return_tensors="pt")
        >>> decoder_inputs = tokenizer(
        ...     "<pad> Studien haben gezeigt dass es hilfreich ist einen Hund zu besitzen",
        ...     return_tensors="pt",
        ...     add_special_tokens=False,
        ... )
        >>> outputs = model(input_ids=inputs.input_ids, decoder_input_ids=decoder_inputs.input_ids)

        >>> last_hidden_states = outputs.last_hidden_state
        >>> list(last_hidden_states.shape)
        [1, 26, 512]
        ```"""
        # If encoder_outputs are not given, pass the inputs to the encoder
        if encoder_outputs is None:
            encoder_outputs = self.encoder(
                input_ids=input_ids,
                attention_mask=attention_mask,
                inputs_embeds=inputs_embeds,
                **kwargs,
            )
        # If the user passed a tuple for encoder_outputs, we wrap it in a BaseModelOutput
        elif 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,
            **kwargs,
        )

        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 Marian Model with a language modeling head. Can be used for summarization.
    """
)
class MarianMTModel(MarianPreTrainedModel, GenerationMixin):
    base_model_prefix = "model"
    _keys_to_ignore_on_load_missing = [
        "final_logits_bias",
        "model.encoder.embed_positions.weight",
        "model.decoder.embed_positions.weight",
    ]
    _keys_to_ignore_on_save = ["model.encoder.embed_positions.weight", "model.decoder.embed_positions.weight"]
    _tied_weights_keys = {"lm_head.weight": "model.decoder.embed_tokens.weight"}

    def __init__(self, config: MarianConfig):
        super().__init__(config)
        self.model = MarianModel(config)
        if self.config.share_encoder_decoder_embeddings:
            self._tied_weights_keys = {
                "lm_head.weight": "model.shared.weight",
                "model.decoder.embed_tokens.weight": "model.shared.weight",
                "model.encoder.embed_tokens.weight": "model.shared.weight",
            }

        target_vocab_size = config.vocab_size if config.share_encoder_decoder_embeddings else config.decoder_vocab_size
        self.register_buffer("final_logits_bias", torch.zeros((1, target_vocab_size)))
        self.lm_head = nn.Linear(config.d_model, target_vocab_size, 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)
        if self.config.share_encoder_decoder_embeddings:
            self._resize_final_logits_bias(new_num_tokens)
        return new_embeddings

    # NOTE: `_resize_token_embeddings` was rewritten in the base class, *args exists to absorb the extra arg
    def _resize_token_embeddings(self, new_num_tokens: int, pad_to_multiple_of=None, *args) -> nn.Embedding:
        old_embeddings = self.get_input_embeddings()
        new_embeddings = self._get_resized_embeddings(old_embeddings, new_num_tokens, pad_to_multiple_of)
        self.set_input_embeddings(new_embeddings)

        new_num_tokens = new_embeddings.weight.shape[0]
        # update config.decoder_vocab_size if embeddings are tied
        if self.config.share_encoder_decoder_embeddings:
            self.config.decoder_vocab_size = new_num_tokens

        # if word embeddings are not tied, make sure that lm head is resized as well
        if (
            self.config.share_encoder_decoder_embeddings
            and self.get_output_embeddings() is not None
            and not self.config.tie_word_embeddings
        ):
            old_lm_head = self.get_output_embeddings()
            new_lm_head = self._get_resized_lm_head(old_lm_head, new_num_tokens)
            self.set_output_embeddings(new_lm_head)

        return self.get_input_embeddings()

    def resize_decoder_token_embeddings(self, new_num_tokens):
        if self.config.share_encoder_decoder_embeddings:
            raise ValueError(
                "`resize_decoder_token_embeddings` should not be called if `config.share_encoder_decoder_embeddings` "
                "is `True`. Please use `resize_token_embeddings` instead."
            )

        old_embeddings = self.model.get_decoder_input_embeddings()
        new_embeddings = self._get_resized_embeddings(old_embeddings, new_num_tokens)
        self.model.set_decoder_input_embeddings(new_embeddings)

        # if word embeddings are not tied, make sure that lm head is resized as well
        if self.get_output_embeddings() is not None and not self.config.tie_word_embeddings:
            old_lm_head = self.get_output_embeddings()
            new_lm_head = self._get_resized_lm_head(old_lm_head, new_num_tokens)
            self.set_output_embeddings(new_lm_head)

        model_embeds = self.model.get_decoder_input_embeddings()

        if new_num_tokens is None:
            return model_embeds

        # Update base model and current model config
        self.config.decoder_vocab_size = new_num_tokens

        # Tie weights again if needed
        self.tie_weights()

        self._resize_final_logits_bias(new_num_tokens)

        return model_embeds

    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)

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

    @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.Tensor | None = None,
        encoder_outputs: tuple[torch.Tensor] | BaseModelOutput | 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,
        **kwargs: Unpack[TransformersKwargs],
    ) -> Seq2SeqLMOutput:
        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)

            Marian uses the `pad_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`).
        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:

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

        >>> src = "fr"  # source language
        >>> trg = "en"  # target language

        >>> model_name = f"Helsinki-NLP/opus-mt-{src}-{trg}"
        >>> model = MarianMTModel.from_pretrained(model_name)
        >>> tokenizer = AutoTokenizer.from_pretrained(model_name)

        >>> sample_text = "où est l'arrêt de bus ?"
        >>> batch = tokenizer([sample_text], return_tensors="pt")

        >>> generated_ids = model.generate(**batch)
        >>> tokenizer.batch_decode(generated_ids, skip_special_tokens=True)[0]
        "Where's the bus stop?"
        ```
        """
        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, self.config.decoder_start_token_id
                )

        outputs: Seq2SeqModelOutput = 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,
            **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.decoder_vocab_size), labels.view(-1))

        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, self.config.decoder_start_token_id)


# Copied from transformers.models.bart.modeling_bart.BartDecoderWrapper with Bart->Marian
class MarianDecoderWrapper(MarianPreTrainedModel):
    """
    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 = MarianDecoder(config)
        self.post_init()

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


# Copied from transformers.models.bart.modeling_bart.BartForCausalLM with Bart->Marian, facebook/bart-base->Helsinki-NLP/opus-mt-fr-en
class MarianForCausalLM(MarianPreTrainedModel, 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 = MarianDecoderWrapper(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, MarianForCausalLM

        >>> tokenizer = AutoTokenizer.from_pretrained("Helsinki-NLP/opus-mt-fr-en")
        >>> model = MarianForCausalLM.from_pretrained("Helsinki-NLP/opus-mt-fr-en")
        >>> 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__ = ["MarianForCausalLM", "MarianModel", "MarianMTModel", "MarianPreTrainedModel"]