# coding=utf-8
# 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 copy
import math
from typing import Callable, Optional, Union

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

from ...activations import ACT2FN
from ...cache_utils import Cache, EncoderDecoderCache
from ...generation import GenerationMixin
from ...modeling_attn_mask_utils import (
    AttentionMaskConverter,
    _prepare_4d_attention_mask,
    _prepare_4d_attention_mask_for_sdpa,
)
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 (
    auto_docstring,
    is_torch_flex_attn_available,
    is_torchdynamo_compiling,
    logging,
)
from .configuration_marian import MarianConfig


if is_torch_flex_attn_available():
    from torch.nn.attention.flex_attention import BlockMask

    from ...integrations.flex_attention import make_flex_block_causal_mask


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: Optional[int] = None) -> None:
        super().__init__(num_positions, embedding_dim)

    def _init_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]))
        self.weight = nn.Parameter(out, requires_grad=False)

    @torch.no_grad()
    def forward(
        self, input_ids_shape: torch.Size, past_key_values_length: int = 0, position_ids: Optional[torch.Tensor] = 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.bart.modeling_bart.eager_attention_forward
def eager_attention_forward(
    module: nn.Module,
    query: torch.Tensor,
    key: torch.Tensor,
    value: torch.Tensor,
    attention_mask: Optional[torch.Tensor],
    scaling: Optional[float] = None,
    dropout: float = 0.0,
    head_mask: Optional[torch.Tensor] = None,
    **kwargs,
):
    if scaling is None:
        scaling = query.size(-1) ** -0.5

    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)

    if head_mask is not None:
        attn_weights = attn_weights * head_mask.view(1, -1, 1, 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: Optional[MarianConfig] = None,
        layer_idx: Optional[int] = 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: Optional[torch.Tensor] = None,
        past_key_value: Optional[Cache] = None,
        attention_mask: Optional[torch.Tensor] = None,
        layer_head_mask: Optional[torch.Tensor] = None,
        output_attentions: bool = False,
        cache_position: Optional[torch.Tensor] = 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, Optional[torch.Tensor], Optional[tuple[torch.Tensor]]]:
        """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
        bsz, tgt_len = hidden_states.shape[:-1]
        src_len = key_value_states.shape[1] if is_cross_attention else tgt_len

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

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

        if past_key_value is not None:
            if isinstance(past_key_value, EncoderDecoderCache):
                is_updated = past_key_value.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_value = past_key_value.cross_attention_cache
                else:
                    curr_past_key_value = past_key_value.self_attention_cache
            else:
                curr_past_key_value = past_key_value

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

            if past_key_value is not None:
                # save all key/value_states to cache to be re-used for fast auto-regressive generation
                cache_position = cache_position if not is_cross_attention else None
                key_states, value_states = curr_past_key_value.update(
                    key_states, value_states, self.layer_idx, {"cache_position": cache_position}
                )
                # set flag that curr layer for cross-attn is already updated so we can re-use in subsequent calls
                if is_cross_attention:
                    past_key_value.is_updated[self.layer_idx] = True

        attention_interface: Callable = eager_attention_forward
        if self.config._attn_implementation != "eager":
            attention_interface = ALL_ATTENTION_FUNCTIONS[self.config._attn_implementation]

        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,
            output_attentions=output_attentions,
            head_mask=layer_head_mask,
            **kwargs,
        )

        attn_output = attn_output.reshape(bsz, tgt_len, -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: Optional[int] = 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,
        layer_head_mask: torch.FloatTensor,
        output_attentions: Optional[bool] = False,
    ) -> tuple[torch.FloatTensor, Optional[torch.FloatTensor]]:
        """
        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.
            layer_head_mask (`torch.FloatTensor`): mask for attention heads in a given layer of size
                `(encoder_attention_heads,)`.
            output_attentions (`bool`, *optional*):
                Whether or not to return the attentions tensors of all attention layers. See `attentions` under
                returned tensors for more detail.
        """
        residual = hidden_states
        hidden_states, attn_weights = self.self_attn(
            hidden_states=hidden_states,
            attention_mask=attention_mask,
            layer_head_mask=layer_head_mask,
            output_attentions=output_attentions,
        )
        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 (
            torch.isinf(hidden_states).any() or torch.isnan(hidden_states).any()
        ):
            clamp_value = torch.finfo(hidden_states.dtype).max - 1000
            hidden_states = torch.clamp(hidden_states, min=-clamp_value, max=clamp_value)

        outputs = (hidden_states,)

        if output_attentions:
            outputs += (attn_weights,)

        return outputs


# Copied from transformers.models.bart.modeling_bart.BartDecoderLayer with Bart->Marian, BART->MARIAN
class MarianDecoderLayer(GradientCheckpointingLayer):
    def __init__(self, config: MarianConfig, layer_idx: Optional[int] = 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: Optional[torch.Tensor] = None,
        encoder_hidden_states: Optional[torch.Tensor] = None,
        encoder_attention_mask: Optional[torch.Tensor] = None,
        layer_head_mask: Optional[torch.Tensor] = None,
        cross_attn_layer_head_mask: Optional[torch.Tensor] = None,
        past_key_value: Optional[Cache] = None,
        output_attentions: Optional[bool] = False,
        use_cache: Optional[bool] = True,
        cache_position: Optional[torch.Tensor] = None,
    ) -> tuple[torch.FloatTensor, Optional[tuple[torch.FloatTensor, torch.FloatTensor]]]:
        """
        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.
            layer_head_mask (`torch.FloatTensor`): mask for attention heads in a given layer of size
                `(encoder_attention_heads,)`.
            cross_attn_layer_head_mask (`torch.FloatTensor`): mask for cross-attention heads in a given layer of
                size `(decoder_attention_heads,)`.
            past_key_value (`Tuple(torch.FloatTensor)`): cached past key and value projection states
            output_attentions (`bool`, *optional*):
                Whether or not to return the attentions tensors of all attention layers. See `attentions` under
                returned tensors for more detail.
            cache_position (`torch.LongTensor` of shape `(sequence_length)`, *optional*):
                Indices depicting the position of the input sequence tokens in the sequence. It is used to update the
                cache in the correct position and to infer the complete sequence length.
        """
        residual = hidden_states

        # Self Attention
        hidden_states, self_attn_weights = self.self_attn(
            hidden_states=hidden_states,
            past_key_value=past_key_value,
            attention_mask=attention_mask,
            layer_head_mask=layer_head_mask,
            output_attentions=output_attentions,
            cache_position=cache_position,
        )
        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
        cross_attn_weights = None
        if encoder_hidden_states is not None:
            residual = hidden_states

            hidden_states, cross_attn_weights = self.encoder_attn(
                hidden_states=hidden_states,
                key_value_states=encoder_hidden_states,
                attention_mask=encoder_attention_mask,
                layer_head_mask=cross_attn_layer_head_mask,
                past_key_value=past_key_value,
                output_attentions=output_attentions,
                cache_position=cache_position,
            )
            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)

        outputs = (hidden_states,)

        if output_attentions:
            outputs += (self_attn_weights, cross_attn_weights)
        return outputs


@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

    def _init_weights(self, module: Union[nn.Linear, nn.Embedding, MarianSinusoidalPositionalEmbedding]):
        std = self.config.init_std
        if isinstance(module, nn.Linear):
            module.weight.data.normal_(mean=0.0, std=std)
            if module.bias is not None:
                module.bias.data.zero_()
        elif isinstance(module, MarianSinusoidalPositionalEmbedding):
            module._init_weight()
        elif isinstance(module, nn.Embedding):
            module.weight.data.normal_(mean=0.0, std=std)
            if module.padding_idx is not None:
                module.weight.data[module.padding_idx].zero_()
        elif isinstance(module, nn.LayerNorm):
            module.weight.data.fill_(1.0)
            module.bias.data.zero_()

    @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

    # Copied from transformers.models.bart.modeling_bart.BartPreTrainedModel._update_full_mask
    def _update_full_mask(
        self,
        attention_mask: Union[torch.Tensor, None],
        inputs_embeds: torch.Tensor,
    ):
        if attention_mask is not None:
            if self.config._attn_implementation == "flash_attention_2":
                attention_mask = attention_mask if 0 in attention_mask else None
            elif self.config._attn_implementation == "sdpa":
                # output_attentions=True & head_mask can not be supported when using SDPA, fall back to
                # the manual implementation that requires a 4D causal mask in all cases.
                # [bsz, seq_len] -> [bsz, 1, tgt_seq_len, src_seq_len]
                attention_mask = _prepare_4d_attention_mask_for_sdpa(attention_mask, inputs_embeds.dtype)
            elif self.config._attn_implementation == "flex_attention":
                if isinstance(attention_mask, torch.Tensor):
                    attention_mask = make_flex_block_causal_mask(attention_mask, is_causal=False)
            else:
                # [bsz, seq_len] -> [bsz, 1, tgt_seq_len, src_seq_len]
                attention_mask = _prepare_4d_attention_mask(attention_mask, inputs_embeds.dtype)

        return attention_mask

    # Copied from transformers.models.bart.modeling_bart.BartPreTrainedModel._update_causal_mask
    def _update_causal_mask(
        self,
        attention_mask: Optional[Union[torch.Tensor, "BlockMask"]],
        input_tensor: torch.Tensor,
        cache_position: torch.Tensor,
        past_key_values: Cache,
    ):
        if self.config._attn_implementation == "flex_attention":
            if isinstance(attention_mask, torch.Tensor):
                attention_mask = make_flex_block_causal_mask(attention_mask)
            # Other attention flavors support in-built causal (when `mask is None`)
            # while we need to create our specific block mask regardless
            elif attention_mask is None:
                attention_mask = make_flex_block_causal_mask(
                    torch.ones(
                        size=(input_tensor.shape[0], input_tensor.shape[1]),
                        device=attention_mask.device,
                    )
                )
            return attention_mask

        if self.config._attn_implementation == "flash_attention_2":
            if attention_mask is not None and (attention_mask == 0.0).any():
                return attention_mask
            return None

        # For SDPA, when possible, we will rely on its `is_causal` argument instead of its `attn_mask` argument, in
        # order to dispatch on Flash Attention 2. This feature is not compatible with static cache, as SDPA will fail
        # to infer the attention mask.
        past_seen_tokens = past_key_values.get_seq_length() if past_key_values is not None else 0
        using_compilable_cache = past_key_values.is_compileable if past_key_values is not None else False

        # When output attentions is True, sdpa implementation's forward method calls the eager implementation's forward
        if self.config._attn_implementation == "sdpa" and not using_compilable_cache:
            if AttentionMaskConverter._ignore_causal_mask_sdpa(
                attention_mask,
                inputs_embeds=input_tensor,
                past_key_values_length=past_seen_tokens,
                is_training=self.training,
            ):
                return None

        dtype = input_tensor.dtype
        sequence_length = input_tensor.shape[1]
        if using_compilable_cache:
            target_length = past_key_values.get_max_cache_shape()
        else:
            target_length = (
                attention_mask.shape[-1]
                if isinstance(attention_mask, torch.Tensor)
                else past_seen_tokens + sequence_length + 1
            )

        # In case the provided `attention` mask is 2D, we generate a causal mask here (4D).
        causal_mask = self._prepare_4d_causal_attention_mask_with_cache_position(
            attention_mask,
            sequence_length=sequence_length,
            target_length=target_length,
            dtype=dtype,
            cache_position=cache_position,
            batch_size=input_tensor.shape[0],
        )

        if (
            self.config._attn_implementation == "sdpa"
            and attention_mask is not None
            and attention_mask.device.type in ["cuda", "xpu", "npu"]
        ):
            # Attend to all tokens in fully masked rows in the causal_mask, for example the relevant first rows when
            # using left padding. This is required by F.scaled_dot_product_attention memory-efficient attention path.
            # Details: https://github.com/pytorch/pytorch/issues/110213
            min_dtype = torch.finfo(dtype).min
            causal_mask = AttentionMaskConverter._unmask_unattended(causal_mask, min_dtype)

        return causal_mask

    @staticmethod
    # Copied from transformers.models.gptj.modeling_gptj.GPTJModel._prepare_4d_causal_attention_mask_with_cache_position
    def _prepare_4d_causal_attention_mask_with_cache_position(
        attention_mask: torch.Tensor,
        sequence_length: int,
        target_length: int,
        dtype: torch.dtype,
        cache_position: torch.Tensor,
        batch_size: int,
        **kwargs,
    ):
        """
        Creates a causal 4D mask of shape `(batch_size, 1, query_length, key_value_length)` from a 2D mask of shape
        `(batch_size, key_value_length)`, or if the input `attention_mask` is already 4D, do nothing.

        Args:
            attention_mask (`torch.Tensor`):
                A 2D attention mask of shape `(batch_size, key_value_length)` or a 4D attention mask of shape
                `(batch_size, 1, query_length, key_value_length)`.
            sequence_length (`int`):
                The sequence length being processed.
            target_length (`int`):
                The target length: when generating with static cache, the mask should be as long as the static cache,
                to account for the 0 padding, the part of the cache that is not filled yet.
            dtype (`torch.dtype`):
                The dtype to use for the 4D attention mask.
            cache_position (`torch.Tensor`):
                Indices depicting the position of the input sequence tokens in the sequence.
            batch_size (`torch.Tensor`):
                Batch size.
        """
        if attention_mask is not None and attention_mask.dim() == 4:
            # In this case we assume that the mask comes already in inverted form and requires no inversion or slicing.
            causal_mask = attention_mask
        else:
            min_dtype = torch.finfo(dtype).min
            causal_mask = torch.full(
                (sequence_length, target_length), fill_value=min_dtype, dtype=dtype, device=cache_position.device
            )
            if sequence_length != 1:
                causal_mask = torch.triu(causal_mask, diagonal=1)
            causal_mask *= torch.arange(target_length, device=cache_position.device) > cache_position.reshape(-1, 1)
            causal_mask = causal_mask[None, None, :, :].expand(batch_size, 1, -1, -1)
            if attention_mask is not None:
                causal_mask = causal_mask.clone()  # copy to contiguous memory for in-place edit
                mask_length = attention_mask.shape[-1]
                padding_mask = causal_mask[:, :, :, :mask_length] + attention_mask[:, None, None, :].to(
                    causal_mask.device
                )
                padding_mask = padding_mask == 0
                causal_mask[:, :, :, :mask_length] = causal_mask[:, :, :, :mask_length].masked_fill(
                    padding_mask, min_dtype
                )

        return causal_mask

    # Copied from transformers.models.bart.modeling_bart.BartPreTrainedModel._update_cross_attn_mask
    def _update_cross_attn_mask(
        self,
        encoder_hidden_states: Union[torch.Tensor, None],
        encoder_attention_mask: Union[torch.Tensor, None],
        input_shape: torch.Size,
        inputs_embeds: torch.Tensor,
    ):
        # expand encoder attention mask
        if encoder_hidden_states is not None and encoder_attention_mask is not None:
            if self.config._attn_implementation == "flash_attention_2":
                encoder_attention_mask = encoder_attention_mask if 0 in encoder_attention_mask else None
            elif self.config._attn_implementation == "sdpa":
                # output_attentions=True & cross_attn_head_mask can not be supported when using SDPA, and we fall back on
                # the manual implementation that requires a 4D causal mask in all cases.
                # [bsz, seq_len] -> [bsz, 1, tgt_seq_len, src_seq_len]
                encoder_attention_mask = _prepare_4d_attention_mask_for_sdpa(
                    encoder_attention_mask,
                    inputs_embeds.dtype,
                    tgt_len=input_shape[-1],
                )
            elif self.config._attn_implementation == "flex_attention":
                if isinstance(encoder_attention_mask, torch.Tensor):
                    encoder_attention_mask = make_flex_block_causal_mask(
                        encoder_attention_mask,
                        query_length=input_shape[-1],
                        is_causal=False,
                    )
            else:
                # [bsz, seq_len] -> [bsz, 1, tgt_seq_len, src_seq_len]
                encoder_attention_mask = _prepare_4d_attention_mask(
                    encoder_attention_mask, inputs_embeds.dtype, tgt_len=input_shape[-1]
                )

        return encoder_attention_mask


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
    """

    def __init__(self, config: MarianConfig, embed_tokens: Optional[nn.Embedding] = None):
        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

        if embed_tokens is not None:
            self.embed_tokens = embed_tokens
        else:
            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()

    def forward(
        self,
        input_ids: Optional[torch.LongTensor] = None,
        attention_mask: Optional[torch.LongTensor] = None,
        head_mask: Optional[torch.Tensor] = None,
        inputs_embeds: Optional[torch.FloatTensor] = None,
        output_attentions: Optional[bool] = None,
        output_hidden_states: Optional[bool] = None,
        return_dict: Optional[bool] = None,
    ) -> Union[tuple[torch.Tensor], 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)
            head_mask (`torch.Tensor` of shape `(encoder_layers, encoder_attention_heads)`, *optional*):
                Mask to nullify selected heads of the attention modules. Mask values selected in `[0, 1]`:

                - 1 indicates the head is **not masked**,
                - 0 indicates the head is **masked**.

            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.
            output_attentions (`bool`, *optional*):
                Whether or not to return the attentions tensors of all attention layers. See `attentions` under
                returned tensors for more detail.
            output_hidden_states (`bool`, *optional*):
                Whether or not to return the hidden states of all layers. See `hidden_states` under returned tensors
                for more detail.
            return_dict (`bool`, *optional*):
                Whether or not to return a [`~utils.ModelOutput`] instead of a plain tuple.
        """
        output_attentions = output_attentions if output_attentions is not None else self.config.output_attentions
        output_hidden_states = (
            output_hidden_states if output_hidden_states is not None else self.config.output_hidden_states
        )
        return_dict = return_dict if return_dict is not None else self.config.use_return_dict

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

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

        attention_mask = self._update_full_mask(
            attention_mask,
            inputs_embeds,
        )

        encoder_states = () if output_hidden_states else None
        all_attentions = () if output_attentions else None

        # check if head_mask has a correct number of layers specified if desired
        if head_mask is not None:
            assert head_mask.size()[0] == (len(self.layers)), (
                f"The head_mask should be specified for {len(self.layers)} layers, but it is for {head_mask.size()[0]}."
            )
        for idx, encoder_layer in enumerate(self.layers):
            if output_hidden_states:
                encoder_states = encoder_states + (hidden_states,)
            # 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 to_drop:
                layer_outputs = (None, None)
            else:
                layer_outputs = encoder_layer(
                    hidden_states,
                    attention_mask,
                    layer_head_mask=(head_mask[idx] if head_mask is not None else None),
                    output_attentions=output_attentions,
                )

                hidden_states = layer_outputs[0]

            if output_attentions:
                all_attentions = all_attentions + (layer_outputs[1],)

        if output_hidden_states:
            encoder_states = encoder_states + (hidden_states,)

        if not return_dict:
            return tuple(v for v in [hidden_states, encoder_states, all_attentions] if v is not None)
        return BaseModelOutput(
            last_hidden_state=hidden_states, hidden_states=encoder_states, attentions=all_attentions
        )


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
    """

    def __init__(self, config: MarianConfig, embed_tokens: Optional[nn.Embedding] = None):
        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

        if embed_tokens is not None:
            self.embed_tokens = embed_tokens
        else:
            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()

    def forward(
        self,
        input_ids: Optional[torch.LongTensor] = None,
        attention_mask: Optional[torch.Tensor] = None,
        encoder_hidden_states: Optional[torch.FloatTensor] = None,
        encoder_attention_mask: Optional[torch.LongTensor] = None,
        head_mask: Optional[torch.Tensor] = None,
        cross_attn_head_mask: Optional[torch.Tensor] = None,
        past_key_values: Optional[Cache] = None,
        inputs_embeds: Optional[torch.FloatTensor] = None,
        use_cache: Optional[bool] = None,
        output_attentions: Optional[bool] = None,
        output_hidden_states: Optional[bool] = None,
        return_dict: Optional[bool] = None,
        cache_position: Optional[torch.Tensor] = None,
    ) -> Union[tuple[torch.Tensor], 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)
            head_mask (`torch.Tensor` of shape `(decoder_layers, decoder_attention_heads)`, *optional*):
                Mask to nullify selected heads of the attention modules. Mask values selected in `[0, 1]`:

                - 1 indicates the head is **not masked**,
                - 0 indicates the head is **masked**.

            cross_attn_head_mask (`torch.Tensor` of shape `(decoder_layers, decoder_attention_heads)`, *optional*):
                Mask to nullify selected heads of the cross-attention modules in the decoder to avoid performing
                cross-attention on hidden heads. Mask values selected in `[0, 1]`:

                - 1 indicates the head is **not masked**,
                - 0 indicates the head is **masked**.

            past_key_values (`Cache`, *optional*, returned when `use_cache=True` is passed or when `config.use_cache=True`):
                Tuple of `tuple(torch.FloatTensor)` of length `config.n_layers`, with each tuple having 2 tensors of
                shape `(batch_size, num_heads, sequence_length, embed_size_per_head)`) and 2 additional tensors of
                shape `(batch_size, num_heads, encoder_sequence_length, embed_size_per_head)`.

                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.
            output_attentions (`bool`, *optional*):
                Whether or not to return the attentions tensors of all attention layers. See `attentions` under
                returned tensors for more detail.
            output_hidden_states (`bool`, *optional*):
                Whether or not to return the hidden states of all layers. See `hidden_states` under returned tensors
                for more detail.
            return_dict (`bool`, *optional*):
                Whether or not to return a [`~utils.ModelOutput`] instead of a plain tuple.
            cache_position (`torch.LongTensor` of shape `(sequence_length)`, *optional*):
                Indices depicting the position of the input sequence tokens in the sequence. It is used to update the
                cache in the correct position and to infer the complete sequence length.
        """
        output_attentions = output_attentions if output_attentions is not None else self.config.output_attentions
        output_hidden_states = (
            output_hidden_states if output_hidden_states is not None else self.config.output_hidden_states
        )
        use_cache = use_cache if use_cache is not None else self.config.use_cache
        return_dict = return_dict if return_dict is not None else self.config.use_return_dict

        if self.gradient_checkpointing and self.training:
            if use_cache:
                logger.warning_once(
                    "`use_cache=True` is incompatible with gradient checkpointing. Setting `use_cache=False`..."
                )
                use_cache = False

        # retrieve input_ids and inputs_embeds
        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")
        elif input_ids is not None:
            input = input_ids
            input_shape = input.shape
            input_ids = input_ids.view(-1, input_shape[-1])
        elif inputs_embeds is not None:
            input_shape = inputs_embeds.size()[:-1]
            input = inputs_embeds[:, :, -1]
        else:
            raise ValueError("You have to specify either decoder_input_ids or decoder_inputs_embeds")

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

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

        # initialize `past_key_values`
        return_legacy_cache = False
        if use_cache and not isinstance(past_key_values, Cache):
            return_legacy_cache = True
            logger.warning_once(
                "Passing a tuple of `past_key_values` is deprecated and will be removed in Transformers v4.58.0. "
                "You should pass an instance of `EncoderDecoderCache` instead, e.g. "
                "`past_key_values=EncoderDecoderCache.from_legacy_cache(past_key_values)`."
            )
            past_key_values = EncoderDecoderCache.from_legacy_cache(past_key_values)

        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
        if cache_position is None:
            cache_position = torch.arange(
                past_key_values_length, past_key_values_length + seq_length, device=inputs_embeds.device
            )

        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 = self._update_causal_mask(
            attention_mask,
            inputs_embeds,
            cache_position,
            self_attn_cache,
        )
        encoder_attention_mask = self._update_cross_attn_mask(
            encoder_hidden_states,
            encoder_attention_mask,
            input_shape,
            inputs_embeds,
        )

        # embed positions
        position_ids = self.embed_positions(
            (batch_size, seq_length), past_key_values_length, position_ids=cache_position
        )
        hidden_states = inputs_embeds + position_ids
        hidden_states = nn.functional.dropout(hidden_states, p=self.dropout, training=self.training)

        # decoder layers
        all_hidden_states = () if output_hidden_states else None
        all_self_attns = () if output_attentions else None
        all_cross_attentions = () if (output_attentions and encoder_hidden_states is not None) else None

        # check if head_mask/cross_attn_head_mask has a correct number of layers specified if desired
        for attn_mask, mask_name in zip([head_mask, cross_attn_head_mask], ["head_mask", "cross_attn_head_mask"]):
            if attn_mask is not None:
                assert attn_mask.size()[0] == (len(self.layers)), (
                    f"The `{mask_name}` should be specified for {len(self.layers)} layers, but it is for"
                    f" {head_mask.size()[0]}."
                )
        for idx, decoder_layer in enumerate(self.layers):
            # add LayerDrop (see https://huggingface.co/papers/1909.11556 for description)
            if output_hidden_states:
                all_hidden_states += (hidden_states,)
            if self.training:
                dropout_probability = torch.rand([])
                if dropout_probability < self.layerdrop:
                    continue

            layer_outputs = decoder_layer(
                hidden_states,
                causal_mask,
                encoder_hidden_states,  # as a positional argument for gradient checkpointing
                encoder_attention_mask=encoder_attention_mask,
                layer_head_mask=(head_mask[idx] if head_mask is not None else None),
                cross_attn_layer_head_mask=(cross_attn_head_mask[idx] if cross_attn_head_mask is not None else None),
                past_key_value=past_key_values,
                output_attentions=output_attentions,
                use_cache=use_cache,
                cache_position=cache_position,
            )
            hidden_states = layer_outputs[0]

            if output_attentions:
                all_self_attns += (layer_outputs[1],)

                if encoder_hidden_states is not None:
                    all_cross_attentions += (layer_outputs[2],)

        # add hidden states from the last decoder layer
        if output_hidden_states:
            all_hidden_states += (hidden_states,)

        if return_legacy_cache:
            past_key_values = past_key_values.to_legacy_cache()

        if not return_dict:
            return tuple(
                v
                for v in [hidden_states, past_key_values, all_hidden_states, all_self_attns, all_cross_attentions]
                if v is not None
            )
        return BaseModelOutputWithPastAndCrossAttentions(
            last_hidden_state=hidden_states,
            past_key_values=past_key_values,
            hidden_states=all_hidden_states,
            attentions=all_self_attns,
            cross_attentions=all_cross_attentions,
        )


@auto_docstring
class MarianModel(MarianPreTrainedModel):
    _tied_weights_keys = ["encoder.embed_tokens.weight", "decoder.embed_tokens.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
        self.shared = nn.Embedding(vocab_size, config.d_model, padding_idx)
        if self.config.share_encoder_decoder_embeddings:
            encoder_embed_tokens = decoder_embed_tokens = self.shared
        else:
            # Since the embeddings are not shared, deepcopy the embeddings here for encoder
            # and decoder to make sure they are not tied.
            encoder_embed_tokens = copy.deepcopy(self.shared)
            decoder_embed_tokens = copy.deepcopy(self.shared)
            self.shared = None

        self.encoder = MarianEncoder(config, encoder_embed_tokens)
        self.decoder = MarianDecoder(config, decoder_embed_tokens)

        # 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 get_encoder(self):
        return self.encoder

    def get_decoder(self):
        return self.decoder

    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

    @auto_docstring
    def forward(
        self,
        input_ids: Optional[torch.LongTensor] = None,
        attention_mask: Optional[torch.Tensor] = None,
        decoder_input_ids: Optional[torch.LongTensor] = None,
        decoder_attention_mask: Optional[torch.Tensor] = None,
        head_mask: Optional[torch.Tensor] = None,
        decoder_head_mask: Optional[torch.Tensor] = None,
        cross_attn_head_mask: Optional[torch.Tensor] = None,
        encoder_outputs: Optional[Union[tuple[torch.Tensor], BaseModelOutput]] = None,
        past_key_values: Optional[Cache] = None,
        inputs_embeds: Optional[torch.FloatTensor] = None,
        decoder_inputs_embeds: Optional[torch.FloatTensor] = None,
        use_cache: Optional[bool] = None,
        output_attentions: Optional[bool] = None,
        output_hidden_states: Optional[bool] = None,
        return_dict: Optional[bool] = None,
        cache_position: Optional[torch.Tensor] = None,
    ) -> 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.
        cross_attn_head_mask (`torch.Tensor` of shape `(decoder_layers, decoder_attention_heads)`, *optional*):
            Mask to nullify selected heads of the cross-attention modules in the decoder. Mask values selected in `[0,
            1]`:

            - 1 indicates the head is **not masked**,
            - 0 indicates the head is **masked**.

        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]
        ```"""
        output_attentions = output_attentions if output_attentions is not None else self.config.output_attentions
        output_hidden_states = (
            output_hidden_states if output_hidden_states is not None else self.config.output_hidden_states
        )
        use_cache = use_cache if use_cache is not None else self.config.use_cache
        return_dict = return_dict if return_dict is not None else self.config.use_return_dict

        if encoder_outputs is None:
            encoder_outputs = self.encoder(
                input_ids=input_ids,
                attention_mask=attention_mask,
                head_mask=head_mask,
                inputs_embeds=inputs_embeds,
                output_attentions=output_attentions,
                output_hidden_states=output_hidden_states,
                return_dict=return_dict,
            )
        # 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_value, 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,
            head_mask=decoder_head_mask,
            cross_attn_head_mask=cross_attn_head_mask,
            past_key_values=past_key_values,
            inputs_embeds=decoder_inputs_embeds,
            use_cache=use_cache,
            output_attentions=output_attentions,
            output_hidden_states=output_hidden_states,
            return_dict=return_dict,
            cache_position=cache_position,
        )

        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 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",
        "encoder.embed_positions.weight",
        "decoder.embed_positions.weight",
    ]
    _keys_to_ignore_on_save = ["model.encoder.embed_positions.weight", "model.decoder.embed_positions.weight"]
    _tied_weights_keys = ["model.encoder.embed_tokens.weight", "model.decoder.embed_tokens.weight", "lm_head.weight"]

    def __init__(self, config: MarianConfig):
        super().__init__(config)
        self.model = MarianModel(config)

        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 get_encoder(self):
        return self.model.get_encoder()

    def get_decoder(self):
        return self.model.get_decoder()

    def resize_token_embeddings(
        self, new_num_tokens: int, pad_to_multiple_of: Optional[int] = 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

    def tie_weights(self):
        """
        Tie the weights between the input embeddings and the output embeddings.

        If the `torchscript` flag is set in the configuration, can't handle parameter sharing so we are cloning the
        weights instead.
        """
        output_embeddings = self.get_output_embeddings()
        if output_embeddings is not None and getattr(self.config, "tie_word_embeddings", True):
            # if embeddings are shared this will return shared embeddings otherwise decoder embed_tokens
            word_embeddings = self.get_decoder().get_input_embeddings()
            self._tie_or_clone_weights(output_embeddings, word_embeddings)

        if getattr(self.config, "is_encoder_decoder", False) and getattr(self.config, "tie_encoder_decoder", False):
            if hasattr(self, self.base_model_prefix):
                self = getattr(self, self.base_model_prefix)
            tied_weights = self._tie_encoder_decoder_weights(
                self.encoder, self.decoder, self.base_model_prefix, "encoder"
            )
            # Setting a dynamic variable instead of `_tied_weights_keys` because it's a class
            # attributed not an instance member, therefore modifying it will modify the entire class
            # Leading to issues on subsequent calls by different tests or subsequent calls.
            self._dynamic_tied_weights_keys = tied_weights

        for module in self.modules():
            if hasattr(module, "_tie_weights"):
                module._tie_weights()

    @auto_docstring
    def forward(
        self,
        input_ids: Optional[torch.LongTensor] = None,
        attention_mask: Optional[torch.Tensor] = None,
        decoder_input_ids: Optional[torch.LongTensor] = None,
        decoder_attention_mask: Optional[torch.Tensor] = None,
        head_mask: Optional[torch.Tensor] = None,
        decoder_head_mask: Optional[torch.Tensor] = None,
        cross_attn_head_mask: Optional[torch.Tensor] = None,
        encoder_outputs: Optional[Union[tuple[torch.Tensor], BaseModelOutput]] = None,
        past_key_values: Optional[Cache] = None,
        inputs_embeds: Optional[torch.FloatTensor] = None,
        decoder_inputs_embeds: Optional[torch.FloatTensor] = None,
        labels: Optional[torch.LongTensor] = None,
        use_cache: Optional[bool] = None,
        output_attentions: Optional[bool] = None,
        output_hidden_states: Optional[bool] = None,
        return_dict: Optional[bool] = None,
        cache_position: Optional[torch.Tensor] = None,
    ) -> 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.
        cross_attn_head_mask (`torch.Tensor` of shape `(decoder_layers, decoder_attention_heads)`, *optional*):
            Mask to nullify selected heads of the cross-attention modules in the decoder. Mask values selected in `[0,
            1]`:

            - 1 indicates the head is **not masked**,
            - 0 indicates the head is **masked**.
        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?"
        ```
        """
        return_dict = return_dict if return_dict is not None else self.config.use_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, self.config.decoder_start_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,
            head_mask=head_mask,
            decoder_head_mask=decoder_head_mask,
            cross_attn_head_mask=cross_attn_head_mask,
            past_key_values=past_key_values,
            inputs_embeds=inputs_embeds,
            decoder_inputs_embeds=decoder_inputs_embeds,
            use_cache=use_cache,
            output_attentions=output_attentions,
            output_hidden_states=output_hidden_states,
            return_dict=return_dict,
            cache_position=cache_position,
        )
        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))

        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, 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)

    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"]

    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

    def set_decoder(self, decoder):
        self.model.decoder = decoder

    def get_decoder(self):
        return self.model.decoder

    @auto_docstring
    def forward(
        self,
        input_ids: Optional[torch.LongTensor] = None,
        attention_mask: Optional[torch.Tensor] = None,
        encoder_hidden_states: Optional[torch.FloatTensor] = None,
        encoder_attention_mask: Optional[torch.FloatTensor] = None,
        head_mask: Optional[torch.Tensor] = None,
        cross_attn_head_mask: Optional[torch.Tensor] = None,
        past_key_values: Optional[Cache] = None,
        inputs_embeds: Optional[torch.FloatTensor] = None,
        labels: Optional[torch.LongTensor] = None,
        use_cache: Optional[bool] = None,
        output_attentions: Optional[bool] = None,
        output_hidden_states: Optional[bool] = None,
        return_dict: Optional[bool] = None,
        cache_position: Optional[torch.LongTensor] = None,
    ) -> Union[tuple, CausalLMOutputWithCrossAttentions]:
        r"""
        cross_attn_head_mask (`torch.Tensor` of shape `(decoder_layers, decoder_attention_heads)`, *optional*):
            Mask to nullify selected heads of the cross-attention modules. Mask values selected in `[0, 1]`:

            - 1 indicates the head is **not masked**,
            - 0 indicates the head is **masked**.
        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", add_cross_attention=False)
        >>> 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
        ```"""

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

        # decoder outputs consists of (dec_features, layer_state, dec_hidden, dec_attn)
        outputs = self.model.decoder(
            input_ids=input_ids,
            attention_mask=attention_mask,
            encoder_hidden_states=encoder_hidden_states,
            encoder_attention_mask=encoder_attention_mask,
            head_mask=head_mask,
            cross_attn_head_mask=cross_attn_head_mask,
            past_key_values=past_key_values,
            inputs_embeds=inputs_embeds,
            use_cache=use_cache,
            output_attentions=output_attentions,
            output_hidden_states=output_hidden_states,
            return_dict=return_dict,
            cache_position=cache_position,
        )

        logits = self.lm_head(outputs[0])

        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))

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

        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"]