# Copyright 2021 The Fairseq 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 TrOCR decoder model (based on RoBERTa)."""

import math

import torch
from torch import nn
from torch.nn import CrossEntropyLoss

from ...activations import ACT2FN
from ...cache_utils import Cache, DynamicCache, EncoderDecoderCache
from ...generation import GenerationMixin
from ...modeling_attn_mask_utils import (
    _prepare_4d_attention_mask,
    _prepare_4d_causal_attention_mask,
)
from ...modeling_layers import GradientCheckpointingLayer
from ...modeling_outputs import BaseModelOutputWithPastAndCrossAttentions, CausalLMOutputWithCrossAttentions
from ...modeling_utils import PreTrainedModel
from ...utils import auto_docstring, logging
from .configuration_trocr import TrOCRConfig


logger = logging.get_logger(__name__)


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

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

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

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

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


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

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

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


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

    def __init__(self, num_positions: int, embedding_dim: int, padding_idx: int | None = None):
        super().__init__()
        self.offset = 2
        self.embedding_dim = embedding_dim
        self.padding_idx = padding_idx
        self.weights = self.get_embedding(num_positions, embedding_dim, padding_idx)

    @staticmethod
    def get_embedding(num_embeddings: int, embedding_dim: int, padding_idx: int | None = None):
        """
        Build sinusoidal embeddings. This matches the implementation in tensor2tensor, but differs slightly from the
        description in Section 3.5 of "Attention Is All You Need".
        """
        half_dim = embedding_dim // 2
        emb = math.log(10000) / (half_dim - 1)
        emb = torch.exp(torch.arange(half_dim, dtype=torch.int64).float() * -emb)
        emb = torch.arange(num_embeddings, dtype=torch.int64).float().unsqueeze(1) * emb.unsqueeze(0)
        emb = torch.cat([torch.sin(emb), torch.cos(emb)], dim=1).view(num_embeddings, -1)
        if embedding_dim % 2 == 1:
            # zero pad
            emb = torch.cat([emb, torch.zeros(num_embeddings, 1)], dim=1)
        if padding_idx is not None:
            emb[padding_idx, :] = 0

        return emb.to(torch.get_default_dtype())

    @torch.no_grad()
    def forward(self, input_ids: torch.Tensor, past_key_values_length: int = 0):
        bsz, seq_len = input_ids.size()
        # Create the position ids from the input token ids. Any padded tokens remain padded.
        position_ids = self.create_position_ids_from_input_ids(input_ids, self.padding_idx, past_key_values_length).to(
            input_ids.device
        )

        # expand embeddings if needed
        max_pos = self.padding_idx + 1 + seq_len
        if self.weights is None or max_pos > self.weights.size(0):
            # recompute/expand embeddings if needed
            self.weights = self.get_embedding(max_pos, self.embedding_dim, self.padding_idx)

        x = self.weights.index_select(0, position_ids.view(-1)).view(bsz, seq_len, -1).detach()

        return x

    def create_position_ids_from_input_ids(
        self, input_ids: torch.Tensor, padding_idx: int, past_key_values_length: int | None = 0
    ):
        """
        Replace non-padding symbols with their position numbers. Position numbers begin at padding_idx+1. Padding
        symbols are ignored. This is modified from fairseq's `utils.make_positions`.
        """
        # The series of casts and type-conversions here are carefully balanced to both work with ONNX export and XLA.
        mask = input_ids.ne(padding_idx).int()
        incremental_indices = (torch.cumsum(mask, dim=1).type_as(mask) + past_key_values_length) * mask
        return incremental_indices.long() + padding_idx


class TrOCRAttention(nn.Module):
    """Multi-headed attention from 'Attention Is All You Need' paper."""

    def __init__(
        self,
        config,
        embed_dim: int,
        num_heads: int,
        kdim: int | None = None,
        vdim: int | None = None,
        dropout: float | None = 0.0,
        is_decoder: bool | None = False,
        bias: bool | None = True,
        is_cross_attention: bool | None = False,
        layer_idx: bool | None = None,
    ):
        super().__init__()
        self.embed_dim = embed_dim
        self.kdim = kdim if kdim is not None else embed_dim
        self.vdim = vdim if vdim is not None else embed_dim
        self.num_heads = num_heads
        self.dropout = dropout
        self.head_dim = embed_dim // num_heads
        if not (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} and `num_heads`:"
                f" {num_heads})."
            )
        self.scaling = self.head_dim**-0.5
        self.is_decoder = is_decoder
        self.layer_idx = layer_idx

        self.k_proj = nn.Linear(self.kdim, embed_dim, bias=bias)
        self.v_proj = nn.Linear(self.vdim, 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,
        output_attentions: bool | None = False,
        cache_position: torch.Tensor | None = None,
    ) -> tuple[torch.Tensor, torch.Tensor | None, tuple[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
        bsz, tgt_len, embed_dim = hidden_states.size()

        # get query proj
        query_states = self.q_proj(hidden_states) * self.scaling

        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)
            key_states = key_states.view(bsz, -1, self.num_heads, self.head_dim).transpose(1, 2)
            value_states = value_states.view(bsz, -1, self.num_heads, self.head_dim).transpose(1, 2)

            if past_key_values 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_values.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 and isinstance(past_key_values, EncoderDecoderCache):
                    past_key_values.is_updated[self.layer_idx] = True

        proj_shape = (bsz * self.num_heads, -1, self.head_dim)
        query_states = query_states.view(bsz, tgt_len, self.num_heads, self.head_dim).transpose(1, 2)
        query_states = query_states.reshape(*proj_shape)
        key_states = key_states.reshape(*proj_shape)
        value_states = value_states.reshape(*proj_shape)

        src_len = key_states.size(1)
        attn_weights = torch.bmm(query_states, key_states.transpose(1, 2))

        if attn_weights.size() != (bsz * self.num_heads, tgt_len, src_len):
            raise ValueError(
                f"Attention weights should be of size {(bsz * self.num_heads, tgt_len, src_len)}, but is"
                f" {attn_weights.size()}"
            )

        if attention_mask is not None:
            if attention_mask.size() != (bsz, 1, tgt_len, src_len):
                raise ValueError(
                    f"Attention mask should be of size {(bsz, 1, tgt_len, src_len)}, but is {attention_mask.size()}"
                )
            attn_weights = attn_weights.view(bsz, self.num_heads, tgt_len, src_len) + attention_mask
            attn_weights = attn_weights.view(bsz * self.num_heads, tgt_len, src_len)

        attn_weights = nn.functional.softmax(attn_weights, dim=-1)

        if output_attentions:
            # this operation is a bit awkward, but it's required to
            # make sure that attn_weights keeps its gradient.
            # In order to do so, attn_weights have to be reshaped
            # twice and have to be reused in the following
            attn_weights_reshaped = attn_weights.view(bsz, self.num_heads, tgt_len, src_len)
            attn_weights = attn_weights_reshaped.view(bsz * self.num_heads, tgt_len, src_len)
        else:
            attn_weights_reshaped = None

        attn_probs = nn.functional.dropout(attn_weights, p=self.dropout, training=self.training)

        attn_output = torch.bmm(attn_probs, value_states)

        if attn_output.size() != (bsz * self.num_heads, tgt_len, self.head_dim):
            raise ValueError(
                f"`attn_output` should be of size {(bsz, self.num_heads, tgt_len, self.head_dim)}, but is"
                f" {attn_output.size()}"
            )

        attn_output = attn_output.view(bsz, self.num_heads, tgt_len, self.head_dim)
        attn_output = attn_output.transpose(1, 2)
        attn_output = attn_output.reshape(bsz, tgt_len, embed_dim)

        attn_output = self.out_proj(attn_output)

        return attn_output, attn_weights_reshaped


class TrOCRDecoderLayer(GradientCheckpointingLayer):
    def __init__(self, config: TrOCRConfig, layer_idx=None):
        super().__init__()
        self.embed_dim = config.hidden_size

        self.self_attn = TrOCRAttention(
            config,
            embed_dim=self.embed_dim,
            num_heads=config.decoder_attention_heads,
            dropout=config.attention_dropout,
            is_decoder=True,
            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)

        if config.is_decoder:
            self.encoder_attn = TrOCRAttention(
                config,
                embed_dim=self.embed_dim,
                num_heads=config.decoder_attention_heads,
                kdim=config.cross_attention_hidden_size,
                vdim=config.cross_attention_hidden_size,
                dropout=config.attention_dropout,
                is_decoder=True,
                is_cross_attention=True,
                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,
        output_attentions: bool | None = False,
        use_cache: bool | None = True,
        cache_position: torch.Tensor | None = None,
    ):
        """
        Args:
            hidden_states (`torch.FloatTensor`): input to the layer of shape `(batch, seq_len, embed_dim)`
            attention_mask (`torch.FloatTensor`): attention mask of size
                `(batch, 1, tgt_len, src_len)` where padding elements are indicated by very large negative values.
            encoder_hidden_states (`torch.FloatTensor`):
                cross attention input to the layer of shape `(batch, seq_len, embed_dim)`
            encoder_attention_mask (`torch.FloatTensor`): encoder attention mask of size
                `(batch, 1, tgt_len, src_len)` where padding elements are indicated by very large negative values.
            past_key_values (`Cache`): cached past key and value projection states
            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

        # Self Attention
        hidden_states, self_attn_weights = self.self_attn(
            hidden_states=hidden_states,
            past_key_values=past_key_values,
            attention_mask=attention_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,
                past_key_values=past_key_values,
                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 TrOCRPreTrainedModel(PreTrainedModel):
    config: TrOCRConfig
    base_model_prefix = "model"
    supports_gradient_checkpointing = True
    _no_split_modules = ["TrOCRDecoderLayer"]


class TrOCRDecoder(TrOCRPreTrainedModel):
    """
    Transformer decoder consisting of *config.decoder_layers* layers. Each layer is a [`TrOCRDecoderLayer`]

    Args:
        config: TrOCRConfig
    """

    def __init__(self, config: TrOCRConfig):
        super().__init__(config)
        self.dropout = config.dropout
        self.layerdrop = config.decoder_layerdrop
        self.padding_idx = config.pad_token_id
        embed_scale = math.sqrt(config.hidden_size) if config.scale_embedding else 1.0

        self.embed_tokens = TrOCRScaledWordEmbedding(
            config.vocab_size, config.hidden_size, self.padding_idx, embed_scale=embed_scale
        )

        if config.use_learned_position_embeddings:
            self.embed_positions = TrOCRLearnedPositionalEmbedding(config.max_position_embeddings, config.hidden_size)
        else:
            self.embed_positions = TrOCRSinusoidalPositionalEmbedding(
                config.max_position_embeddings + self.padding_idx + 1,
                config.hidden_size,
                self.padding_idx,
            )

        if config.layernorm_embedding:
            self.layernorm_embedding = nn.LayerNorm(config.hidden_size)
        else:
            self.layernorm_embedding = None

        self.layers = nn.ModuleList([TrOCRDecoderLayer(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=None,
        attention_mask=None,
        encoder_hidden_states=None,
        encoder_attention_mask=None,
        past_key_values=None,
        inputs_embeds=None,
        use_cache=None,
        output_attentions=None,
        output_hidden_states=None,
        return_dict=None,
        cache_position=None,
        **kwargs,
    ):
        r"""
        Args:
            input_ids (`torch.LongTensor` of shape `(batch_size, sequence_length)`):
                Indices of input sequence tokens in the vocabulary. Padding will be ignored by default should you
                provide it.

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

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

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

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

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

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

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

                If `past_key_values` are used, the user can optionally input only the last `decoder_input_ids` (those
                that don't have their past key value states given to this model) of shape `(batch_size, 1)` instead of
                all `decoder_input_ids` of shape `(batch_size, sequence_length)`.
            inputs_embeds (`torch.FloatTensor` of shape `(batch_size, sequence_length, hidden_size)`, *optional*):
                Optionally, instead of passing `input_ids` you can choose to directly pass an embedded representation.
                This is useful if you want more control over how to convert `input_ids` indices into associated vectors
                than the model's internal embedding lookup matrix.
            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
        )
        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

        # retrieve input_ids and inputs_embeds
        if input_ids is not None and 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_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 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

        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)
            )

        past_key_values_length = past_key_values.get_seq_length() if past_key_values is not None else 0

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

        if self.config.use_learned_position_embeddings:
            embed_pos = self.embed_positions(input, past_key_values_length=past_key_values_length)
        else:
            embed_pos = self.embed_positions(input_ids, past_key_values_length=past_key_values_length)

        hidden_states = inputs_embeds + embed_pos

        if self.layernorm_embedding is not None:
            hidden_states = self.layernorm_embedding(hidden_states)

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

        input_shape = input.shape

        attention_mask = _prepare_4d_causal_attention_mask(
            attention_mask, input_shape, inputs_embeds, past_key_values_length
        )

        # expand encoder attention mask
        if encoder_hidden_states is not None and encoder_attention_mask is not None:
            # [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]
            )

        # 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

        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,
                attention_mask,
                encoder_hidden_states,  # as a positional argument for gradient checkpointing
                encoder_attention_mask=encoder_attention_mask,
                past_key_values=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 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(
    custom_intro="""
    The TrOCR Model with a language modeling head. Can be used for summarization.
    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.
    """
)
class TrOCRDecoderWrapper(TrOCRPreTrainedModel):
    def __init__(self, config):
        super().__init__(config)
        self.decoder = TrOCRDecoder(config)
        self.post_init()

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


@auto_docstring(
    custom_intro="""
    The TrOCR Decoder with a language modeling head. Can be used as the decoder part of [`EncoderDecoderModel`] and
    """
)
class TrOCRForCausalLM(TrOCRPreTrainedModel, GenerationMixin):
    _tied_weights_keys = {"output_projection.weight": "model.decoder.embed_tokens.weight"}

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

        self.output_projection = 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 get_output_embeddings(self):
        return self.output_projection

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

    @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,
        labels: torch.LongTensor | None = None,
        use_cache: bool | None = None,
        output_attentions: bool | None = None,
        output_hidden_states: bool | None = None,
        return_dict: bool | None = None,
        cache_position: torch.Tensor | None = None,
        **kwargs,
    ) -> 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 (
        ...     TrOCRConfig,
        ...     TrOCRProcessor,
        ...     TrOCRForCausalLM,
        ...     ViTConfig,
        ...     ViTModel,
        ...     VisionEncoderDecoderModel,
        ... )
        >>> import httpx
        >>> from io import BytesIO
        >>> from PIL import Image

        >>> # TrOCR is a decoder model and should be used within a VisionEncoderDecoderModel
        >>> # init vision2text model with random weights
        >>> encoder = ViTModel(ViTConfig())
        >>> decoder = TrOCRForCausalLM(TrOCRConfig())
        >>> model = VisionEncoderDecoderModel(encoder=encoder, decoder=decoder)

        >>> # If you want to start from the pretrained model, load the checkpoint with `VisionEncoderDecoderModel`
        >>> processor = TrOCRProcessor.from_pretrained("microsoft/trocr-base-handwritten")
        >>> model = VisionEncoderDecoderModel.from_pretrained("microsoft/trocr-base-handwritten")

        >>> # load image from the IAM dataset
        >>> url = "https://fki.tic.heia-fr.ch/static/img/a01-122-02.jpg"
        >>> with httpx.stream("GET", url) as response:
        ...     image = Image.open(BytesIO(response.read())).convert("RGB")
        >>> pixel_values = processor(image, return_tensors="pt").pixel_values
        >>> text = "industry, ' Mr. Brown commented icily. ' Let us have a"

        >>> # training
        >>> model.config.decoder_start_token_id = processor.tokenizer.eos_token_id
        >>> model.config.pad_token_id = processor.tokenizer.pad_token_id
        >>> model.config.vocab_size = model.config.decoder.vocab_size

        >>> labels = processor.tokenizer(text, return_tensors="pt").input_ids
        >>> outputs = model(pixel_values, labels=labels)
        >>> loss = outputs.loss
        >>> round(loss.item(), 2)
        5.30

        >>> # inference
        >>> generated_ids = model.generate(pixel_values)
        >>> generated_text = processor.batch_decode(generated_ids, skip_special_tokens=True)[0]
        >>> generated_text
        'industry, " Mr. Brown commented icily. " Let us have a'
        ```"""

        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,
            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.output_projection(outputs[0])

        loss = None
        if labels is not None:
            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__ = ["TrOCRForCausalLM", "TrOCRPreTrainedModel"]