# coding=utf-8
# Copyright 2023 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 VipLlava model."""

from dataclasses import dataclass
from typing import List, Optional, Tuple, Union

import torch
import torch.utils.checkpoint
from torch import nn

from ...activations import ACT2FN
from ...generation import GenerationMixin
from ...modeling_outputs import ModelOutput
from ...modeling_utils import PreTrainedModel
from ...utils import (
    add_start_docstrings,
    add_start_docstrings_to_model_forward,
    logging,
    replace_return_docstrings,
)
from ..auto import AutoModel, AutoModelForCausalLM
from .configuration_vipllava import VipLlavaConfig


logger = logging.get_logger(__name__)

_CONFIG_FOR_DOC = "VipLlavaConfig"


@dataclass
# Copied from transformers.models.llava.modeling_llava.LlavaCausalLMOutputWithPast with Llava->VipLlava
class VipLlavaCausalLMOutputWithPast(ModelOutput):
    """
    Base class for VipLlava causal language model (or autoregressive) outputs.

    Args:
        loss (`torch.FloatTensor` of shape `(1,)`, *optional*, returned when `labels` is provided):
            Language modeling loss (for next-token prediction).
        logits (`torch.FloatTensor` of shape `(batch_size, sequence_length, config.vocab_size)`):
            Prediction scores of the language modeling head (scores for each vocabulary token before SoftMax).
        past_key_values (`tuple(tuple(torch.FloatTensor))`, *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)`)

            Contains pre-computed hidden-states (key and values in the self-attention blocks) that can be used (see
            `past_key_values` input) to speed up sequential decoding.
        hidden_states (`tuple(torch.FloatTensor)`, *optional*, returned when `output_hidden_states=True` is passed or when `config.output_hidden_states=True`):
            Tuple of `torch.FloatTensor` (one for the output of the embeddings, if the model has an embedding layer, +
            one for the output of each layer) of shape `(batch_size, sequence_length, hidden_size)`.

            Hidden-states of the model at the output of each layer plus the optional initial embedding outputs.
        attentions (`tuple(torch.FloatTensor)`, *optional*, returned when `output_attentions=True` is passed or when `config.output_attentions=True`):
            Tuple of `torch.FloatTensor` (one for each layer) of shape `(batch_size, num_heads, sequence_length,
            sequence_length)`.

            Attentions weights after the attention softmax, used to compute the weighted average in the self-attention
            heads.
        image_hidden_states (`torch.FloatTensor`, *optional*):
            A `torch.FloatTensor` of size (batch_size, num_images, sequence_length, hidden_size)`.
            image_hidden_states of the model produced by the vision encoder and after projecting the last hidden state.
    """

    loss: Optional[torch.FloatTensor] = None
    logits: torch.FloatTensor = None
    past_key_values: Optional[List[torch.FloatTensor]] = None
    hidden_states: Optional[Tuple[torch.FloatTensor]] = None
    attentions: Optional[Tuple[torch.FloatTensor]] = None
    image_hidden_states: Optional[torch.FloatTensor] = None


class VipLlavaMultiModalProjector(nn.Module):
    def __init__(self, config: VipLlavaConfig):
        super().__init__()
        self.projector_layernorm = nn.LayerNorm(
            len(config.vision_feature_layers) * config.vision_config.hidden_size, eps=config.projector_layernorm_eps
        )

        self.linear_1 = nn.Linear(
            len(config.vision_feature_layers) * config.vision_config.hidden_size,
            config.text_config.hidden_size,
            bias=True,
        )
        self.act = ACT2FN[config.projector_hidden_act]
        self.linear_2 = nn.Linear(config.text_config.hidden_size, config.text_config.hidden_size, bias=True)

    def forward(self, hidden_states):
        hidden_states = self.projector_layernorm(hidden_states)
        hidden_states = self.linear_1(hidden_states)
        hidden_states = self.act(hidden_states)
        hidden_states = self.linear_2(hidden_states)
        return hidden_states


VIPLLAVA_START_DOCSTRING = r"""
    This model inherits from [`PreTrainedModel`]. Check the superclass documentation for the generic methods the
    library implements for all its model (such as downloading or saving, resizing the input embeddings, pruning heads
    etc.)

    This model is also a PyTorch [torch.nn.Module](https://pytorch.org/docs/stable/nn.html#torch.nn.Module) subclass.
    Use it as a regular PyTorch Module and refer to the PyTorch documentation for all matter related to general usage
    and behavior.

    Parameters:
        config ([`VipLlavaConfig`] or [`VipLlavaVisionConfig`]):
            Model configuration class with all the parameters of the model. Initializing with a config file does not
            load the weights associated with the model, only the configuration. Check out the
            [`~PreTrainedModel.from_pretrained`] method to load the model weights.
"""


@add_start_docstrings(
    "The bare VipLlava Model outputting raw hidden-states without any specific head on top.",
    VIPLLAVA_START_DOCSTRING,
)
# Copied from transformers.models.llava.modeling_llava.LlavaPreTrainedModel with Llava->VipLlava,llava->vipllava
class VipLlavaPreTrainedModel(PreTrainedModel):
    config_class = VipLlavaConfig
    base_model_prefix = "model"
    supports_gradient_checkpointing = True
    _no_split_modules = ["VipLlavaVisionAttention"]
    _skip_keys_device_placement = "past_key_values"
    _supports_cache_class = True
    _supports_flash_attn_2 = True
    _supports_sdpa = True

    def _init_weights(self, module):
        # important: this ported version of VipLlava isn't meant for training from scratch - only
        # inference and fine-tuning - so the proper init weights code has been removed - the original codebase
        # https://github.com/haotian-liu/LLaVA/tree/main/vipllava should serve for that purpose
        std = (
            self.config.initializer_range
            if hasattr(self.config, "initializer_range")
            else self.config.text_config.initializer_range
        )

        if hasattr(module, "class_embedding"):
            module.class_embedding.data.normal_(mean=0.0, std=std)

        if isinstance(module, (nn.Linear, nn.Conv2d)):
            module.weight.data.normal_(mean=0.0, std=std)
            if module.bias is not None:
                module.bias.data.zero_()
        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_()


VIPLLAVA_INPUTS_DOCSTRING = 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)
        pixel_values (`torch.FloatTensor` of shape `(batch_size, num_channels, image_size, image_size)):
            The tensors corresponding to the input images. Pixel values can be obtained using
            [`AutoImageProcessor`]. See [`CLIPImageProcessor.__call__`] for details ([]`LlavaProcessor`] uses
            [`CLIPImageProcessor`] for processing images).
        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)

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

            If `past_key_values` is used, optionally only the last `decoder_input_ids` have to be input (see
            `past_key_values`).

            If you want to change padding behavior, you should read [`modeling_opt._prepare_decoder_attention_mask`]
            and modify to your needs. See diagram 1 in [the paper](https://arxiv.org/abs/1910.13461) for more
            information on the default strategy.

            - 1 indicates the head is **not masked**,
            - 0 indicates the head is **masked**.
        position_ids (`torch.LongTensor` of shape `(batch_size, sequence_length)`, *optional*):
            Indices of positions of each input sequence tokens in the position embeddings. Selected in the range `[0,
            config.n_positions - 1]`. [What are position IDs?](../glossary#position-ids)
        past_key_values (`tuple(tuple(torch.FloatTensor))`, *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.
        use_cache (`bool`, *optional*):
            If set to `True`, `past_key_values` key value states are returned and can be used to speed up decoding (see
            `past_key_values`).
        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. Contrarily to `position_ids`,
            this tensor is not affected by padding. It is used to update the cache in the correct position and to infer
            the complete sequence length.
"""


@add_start_docstrings(
    """The VIPLLAVA model which consists of a vision backbone and a language model.""",
    VIPLLAVA_START_DOCSTRING,
)
# Copied from transformers.models.llava.modeling_llava.LlavaForConditionalGeneration with LLAVA->VIPLLAVA,Llava->VipLlava
class VipLlavaForConditionalGeneration(VipLlavaPreTrainedModel, GenerationMixin):
    def __init__(self, config: VipLlavaConfig):
        super().__init__(config)
        self.vision_tower = AutoModel.from_config(config.vision_config)

        self.multi_modal_projector = VipLlavaMultiModalProjector(config)
        self.vocab_size = config.text_config.vocab_size
        self.language_model = AutoModelForCausalLM.from_config(config.text_config)
        self.pad_token_id = self.config.pad_token_id if self.config.pad_token_id is not None else -1
        self.post_init()

    def get_input_embeddings(self):
        return self.language_model.get_input_embeddings()

    def set_input_embeddings(self, value):
        self.language_model.set_input_embeddings(value)

    def get_output_embeddings(self):
        return self.language_model.get_output_embeddings()

    def set_output_embeddings(self, new_embeddings):
        self.language_model.set_output_embeddings(new_embeddings)

    def set_decoder(self, decoder):
        self.language_model.set_decoder(decoder)

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

    def tie_weights(self):
        return self.language_model.tie_weights()

    def resize_token_embeddings(self, new_num_tokens: Optional[int] = None, pad_to_multiple_of=None) -> nn.Embedding:
        model_embeds = self.language_model.resize_token_embeddings(new_num_tokens, pad_to_multiple_of)
        # update vocab size
        self.config.text_config.vocab_size = model_embeds.num_embeddings
        self.vocab_size = model_embeds.num_embeddings
        return model_embeds

    # Ignore copy
    def get_image_features(self, pixel_values: torch.FloatTensor, vision_feature_layers: List[int]):
        """
        Obtains image last hidden states from the vision tower and apply multimodal projection.

        Args:
            pixel_values (`torch.FloatTensor]` of shape `(batch_size, channels, height, width)`)
               The tensors corresponding to the input images.
            vision_feature_layers (`List[int]`):
                The list og indexes of the layers to select the vision feature.
        Returns:
            image_features (`torch.Tensor`): Image feature tensor of shape `(num_images, image_length, embed_dim)`).
        """
        image_outputs = self.vision_tower(pixel_values, output_hidden_states=True)

        # For VIP-llava, the image features are computed this way
        # We select the features from index 1: for the layers -2, -5, -8, -11 and 6
        image_features = [image_outputs.hidden_states[index][:, 1:] for index in vision_feature_layers]
        image_features = torch.cat(image_features, dim=-1)
        image_features = self.multi_modal_projector(image_features)
        return image_features

    def _merge_input_ids_with_image_features(self, image_features, inputs_embeds, input_ids, attention_mask, labels):
        num_images, num_image_patches, embed_dim = image_features.shape
        batch_size, sequence_length = input_ids.shape
        left_padding = not torch.sum(input_ids[:, -1] == torch.tensor(self.pad_token_id))
        # 1. Create a mask to know where special image tokens are
        special_image_token_mask = input_ids == self.config.image_token_index
        num_special_image_tokens = torch.sum(special_image_token_mask, dim=-1)
        # Compute the maximum embed dimension
        max_embed_dim = (num_special_image_tokens.max() * (num_image_patches - 1)) + sequence_length
        batch_indices, non_image_indices = torch.where(input_ids != self.config.image_token_index)

        # 2. Compute the positions where text should be written
        # Calculate new positions for text tokens in merged image-text sequence.
        # `special_image_token_mask` identifies image tokens. Each image token will be replaced by `nb_text_tokens_per_images - 1` text tokens.
        # `torch.cumsum` computes how each image token shifts subsequent text token positions.
        # - 1 to adjust for zero-based indexing, as `cumsum` inherently increases indices by one.
        new_token_positions = torch.cumsum((special_image_token_mask * (num_image_patches - 1) + 1), -1) - 1
        nb_image_pad = max_embed_dim - 1 - new_token_positions[:, -1]
        if left_padding:
            new_token_positions += nb_image_pad[:, None]  # offset for left padding
        text_to_overwrite = new_token_positions[batch_indices, non_image_indices]

        # 3. Create the full embedding, already padded to the maximum position
        final_embedding = torch.zeros(
            batch_size, max_embed_dim, embed_dim, dtype=inputs_embeds.dtype, device=inputs_embeds.device
        )
        final_attention_mask = torch.zeros(
            batch_size, max_embed_dim, dtype=attention_mask.dtype, device=inputs_embeds.device
        )
        if labels is not None:
            final_labels = torch.full(
                (batch_size, max_embed_dim), self.config.ignore_index, dtype=input_ids.dtype, device=input_ids.device
            )
        # In case the Vision model or the Language model has been offloaded to CPU, we need to manually
        # set the corresponding tensors into their correct target device.
        target_device = inputs_embeds.device
        batch_indices, non_image_indices, text_to_overwrite = (
            batch_indices.to(target_device),
            non_image_indices.to(target_device),
            text_to_overwrite.to(target_device),
        )
        attention_mask = attention_mask.to(target_device)

        # 4. Fill the embeddings based on the mask. If we have ["hey" "<image>", "how", "are"]
        # we need to index copy on [0, 577, 578, 579] for the text and [1:576] for the image features
        final_embedding[batch_indices, text_to_overwrite] = inputs_embeds[batch_indices, non_image_indices]
        final_attention_mask[batch_indices, text_to_overwrite] = attention_mask[batch_indices, non_image_indices]
        if labels is not None:
            final_labels[batch_indices, text_to_overwrite] = labels[batch_indices, non_image_indices]

        # 5. Fill the embeddings corresponding to the images. Anything that is not `text_positions` needs filling (#29835)
        image_to_overwrite = torch.full(
            (batch_size, max_embed_dim), True, dtype=torch.bool, device=inputs_embeds.device
        )
        image_to_overwrite[batch_indices, text_to_overwrite] = False
        image_to_overwrite &= image_to_overwrite.cumsum(-1) - 1 >= nb_image_pad[:, None].to(target_device)

        if image_to_overwrite.sum() != image_features.shape[:-1].numel():
            raise ValueError(
                f"The input provided to the model are wrong. The number of image tokens is {torch.sum(special_image_token_mask)} while"
                f" the number of image given to the model is {num_images}. This prevents correct indexing and breaks batch generation."
            )

        final_embedding[image_to_overwrite] = image_features.contiguous().reshape(-1, embed_dim).to(target_device)
        final_attention_mask |= image_to_overwrite
        position_ids = (final_attention_mask.cumsum(-1) - 1).masked_fill_((final_attention_mask == 0), 1)

        # 6. Mask out the embedding at padding positions, as we later use the past_key_value value to determine the non-attended tokens.
        batch_indices, pad_indices = torch.where(input_ids == self.pad_token_id)
        indices_to_mask = new_token_positions[batch_indices, pad_indices]

        final_embedding[batch_indices, indices_to_mask] = 0

        if labels is None:
            final_labels = None

        return final_embedding, final_attention_mask, final_labels, position_ids

    @add_start_docstrings_to_model_forward(VIPLLAVA_INPUTS_DOCSTRING)
    @replace_return_docstrings(output_type=VipLlavaCausalLMOutputWithPast, config_class=_CONFIG_FOR_DOC)
    # Ignore copy
    def forward(
        self,
        input_ids: torch.LongTensor = None,
        pixel_values: torch.FloatTensor = None,
        attention_mask: Optional[torch.Tensor] = None,
        position_ids: Optional[torch.LongTensor] = None,
        past_key_values: Optional[List[torch.FloatTensor]] = None,
        inputs_embeds: Optional[torch.FloatTensor] = None,
        vision_feature_layers: Optional[List[int]] = 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,
        num_logits_to_keep: int = 0,
    ) -> Union[Tuple, VipLlavaCausalLMOutputWithPast]:
        r"""
        Args:
            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]`.

            num_logits_to_keep (`int`, *optional*):
                Calculate logits for the last `num_logits_to_keep` tokens. If `0`, calculate logits for all
                `input_ids` (special case). Only last token logits are needed for generation, and calculating them only for that
                token can save memory, which becomes pretty significant for long sequences or large vocabulary size.


        Returns:

        Example:

        ```python
        >>> import torch
        >>> from PIL import Image
        >>> import requests
        >>> from transformers import AutoProcessor, VipLlavaForConditionalGeneration

        >>> model = VipLlavaForConditionalGeneration.from_pretrained("llava-hf/vip-llava-7b-hf", device_map="auto", torch_dtype=torch.float16)
        >>> processor = AutoProcessor.from_pretrained("llava-hf/vip-llava-7b-hf")

        >>> prompt = "A chat between a curious human and an artificial intelligence assistant. The assistant gives helpful, detailed, and polite answers to the human's questions.###Human: <image>\n{}###Assistant:"
        >>> question = "Can you please describe this image?"
        >>> prompt = prompt.format(question)
        >>> url = "https://huggingface.co/datasets/huggingface/documentation-images/resolve/main/diffusers/compel-neg.png"
        >>> image = Image.open(requests.get(url, stream=True).raw)

        >>> inputs = processor(text=text, images=image, return_tensors="pt").to(0, torch.float16)

        >>> # Generate
        >>> generate_ids = model.generate(**inputs, max_new_tokens=20)
        >>> processor.decode(generate_ids[0][len(inputs["input_ids"][0]):], skip_special_tokens=True)
        The image features a brown and white cat sitting on a green surface, with a red ball in its
        ```"""

        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
        vision_feature_layers = (
            vision_feature_layers if vision_feature_layers is not None else self.config.vision_feature_layers
        )

        if (input_ids is None) ^ (inputs_embeds is not None):
            raise ValueError("You must specify exactly one of input_ids or inputs_embeds")

        if pixel_values is not None and inputs_embeds is not None:
            raise ValueError(
                "You cannot specify both pixel_values and inputs_embeds at the same time, and must specify either one"
            )

        legacy_processing = False
        if inputs_embeds is None:
            inputs_embeds = self.get_input_embeddings()(input_ids)

            # if the number of image tokens is more than image embeddings seq length, then prob we expanded it in processing
            # not very reliable, but we don't expect one to actually pass 500+ images for one prompt
            # In case we're in decoding stage, legacy behavior is checked by presence of pixel values even if use_cache=True
            legacy_processing = (
                (input_ids == self.config.image_token_index).sum(1).max() < self.config.image_seq_length
            ) or (input_ids.shape[-1] == 1 and pixel_values is not None)

        if pixel_values is not None:
            image_features = self.get_image_features(
                pixel_values=pixel_values, vision_feature_layers=vision_feature_layers
            )

            if legacy_processing:
                logger.warning_once(
                    "Expanding inputs for image tokens in VipLLaVa should be done in processing. "
                    "Please add `patch_size` and `vision_feature_select_strategy` to the model's image processing config. "
                    "Using processors without these attributes in the config is deprecated and will throw an error in v4.47."
                )
                # prefill stage vs decoding stage (legacy behavior copied)
                if input_ids.shape[1] != 1:
                    inputs_embeds, attention_mask, labels, position_ids = self._merge_input_ids_with_image_features(
                        image_features, inputs_embeds, input_ids, attention_mask, labels
                    )
                    cache_position = torch.arange(attention_mask.shape[1], device=attention_mask.device)
                else:
                    # Retrieve the first layer to inspect the logits and mask out the hidden states
                    # that are set to 0
                    first_layer_past_key_value = past_key_values[0][0][:, :, :, 0]

                    # Sum all dimensions of head_dim (-1) to avoid random errors such as: https://github.com/huggingface/transformers/pull/28032#issuecomment-1863691941
                    batch_index, non_attended_tokens = torch.where(first_layer_past_key_value.float().sum(-2) == 0)

                    target_length = input_ids.shape[1]
                    past_length = first_layer_past_key_value.shape[-1]

                    extended_attention_mask = torch.ones(
                        (attention_mask.shape[0], past_length),
                        dtype=attention_mask.dtype,
                        device=attention_mask.device,
                    )

                    # Filter out only the tokens that can be un-attended, this can happen
                    # in the case one uses Llava + Fused modules where the cache on the
                    # first iteration is already big enough, or if one passes custom cache
                    valid_indices = non_attended_tokens < extended_attention_mask.size(-1)
                    new_batch_index = batch_index[valid_indices]
                    new_non_attended_tokens = non_attended_tokens[valid_indices]

                    # Zero-out the places where we don't need to attend
                    extended_attention_mask[new_batch_index, new_non_attended_tokens] = 0

                    attention_mask = torch.cat((extended_attention_mask, attention_mask[:, -target_length:]), dim=1)
                    position_ids = torch.sum(attention_mask, dim=1).unsqueeze(-1) - 1
                    cache_position = torch.arange(attention_mask.shape[1], device=attention_mask.device)[
                        -target_length:
                    ]

            # TODO: @raushan retain only the new behavior after v4.47
            else:
                n_image_tokens = (input_ids == self.config.image_token_index).sum(dim=-1)[0].item()
                n_image_features = image_features.shape[1]
                if n_image_tokens != n_image_features:
                    raise ValueError(
                        f"Image features and image tokens do not match: tokens: {n_image_tokens}, features {n_image_features}"
                    )
                special_image_mask = (
                    (input_ids == self.config.image_token_index)
                    .unsqueeze(-1)
                    .expand_as(inputs_embeds)
                    .to(inputs_embeds.device)
                )
                image_features = image_features.to(inputs_embeds.device, inputs_embeds.dtype)
                inputs_embeds = inputs_embeds.masked_scatter(special_image_mask, image_features)

        outputs = self.language_model(
            attention_mask=attention_mask,
            position_ids=position_ids,
            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,
            num_logits_to_keep=num_logits_to_keep,
        )

        logits = outputs[0]

        loss = None
        if labels is not None:
            # Shift so that tokens < n predict n
            if attention_mask is not None:
                shift_attention_mask = attention_mask[:, -(logits.shape[1] - 1) :].to(logits.device)
                shift_logits = logits[..., :-1, :][shift_attention_mask.to(logits.device) != 0].contiguous()
                shift_labels = labels[..., 1:][shift_attention_mask.to(labels.device) != 0].contiguous()
            else:
                shift_logits = logits[..., :-1, :].contiguous()
                shift_labels = labels[..., 1:].contiguous()
            # Flatten the tokens
            loss_fct = nn.CrossEntropyLoss()
            loss = loss_fct(
                shift_logits.view(-1, shift_logits.size(-1)), shift_labels.view(-1).to(shift_logits.device)
            )

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

        return VipLlavaCausalLMOutputWithPast(
            loss=loss,
            logits=logits,
            past_key_values=outputs.past_key_values,
            hidden_states=outputs.hidden_states,
            attentions=outputs.attentions,
            image_hidden_states=image_features if pixel_values is not None else None,
        )

    def prepare_inputs_for_generation(
        self,
        input_ids,
        past_key_values=None,
        inputs_embeds=None,
        pixel_values=None,
        attention_mask=None,
        cache_position=None,
        num_logits_to_keep=None,
        **kwargs,
    ):
        # Overwritten -- in specific circumstances we don't want to forward image inputs to the model

        # Trigger the new behavior if we have more than image embeddings seq length tokens for images
        legacy_processing = (
            input_ids is not None
            and (input_ids == self.config.image_token_index).sum(1).max() < self.config.image_seq_length
        )

        model_inputs = self.language_model.prepare_inputs_for_generation(
            input_ids,
            past_key_values=past_key_values,
            inputs_embeds=inputs_embeds,
            attention_mask=attention_mask,
            cache_position=cache_position,
            num_logits_to_keep=num_logits_to_keep,
            **kwargs,
        )

        if legacy_processing or cache_position[0] == 0:
            # If we're in cached decoding stage, pixel values should be None because input ids do not contain special image token anymore
            # Otherwise we need pixel values to be passed to model
            model_inputs["pixel_values"] = pixel_values

        return model_inputs