#                🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨
#           This file was automatically generated from src/transformers/models/llava_next_video/modular_llava_next_video.py.
#               Do NOT edit this file manually as any edits will be overwritten by the generation of
#             the file from the modular. If any change should be done, please apply the change to the
#                          modular_llava_next_video.py file directly. One of our CI enforces this.
#                🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨
# coding=utf-8
# Copyright 2024 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.

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

import numpy as np
import torch
from torch import nn

from ...activations import ACT2FN
from ...generation import GenerationMixin
from ...image_processing_utils import select_best_resolution
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_llava_next_video import LlavaNextVideoConfig


logger = logging.get_logger(__name__)

_CONFIG_FOR_DOC = "LlavaNextVideoConfig"


@dataclass
class LlavaNextVideoCausalLMOutputWithPast(ModelOutput):
    """
    Base class for LlavaNextVideo 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_patches, num_images, sequence_length, hidden_size)`.
            image_hidden_states of the model produced by the vision encoder and after projecting the last hidden state.

        video_hidden_states (`torch.FloatTensor`, *optional*):
            A `torch.FloatTensor`  of size `(batch_size * num_frames, num_videos, sequence_length, hidden_size)`.
            video_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

    video_hidden_states: Optional[torch.FloatTensor] = None


class LlavaNextVideoPooler(nn.Module):
    def __init__(self, config):
        super().__init__()

        mode = config.spatial_pool_mode
        stride = config.spatial_pool_stride
        out_channels = getattr(config, "spatial_pool_out_channels", config.vision_config.hidden_size)
        self.image_size = (config.vision_config.image_size // config.vision_config.patch_size) ** 2

        if mode == "average":
            self.pool = nn.AvgPool2d(kernel_size=stride, stride=stride)
        elif mode == "max":
            self.pool = nn.MaxPool2d(kernel_size=stride, stride=stride)
        elif mode == "conv":
            self.pool = nn.Conv2d(
                in_channels=config.vision_config.hidden_size,
                out_channels=out_channels,
                kernel_size=stride,
                stride=stride,
            )
        else:
            raise ValueError(f"Unknown pooling mode: {mode}. Has to be one of [`average`, `max`, `conv`]")

    def forward(self, image_features):
        ori_width = int(math.sqrt(image_features.shape[1] * self.image_size // self.image_size))
        ori_height = int(ori_width * self.image_size // self.image_size)

        batch_size, _, dim = image_features.shape
        image_features_spatial = image_features.view(batch_size, ori_height, ori_height, dim).permute(0, 3, 1, 2)
        image_features_spatial_pool = self.pool(image_features_spatial)

        return image_features_spatial_pool.flatten(2).transpose(1, 2).contiguous()


class LlavaNextVideoMultiModalProjector(nn.Module):
    def __init__(self, config: LlavaNextVideoConfig):
        super().__init__()

        self.linear_1 = nn.Linear(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, image_features):
        hidden_states = self.linear_1(image_features)
        hidden_states = self.act(hidden_states)
        hidden_states = self.linear_2(hidden_states)
        return hidden_states


LLAVA_NEXT_VIDEO_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 ([`LlavaNextVideoConfig`] or [`LlavaNextVideoVisionConfig`]):
            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 LLaMA Model outputting raw hidden-states without any specific head on top.",
    LLAVA_NEXT_VIDEO_START_DOCSTRING,
)
class LlavaNextVideoPreTrainedModel(PreTrainedModel):
    config_class = LlavaNextVideoConfig
    base_model_prefix = "model"
    supports_gradient_checkpointing = True
    _no_split_modules = ["LlavaNextVideoVisionAttention"]
    _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 LlavaNextVideo 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/llava_next_video 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_()


def get_anyres_image_grid_shape(image_size, grid_pinpoints, patch_size):
    """
    Calculate the shape of the image patch grid after the preprocessing for images of any resolution.

    Args:
        image_size (`tuple`):
            The size of the input image in the format (width, height).
        grid_pinpoints (`List`):
            A list containing possible resolutions. Each item in the list should be a tuple or list
            of the form `(height, width)`.
        patch_size (`int`):
            The size of each image patch.

    Returns:
        tuple: The shape of the image patch grid in the format (width, height).
    """
    if not isinstance(grid_pinpoints, list):
        raise TypeError("grid_pinpoints should be a list of tuples or lists")

    # ! VERY IMPORTANT if image_size is tensor, must convert to into tuple, otherwise it will cause wrong calculate
    if not isinstance(image_size, (list, tuple)):
        if not isinstance(image_size, (torch.Tensor, np.ndarray)):
            raise TypeError(
                f"image_size invalid type: {type(image_size)} not valid, should be either list, tuple, np.ndarray or tensor"
            )
        image_size = image_size.tolist()

    height, width = select_best_resolution(image_size, grid_pinpoints)
    return height // patch_size, width // patch_size


def image_size_to_num_patches(image_size, grid_pinpoints, patch_size: int):
    """
    Calculate the number of patches after the preprocessing for images of any resolution.

    Args:
        image_size (`torch.LongTensor` or `np.ndarray` or `Tuple[int, int]`):
            The size of the input image in the format (height, width). ?
        grid_pinpoints (`List`):
            A list containing possible resolutions. Each item in the list should be a tuple or list
            of the form `(height, width)`.
        patch_size (`int`):
            The size of each image patch.

    Returns:
        int: the number of patches
    """
    if not isinstance(grid_pinpoints, list):
        raise TypeError("grid_pinpoints should be a list of tuples or lists")

    # ! VERY IMPORTANT if image_size is tensor, must convert to into tuple, otherwise it will cause wrong calculate
    if not isinstance(image_size, (list, tuple)):
        if not isinstance(image_size, (torch.Tensor, np.ndarray)):
            raise TypeError(f"image_size invalid type {type(image_size)} with value {image_size}")
        image_size = image_size.tolist()

    best_resolution = select_best_resolution(image_size, grid_pinpoints)
    height, width = best_resolution
    num_patches = 0
    # consider change to ceil(height/patch_size)*ceil(width/patch_size) + 1
    for i in range(0, height, patch_size):
        for j in range(0, width, patch_size):
            num_patches += 1
    # add the base patch
    num_patches += 1
    return num_patches


def unpad_image(tensor, original_size):
    """
    Unpads a PyTorch tensor of a padded and resized image.

    Args:
        tensor (`torch.Tensor`):
            The image tensor, assumed to be of shape (num_channels, height, width).
        original_size (`tuple`):
            The original size of the image (height, width).

    Returns:
        `torch.Tensor`: The unpadded image tensor.
    """
    if not isinstance(original_size, (list, tuple)):
        if not isinstance(original_size, (torch.Tensor, np.ndarray)):
            raise TypeError(
                f"image_size invalid type: {type(original_size)} not valid, should be either list, tuple, np.ndarray or tensor"
            )
        original_size = original_size.tolist()
    original_height, original_width = original_size
    current_height, current_width = tensor.shape[1:]

    original_aspect_ratio = original_width / original_height
    current_aspect_ratio = current_width / current_height

    if original_aspect_ratio > current_aspect_ratio:
        scale_factor = current_width / original_width
        new_height = int(round(original_height * scale_factor, 7))
        padding = (current_height - new_height) // 2
        unpadded_tensor = tensor[:, padding : current_height - padding, :]
    else:
        scale_factor = current_height / original_height
        new_width = int(round(original_width * scale_factor, 7))
        padding = (current_width - new_width) // 2
        unpadded_tensor = tensor[:, :, padding : current_width - padding]

    return unpadded_tensor


LLAVA_NEXT_VIDEO_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 [`LlavaNextVideoImageProcessor.__call__`] for details. [`LlavaProcessor`] uses
            [`LlavaNextVideoImageProcessor`] for processing images.
        image_sizes (`torch.LongTensor` of shape `(batch_size, 2)`, *optional*):
            The sizes of the images in the batch, being (height, width) for each image.
        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.
        vision_feature_layer (`int`, *optional*, defaults to -2):
            The index of the layer to select the vision feature.
        vision_feature_select_strategy (`str`, *optional*, defaults to `"default"`):
            The feature selection strategy used to select the vision feature from the vision backbone.
            Can be one of `"default"` or `"full"`. If `"default"`, the CLS token is removed from the vision features.
            If `"full"`, the full vision features are used.
        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 LLAVA-NeXT model which consists of a vision backbone and a language model.""",
    LLAVA_NEXT_VIDEO_START_DOCSTRING,
)
class LlavaNextVideoForConditionalGeneration(LlavaNextVideoPreTrainedModel, GenerationMixin):
    def __init__(
        self,
        config: LlavaNextVideoConfig,
    ):
        super().__init__(config)
        self.vision_tower = AutoModel.from_config(config.vision_config)

        self.multi_modal_projector = LlavaNextVideoMultiModalProjector(config)
        embed_std = 1 / math.sqrt(config.text_config.hidden_size)
        self.image_newline = nn.Parameter(torch.randn(config.text_config.hidden_size, dtype=self.dtype) * embed_std)

        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._padding_side = "left"  # set it to left by default, user can use setter to change padding_sides
        self.vision_resampler = LlavaNextVideoPooler(config)
        self.post_init()

    @property
    def padding_side(self):
        return self._padding_side

    @padding_side.setter
    def padding_side(self, padding_side: str):
        if padding_side not in ["left", "right"]:
            raise ValueError(f"{padding_side} is not `left` or `right`.")
        self._padding_side = padding_side

    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

    def _merge_input_ids_with_image_features(
        self,
        image_features,
        feature_lens,
        inputs_embeds,
        input_ids,
        attention_mask,
        position_ids=None,
        labels=None,
        image_token_index=None,
        ignore_index=-100,
    ):
        """
        Merge input_ids with with image features into final embeddings

        Args:
            image_features (`torch.Tensor` of shape `(all_feature_lens, embed_dim)`):
                All vision vectors of all images in the batch
            feature_lens (`torch.LongTensor` of shape `(num_images)`):
                The length of visual embeddings of each image as stacked in `image_features`
            inputs_embeds (`torch.Tensor` of shape `(batch_size, sequence_length, embed_dim)`):
                Token embeddings before merging with visual embeddings
            input_ids (`torch.LongTensor` of shape `(batch_size, sequence_length)`):
                Input_ids of tokens, possibly filled with image token
            attention_mask (`torch.LongTensor` of shape `(batch_size, sequence_length)`):
                Mask to avoid performing attention on padding token indices.
            position_ids (`torch.LongTensor` of shape `(batch_size, sequence_length)`):
                Indices of positions of each input sequence tokens in the position embeddings. Selected in the range `[0,
                config.n_positions - 1]`.
            labels (`torch.Tensor` of shape `(batch_size, sequence_length)`, *optional*)
                :abels need to be recalculated to support training (if provided)
            image_token_index (`int`, *optional*)
                Token id used to indicate the special "image" token. Defaults to `config.image_token_index`
            ignore_index (`int`, *optional*)
                Value that is used to pad `labels` and will be ignored when calculated loss. Default: -100.
        Returns:
            final_embedding, final_attention_mask, position_ids, final_labels

        Explanation:
            each image has variable length embeddings, with length specified by feature_lens
            image_features is concatenation of all visual embed vectors
            task: fill each <image> with the correct number of visual embeddings
            Example:
                X (5 patches), Y (3 patches), Z (8)
                X, Y are in the same sequence (in-context learning)
            if right padding
                input_ids: [
                    a b c d e f X g h i j k Y l m
                    o p q r Z s t u v _ _ _ _ _ _
                ]
                input_ids should be: [
                    a b c d e f X X X X X g h i j k Y Y Y l m
                    o p q r Z Z Z Z Z Z Z Z s t u v _ _ _ _ _
                ]
                labels should be: [
                    a b c d e f _ _ _ _ _ g h i j k _ _ _ l m
                    o p q r _ _ _ _ _ _ _ _ s t u v _ _ _ _ _
                ]
            elif left padding
                input_ids: [
                    a b c d e f X g h i j k Y l m
                    _ _ _ _ _ _ o p q r Z s t u v
                ]
                input_ids should be: [
                    a b c d e f X X X X X g h i j k Y Y Y l m
                    _ _ _ _ _ o p q r Z Z Z Z Z Z Z Z s t u v
                ]
                labels should be: [
                    a b c d e f _ _ _ _ _ g h i j k _ _ _ l m
                    _ _ _ _ _ o p q r _ _ _ _ _ _ _ _ s t u v
                ]
            Edge cases:
                * If tokens are same but image token sizes are different, then cannot infer left or right padding
                ```python
                cat_img = Image.open(requests.get("http://images.cocodataset.org/val2017/000000039769.jpg", stream=True).raw)
                chart_img = Image.open(requests.get("https://github.com/haotian-liu/LLaVA/blob/1a91fc274d7c35a9b50b3cb29c4247ae5837ce39/images/llava_v1_5_radar.jpg?raw=true", stream=True).raw)
                prompts = [
                    "[INST] <image>\nWhat is shown in this image? [/INST]",
                    "[INST] <image>\nWhat is shown in this image? [/INST]",
                ]
                inputs = processor(prompts, [chart_img, cat_img], return_tensors='pt', padding=True).to("cuda")
                    chart_img has 2634 tokens, while cat_img has 2340 tokens
                ```

                input_ids: [
                    a b c d X g h
                    i j Y k l m n
                ]
                where X is 3 tokens while Y is 5, this mean after merge
                if left-padding (batched generation)
                    input_ids should be: [
                        _ _ a b c d X X X g h
                        i j Y Y Y Y Y k l m n
                    ]
                elif (right padding) (training)
                    input_ids should be: [
                        a b c d X X X g h _ _
                        i j Y Y Y Y Y k l m n
                    ]
        """
        image_token_index = image_token_index if image_token_index is not None else self.config.image_token_index
        ignore_index = ignore_index if ignore_index is not None else self.config.ignore_index

        if self.training and self.padding_side == "left":
            logger.warning_once(
                "Padding side is set to 'left' but the model is in training mode. For training "
                "it is recommended to set `model.padding_side='right' and `processor.tokenizer.padding_side='right'`. "
                "If that's intended, ignore this warning"
            )
        if not self.training and self.padding_side == "right":
            logger.warning_once(
                "Padding side is set to 'right' but the model is in inference mode. For correct "
                "generation results, please set `model.padding_side='left'` and `processor.tokenizer.padding_side='left'`. "
                "If that's intended, ignore this warning"
            )

        with torch.no_grad():
            # ! in llava 1.6, number of patches is variable
            num_images = feature_lens.size(0)
            num_image_features, embed_dim = image_features.shape
            if feature_lens.sum() != num_image_features:
                raise ValueError(f"{feature_lens=} / {feature_lens.sum()} != {image_features.shape=}")
            batch_size = input_ids.shape[0]
            _left_padding = torch.any(attention_mask[:, 0] == 0)
            _right_padding = torch.any(attention_mask[:, -1] == 0)

            left_padding = self.padding_side == "left"
            if batch_size > 1:
                if _left_padding and _right_padding:
                    raise ValueError(f"both side of attention_mask has zero, invalid. {attention_mask}")
                elif _right_padding and left_padding:
                    left_padding = False
                elif _left_padding and not left_padding:
                    left_padding = True

            # Whether to turn off right padding
            # 1. Create a mask to know where special image tokens are
            special_image_token_mask = input_ids == image_token_index
            # special_image_token_mask: [bsz, seqlen]
            num_special_image_tokens = torch.sum(special_image_token_mask, dim=-1)
            # num_special_image_tokens: [bsz]
            # Reserve for padding of num_images
            total_num_special_image_tokens = torch.sum(special_image_token_mask)
            if total_num_special_image_tokens != num_images:
                raise ValueError(
                    f"Number of image tokens in input_ids ({total_num_special_image_tokens}) different from num_images ({num_images})."
                )
            # Compute the maximum embed dimension
            # max_image_feature_lens is max_feature_lens per batch
            feature_lens = feature_lens.to(input_ids.device)
            feature_lens_batch = feature_lens.split(num_special_image_tokens.tolist(), dim=0)
            feature_lens_batch_sum = torch.tensor([x.sum() for x in feature_lens_batch], device=input_ids.device)
            embed_sequence_lengths = (
                (attention_mask == 1).long().sum(-1) - num_special_image_tokens + feature_lens_batch_sum
            )
            max_embed_dim = embed_sequence_lengths.max()

            batch_indices, non_image_indices = torch.where((input_ids != image_token_index) & (attention_mask == 1))
            # 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` 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.
            # ! instead of special_image_token_mask * (num_image_patches - 1)
            #   special_image_token_mask * (num_feature_len - 1)
            special_image_token_mask = special_image_token_mask.long()
            special_image_token_mask[special_image_token_mask == 1] = feature_lens - 1
            new_token_positions = torch.cumsum((special_image_token_mask + 1), -1) - 1
            if left_padding:
                # shift right token positions so that they are ending at the same number
                # the below here was incorrect? new_token_positions += new_token_positions[:, -1].max() - new_token_positions[:, -1:]
                new_token_positions += max_embed_dim - 1 - new_token_positions[:, -1:]

            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
        )
        final_input_ids = torch.full(
            (batch_size, max_embed_dim), self.pad_token_id, dtype=input_ids.dtype, device=inputs_embeds.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)
        input_ids = input_ids.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]
        final_input_ids[batch_indices, text_to_overwrite] = input_ids[batch_indices, non_image_indices]
        final_labels = None
        if labels is not None:
            labels = labels.to(target_device)
            final_labels = torch.full_like(final_attention_mask, ignore_index).to(torch.long)
            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)
        with torch.no_grad():
            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
            embed_indices = torch.arange(max_embed_dim).unsqueeze(0).to(target_device)
            embed_indices = embed_indices.expand(batch_size, max_embed_dim)
            embed_seq_lens = embed_sequence_lengths[:, None].to(target_device)

            if left_padding:
                # exclude padding on the left
                max_embed_dim = max_embed_dim.to(target_device)
                val = (max_embed_dim - embed_indices) <= embed_seq_lens
            else:
                # exclude padding on the right
                val = embed_indices < embed_seq_lens
            image_to_overwrite &= val

            if image_to_overwrite.sum() != num_image_features:
                raise ValueError(
                    f"{image_to_overwrite.sum()=} != {num_image_features=} The input provided to the model are wrong. "
                    f"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}. "
                    f"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)

        return final_embedding, final_attention_mask, position_ids, final_labels, final_input_ids

    def pack_image_features(self, image_features, image_sizes, vision_feature_select_strategy, image_newline=None):
        """
        Reshape, unpad and then pack each image_feature into a single image_features tensor containing all visual vectors.

        Args:
            image_features (`List[torch.Tensor]` of length num_images, each of shape `(num_patches, image_length, embed_dim)`)
                List of image feature tensor, each contains all the visual feature of all patches.
            image_sizes (`torch.Tensor` of shape `(num_images, 2)`)
                Actual image size of each images (H, W).
            vision_feature_select_strategy (`str`)
                The feature selection strategy used to select the vision feature from the vision backbone.
            image_newline (`torch.Tensor` of shape `(embed_dim)`)
                New line embedding vector.
        Returns:
            image_features (`torch.Tensor` of shape `(all_feat_len, embed_dim)`)
            feature_lens (`List[int]`)
                token length of each image in image_features
        """
        new_image_features = []
        feature_lens = []
        for image_idx, image_feature in enumerate(image_features):
            if image_feature.shape[0] > 1:
                base_image_feature = image_feature[0]
                image_feature = image_feature[1:]
                height = width = self.config.vision_config.image_size // self.config.vision_config.patch_size

                if vision_feature_select_strategy == "default":
                    expected_num_patches = height * width
                elif vision_feature_select_strategy == "full":
                    expected_num_patches = height * width + 1
                if expected_num_patches != base_image_feature.shape[0]:
                    raise ValueError("The number of patches is not consistent with the image size.")

                num_patch_height, num_patch_width = get_anyres_image_grid_shape(
                    image_sizes[image_idx],
                    self.config.image_grid_pinpoints,
                    self.config.vision_config.image_size,
                )
                image_feature = image_feature.view(num_patch_height, num_patch_width, height, width, -1)
                image_feature = image_feature.permute(4, 0, 2, 1, 3).contiguous()
                image_feature = image_feature.flatten(1, 2).flatten(2, 3)
                image_feature = unpad_image(image_feature, image_sizes[image_idx])
                if image_newline is not None:
                    image_feature = torch.cat(
                        (
                            image_feature,
                            image_newline[:, None, None].expand(*image_feature.shape[:-1], 1).to(image_feature.dtype),
                        ),
                        dim=-1,
                    )
                image_feature = image_feature.flatten(1, 2).transpose(0, 1)
                image_feature = torch.cat((base_image_feature, image_feature), dim=0)
            else:
                image_feature = image_feature[0]
                if image_newline is not None:
                    image_feature = torch.cat((image_feature, image_newline[None].to(image_feature)), dim=0)
            new_image_features.append(image_feature)
            feature_lens.append(image_feature.size(0))
        image_features = torch.cat(new_image_features, dim=0)
        feature_lens = torch.tensor(feature_lens, dtype=torch.long, device=image_features.device)
        return image_features, feature_lens

    def get_image_features(
        self,
        pixel_values: torch.FloatTensor,
        image_sizes: torch.Tensor,
        vision_feature_layer: int,
        vision_feature_select_strategy: str,
    ):
        """
        Obtains image last hidden states from the vision tower and apply multimodal projection.

        Args:
            pixel_values (`torch.FloatTensor]` of shape `(batch_size, num_patches, channels, height, width)`)
               The tensors corresponding to the input images.
            image_sizes (`torch.Tensor` of shape `(num_images, 2)`)
                Actual image size of each images (H, W).
            vision_feature_layer (`int`):
                The index of the layer to select the vision feature.
            vision_feature_select_strategy (`str`):
                The feature selection strategy used to select the vision feature from the vision backbone.
                Can be one of `"default"` or `"full"`
        Returns:
            image_features (List[`torch.Tensor`]): List of image feature tensor, each contains all the visual feature of all patches
            and are of shape `(num_patches, image_length, embed_dim)`).
        """
        # ! infer image_num_patches from image_sizes
        image_num_patches = [
            image_size_to_num_patches(
                image_size=imsize,
                grid_pinpoints=self.config.image_grid_pinpoints,
                patch_size=self.config.vision_config.image_size,
            )
            for imsize in image_sizes
        ]
        if pixel_values.dim() == 5:
            # stacked if input is (batch_size, num_patches, num_channels, height, width)
            _pixel_values_list = [pix_val[:num_patch] for pix_val, num_patch in zip(pixel_values, image_num_patches)]
            pixel_values = torch.cat(_pixel_values_list, dim=0)
        elif pixel_values.dim() != 4:
            # otherwise has to be stacked from list of (num_patches, num_channels, height, width)
            raise ValueError(f"pixel_values of shape {pixel_values.shape}, expect to be of 4 or 5 dimensions")

        image_features = self.vision_tower(pixel_values, output_hidden_states=True)
        selected_image_feature = image_features.hidden_states[vision_feature_layer]
        if vision_feature_select_strategy == "default":
            selected_image_feature = selected_image_feature[:, 1:]
        elif vision_feature_select_strategy == "full":
            selected_image_feature = selected_image_feature
        image_features = self.multi_modal_projector(selected_image_feature)
        image_features = torch.split(image_features, image_num_patches, dim=0)
        return image_features

    @add_start_docstrings_to_model_forward(LLAVA_NEXT_VIDEO_INPUTS_DOCSTRING)
    @replace_return_docstrings(output_type=LlavaNextVideoCausalLMOutputWithPast, config_class=_CONFIG_FOR_DOC)
    def forward(
        self,
        input_ids: torch.LongTensor = None,
        pixel_values: torch.FloatTensor = None,
        pixel_values_videos: torch.FloatTensor = None,
        image_sizes: Optional[torch.LongTensor] = 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_layer: Optional[int] = None,
        vision_feature_select_strategy: Optional[str] = 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, LlavaNextVideoCausalLMOutputWithPast]:
        r"""
        Args:
            pixel_values_videos (`torch.FloatTensor` of shape `(batch_size, num_frames, num_channels, image_size, image_size)):
                The tensors corresponding to the input videos. Pixel values can be obtained using
                [`AutoImageProcessor`]. See [`LlavaNextVideoVideoProcessor.__call__`] for details. [`LlavaProcessor`] uses
                [`LlavaNextVideoVideoProcessor`] for processing videos.
            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
        >>> from PIL import Image
        >>> import requests
        >>> import av
        >>> from transformers import AutoProcessor, LlavaNextVideoForConditionalGeneration

        >>> def read_video_pyav(container, indices):
        ...     '''
        ...     Decode the video with PyAV decoder.
        ...     Args:
        ...         container (`av.container.input.InputContainer`): PyAV container.
        ...         indices (`List[int]`): List of frame indices to decode.
        ...     Returns:
        ...         result (np.ndarray): np array of decoded frames of shape (num_frames, height, width, 3).
        ...     '''
        ...     frames = []
        ...     container.seek(0)
        ...     start_index = indices[0]
        ...     end_index = indices[-1]
        ...     for i, frame in enumerate(container.decode(video=0)):
        ...         if i > end_index:
        ...             break
        ...         if i >= start_index and i in indices:
        ...             frames.append(frame)
        ...     return np.stack([x.to_ndarray(format="rgb24") for x in frames])

        >>> model = LlavaNextVideoForConditionalGeneration.from_pretrained("llava-hf/LLaVA-NeXT-Video-7B-hf", device_map="auto")
        >>> processor = AutoProcessor.from_pretrained("llava-hf/LLaVA-NeXT-Video-7B-hf")

        >>> prompt = "USER: <video>\nWhy is this video funny? ASSISTANT:"
        >>> video_path = hf_hub_download(repo_id="raushan-testing-hf/videos-test", filename="sample_demo_1.mp4", repo_type="dataset")
        >>> container = av.open(video_path)

        >>> # sample uniformly 8 frames from the video (model was trained with 32 frames per video, but this video is short)
        >>> total_frames = container.streams.video[0].frames
        >>> indices = np.arange(0, total_frames, total_frames / 8).astype(int)
        >>> clip = read_video_pyav(container, indices)
        >>> inputs_video = processor(text=prompt, videos=clip, return_tensors="pt").to(model.device)

        >>> # load an image to generate from an image
        >>> prompt = "USER:<image>\nWhat is shown in this image? ASSISTANT:"
        >>> url = "https://www.ilankelman.org/stopsigns/australia.jpg"
        >>> image = Image.open(requests.get(url, stream=True).raw)
        >>> inputs_image = processor(text=prompt, images=image, return_tensors="pt").to(model.device)

        >>> # Generate from video
        >>> generate_ids = model.generate(**inputs_video, max_length=50)
        >>> processor.batch_decode(generate_ids, skip_special_tokens=True, clean_up_tokenization_spaces=False)[0]
        "USER:\nWhy is this video funny? ASSISTANT: The humor in this video comes from the unexpected and endearing sight of a baby wearing glasses and (...)"

        >>> # Generate from image
        >>> generate_ids = model.generate(**inputs_image, max_length=30)
        >>> processor.batch_decode(generate_ids, skip_special_tokens=True, clean_up_tokenization_spaces=False)[0]
        "USER: \nWhat's the content of the image? ASSISTANT: The image shows a red stop sign on a pole, with a traditional Chinese archway (...)"
        ```"""

        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
        self.vision_feature_layer = (
            vision_feature_layer if vision_feature_layer is not None else self.config.vision_feature_layer
        )
        self.vision_feature_select_strategy = (
            vision_feature_select_strategy
            if vision_feature_select_strategy is not None
            else self.config.vision_feature_select_strategy
        )

        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 or pixel_values_videos is not None) and inputs_embeds is not None:
            raise ValueError(
                "You cannot specify both `pixel_values`/`pixel_values_videos` 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/video 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
            img_token_not_enough = (input_ids == self.config.image_token_index).sum(
                1
            ).max() < self.config.image_seq_length
            video_token_not_enough = (input_ids == self.config.video_token_index).sum(
                1
            ).max() < self.config.video_seq_length
            inputs_not_expanded = (img_token_not_enough and pixel_values is not None) or (
                video_token_not_enough and pixel_values_videos is not None
            )
            pixels_present = input_ids.shape[-1] == 1 and (pixel_values is not None or pixel_values_videos is not None)
            legacy_processing = inputs_not_expanded or pixels_present

        image_features = feature_lens = None
        if pixel_values is not None and pixel_values.size(0) > 0:
            image_features = self.get_image_features(
                pixel_values,
                image_sizes,
                vision_feature_layer=self.vision_feature_layer,
                vision_feature_select_strategy=self.vision_feature_select_strategy,
            )
            image_features, feature_lens = self.pack_image_features(
                image_features,
                image_sizes,
                self.vision_feature_select_strategy,
                image_newline=self.image_newline,
            )

        video_features = video_feature_lens = None
        if pixel_values_videos is not None and pixel_values_videos.size(0) > 0:
            video_features = self.get_video_features(
                pixel_values_videos,
                vision_feature_layer=self.vision_feature_layer,
                vision_feature_select_strategy=self.vision_feature_select_strategy,
            )
            video_features = [feature.flatten(0, 1) for feature in video_features]
            video_feature_lens = [feature.size(0) for feature in video_features]
            video_features = torch.cat(video_features, dim=0)
            video_feature_lens = torch.tensor(video_feature_lens, dtype=torch.long, device=video_features.device)

        if legacy_processing:
            logger.warning_once(
                "Expanding inputs for image.video tokens in LLaVa-NeXT-Video should be done in processing. "
                "Please add `patch_size` and `vision_feature_select_strategy` to the model's processing config or set directly "
                "with `processor.patch_size = {{patch_size}}` and processor.vision_feature_select_strategy = {{vision_feature_select_strategy}}`. "
                "Using processors without these attributes in the config is deprecated and will throw an error in v4.47."
            )
            if input_ids.shape[1] != 1:
                iterator = (
                    (image_features, feature_lens, self.config.image_token_index),
                    (video_features, video_feature_lens, self.config.video_token_index),
                )
                for features, lens, special_token in iterator:
                    if features is not None:
                        (
                            inputs_embeds,
                            attention_mask,
                            position_ids,
                            labels,
                            input_ids,
                        ) = self._merge_input_ids_with_image_features(
                            features,
                            lens,
                            inputs_embeds,
                            input_ids,
                            attention_mask,
                            position_ids,
                            labels=labels,
                            image_token_index=special_token,
                        )
                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 (-2) 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)
                # Get the target length
                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
                # if 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:
            if image_features is not None:
                n_image_tokens = (input_ids == self.config.image_token_index).sum().item()
                n_image_features = image_features.shape[0]

                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)
            if video_features is not None:
                n_video_tokens = (input_ids == self.config.video_token_index).sum().item()
                n_video_features = video_features.shape[0]
                if n_video_tokens != n_video_features:
                    raise ValueError(
                        f"Video features and video tokens do not match: tokens: {n_video_tokens}, features {n_video_features}"
                    )
                special_image_mask = (
                    (input_ids == self.config.video_token_index)
                    .unsqueeze(-1)
                    .expand_as(inputs_embeds)
                    .to(inputs_embeds.device)
                )
                video_features = video_features.to(inputs_embeds.device, inputs_embeds.dtype)
                inputs_embeds = inputs_embeds.masked_scatter(special_image_mask, video_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:
                # we use the input attention mask to shift the logits and labels, because it is 2D.
                # we also crop attn mask in case it is longer, which happens in PrefixTuning with peft
                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 LlavaNextVideoCausalLMOutputWithPast(
            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,
            video_hidden_states=video_features if pixel_values_videos is not None else None,
        )

    def prepare_inputs_for_generation(
        self,
        input_ids,
        past_key_values=None,
        inputs_embeds=None,
        pixel_values=None,
        pixel_values_videos=None,
        image_sizes=None,
        attention_mask=None,
        cache_position=None,
        num_logits_to_keep=None,
        **kwargs,
    ):
        # Overwritten -- extra custom processing

        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 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
        if cache_position[0] == 0:
            model_inputs["pixel_values"] = pixel_values
            model_inputs["pixel_values_videos"] = pixel_values_videos
            model_inputs["image_sizes"] = image_sizes

        return model_inputs

    def get_video_features(
        self, pixel_values: torch.FloatTensor, vision_feature_layer: int, vision_feature_select_strategy: str
    ):
        """
        Obtains video last hidden states from the vision tower and apply multimodal projection.

        Args:
            pixel_values (`torch.FloatTensor]` of shape `(batch_size, num_frames, channels, height, width)`)
               The tensors corresponding to the input video.
            vision_feature_layer (`int`):
                The index of the layer to select the vision feature.
            vision_feature_select_strategy (`str`):
                The feature selection strategy used to select the vision feature from the vision backbone.
                Can be one of `"default"` or `"full"`
        Returns:
            video_features (List[`torch.Tensor`]): List of video feature tensor, each contains all the visual feature of all patches
            and are of shape `(num_videos, video_length, embed_dim)`).
        """
        batch_size, frames, channels, height, width = pixel_values.shape
        pixel_values = pixel_values.reshape(batch_size * frames, channels, height, width)
        video_features = self.vision_tower(pixel_values, output_hidden_states=True)
        selected_video_features = video_features.hidden_states[vision_feature_layer]
        if vision_feature_select_strategy == "default":
            selected_video_features = selected_video_features[:, 1:]
        elif vision_feature_select_strategy == "full":
            selected_video_features = selected_video_features

        # Same as image features except that video has pooling layer
        video_features = self.vision_resampler(selected_video_features)
        video_features = self.multi_modal_projector(video_features)
        video_features = torch.split(video_features, frames, dim=0)
        return video_features
