modeling_omchat.py

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

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

from transformers.configuration_utils import PretrainedConfig
from .configuration_omchat import OmChatConfig

from transformers import Qwen2Config, Qwen2Model, Qwen2ForCausalLM, AutoConfig, AutoModelForCausalLM
from transformers.utils import logging
from transformers.modeling_outputs import ModelOutput
from transformers.utils import (
    add_start_docstrings,
    add_start_docstrings_to_model_forward,
    logging,
    replace_return_docstrings,
)



logger = logging.get_logger(__name__)

_CONFIG_FOR_DOC = "OmChatConfig"

from typing import Optional, Tuple, Union

import torch
import torch.nn.functional as F
import torch.utils.checkpoint
from einops import rearrange
from timm.models.layers import DropPath
from torch import nn
from transformers.activations import ACT2FN
from transformers.modeling_outputs import (BaseModelOutput,
                                           BaseModelOutputWithPooling)
from transformers.modeling_utils import PreTrainedModel
from transformers.utils import logging

from .configuration_omchat import InternVisionConfig

try:
    from .flash_attention import FlashAttention
    has_flash_attn = True
except:
    print('FlashAttention is not installed.')
    has_flash_attn = False


logger = logging.get_logger(__name__)


class InternRMSNorm(nn.Module):
    def __init__(self, hidden_size, eps=1e-6):
        super().__init__()
        self.weight = nn.Parameter(torch.ones(hidden_size))
        self.variance_epsilon = eps

    def forward(self, hidden_states):
        input_dtype = hidden_states.dtype
        hidden_states = hidden_states.to(torch.float32)
        variance = hidden_states.pow(2).mean(-1, keepdim=True)
        hidden_states = hidden_states * torch.rsqrt(variance + self.variance_epsilon)
        return self.weight * hidden_states.to(input_dtype)


try:
    from apex.normalization import FusedRMSNorm

    InternRMSNorm = FusedRMSNorm  # noqa

    logger.info('Discovered apex.normalization.FusedRMSNorm - will use it instead of InternRMSNorm')
except ImportError:
    # using the normal InternRMSNorm
    pass
except Exception:
    logger.warning('discovered apex but it failed to load, falling back to InternRMSNorm')
    pass


class InternVisionEmbeddings(nn.Module):
    def __init__(self, config: InternVisionConfig):
        super().__init__()
        self.config = config
        self.embed_dim = config.hidden_size
        self.image_size = config.image_size
        self.patch_size = config.patch_size

        self.class_embedding = nn.Parameter(
            torch.randn(1, 1, self.embed_dim),
        )

        self.patch_embedding = nn.Conv2d(
            in_channels=3, out_channels=self.embed_dim, kernel_size=self.patch_size, stride=self.patch_size
        )

        self.num_patches = (self.image_size // self.patch_size) ** 2
        self.num_positions = self.num_patches + 1

        self.position_embedding = nn.Parameter(torch.randn(1, self.num_positions, self.embed_dim))

    def _get_pos_embed(self, pos_embed, H, W):
        target_dtype = pos_embed.dtype
        pos_embed = pos_embed.float().reshape(
            1, self.image_size // self.patch_size, self.image_size // self.patch_size, -1).permute(0, 3, 1, 2)
        pos_embed = F.interpolate(pos_embed, size=(H, W), mode='bicubic', align_corners=False).\
            reshape(1, -1, H * W).permute(0, 2, 1).to(target_dtype)
        return pos_embed

    def forward(self, pixel_values: torch.FloatTensor) -> torch.Tensor:
        target_dtype = self.patch_embedding.weight.dtype
        patch_embeds = self.patch_embedding(pixel_values)  # shape = [*, channel, width, height]
        batch_size, _, height, width = patch_embeds.shape
        patch_embeds = patch_embeds.flatten(2).transpose(1, 2)
        class_embeds = self.class_embedding.expand(batch_size, 1, -1).to(target_dtype)
        embeddings = torch.cat([class_embeds, patch_embeds], dim=1)
        position_embedding = torch.cat([
            self.position_embedding[:, :1, :],
            self._get_pos_embed(self.position_embedding[:, 1:, :], height, width)
        ], dim=1)
        embeddings = embeddings + position_embedding.to(target_dtype)
        return embeddings


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

    def __init__(self, config: InternVisionConfig):
        super().__init__()
        self.config = config
        self.embed_dim = config.hidden_size
        self.num_heads = config.num_attention_heads
        self.use_flash_attn = config.use_flash_attn and has_flash_attn
        if config.use_flash_attn and not has_flash_attn:
            print('Warning: Flash Attention is not available, use_flash_attn is set to False.')
        self.head_dim = self.embed_dim // self.num_heads
        if self.head_dim * self.num_heads != self.embed_dim:
            raise ValueError(
                f'embed_dim must be divisible by num_heads (got `embed_dim`: {self.embed_dim} and `num_heads`:'
                f' {self.num_heads}).'
            )

        self.scale = self.head_dim ** -0.5
        self.qkv = nn.Linear(self.embed_dim, 3 * self.embed_dim, bias=config.qkv_bias)
        self.attn_drop = nn.Dropout(config.attention_dropout)
        self.proj_drop = nn.Dropout(config.dropout)

        self.qk_normalization = config.qk_normalization

        if self.qk_normalization:
            self.q_norm = InternRMSNorm(self.embed_dim, eps=config.layer_norm_eps)
            self.k_norm = InternRMSNorm(self.embed_dim, eps=config.layer_norm_eps)

        if self.use_flash_attn:
            self.inner_attn = FlashAttention(attention_dropout=config.attention_dropout)
        self.proj = nn.Linear(self.embed_dim, self.embed_dim)

    def _naive_attn(self, x):
        B, N, C = x.shape
        qkv = self.qkv(x).reshape(B, N, 3, self.num_heads, C // self.num_heads).permute(2, 0, 3, 1, 4)
        q, k, v = qkv.unbind(0)  # make torchscript happy (cannot use tensor as tuple)

        if self.qk_normalization:
            B_, H_, N_, D_ = q.shape
            q = self.q_norm(q.transpose(1, 2).flatten(-2, -1)).view(B_, N_, H_, D_).transpose(1, 2)
            k = self.k_norm(k.transpose(1, 2).flatten(-2, -1)).view(B_, N_, H_, D_).transpose(1, 2)

        attn = ((q * self.scale) @ k.transpose(-2, -1))
        attn = attn.softmax(dim=-1)
        attn = self.attn_drop(attn)

        x = (attn @ v).transpose(1, 2).reshape(B, N, C)
        x = self.proj(x)
        x = self.proj_drop(x)
        return x

    def _flash_attn(self, x, key_padding_mask=None, need_weights=False):
        qkv = self.qkv(x)
        qkv = rearrange(qkv, 'b s (three h d) -> b s three h d', three=3, h=self.num_heads)

        if self.qk_normalization:
            q, k, v = qkv.unbind(2)
            q = self.q_norm(q.flatten(-2, -1)).view(q.shape)
            k = self.k_norm(k.flatten(-2, -1)).view(k.shape)
            qkv = torch.stack([q, k, v], dim=2)

        context, _ = self.inner_attn(
            qkv, key_padding_mask=key_padding_mask, need_weights=need_weights, causal=False
        )
        outs = self.proj(rearrange(context, 'b s h d -> b s (h d)'))
        outs = self.proj_drop(outs)
        return outs

    def forward(self, hidden_states: torch.Tensor) -> torch.Tensor:
        x = self._naive_attn(hidden_states) if not self.use_flash_attn else self._flash_attn(hidden_states)
        return x


class InternMLP(nn.Module):
    def __init__(self, config: InternVisionConfig):
        super().__init__()
        self.config = config
        self.act = ACT2FN[config.hidden_act]
        self.fc1 = nn.Linear(config.hidden_size, config.intermediate_size)
        self.fc2 = nn.Linear(config.intermediate_size, config.hidden_size)

    def forward(self, hidden_states: torch.Tensor) -> torch.Tensor:
        hidden_states = self.fc1(hidden_states)
        hidden_states = self.act(hidden_states)
        hidden_states = self.fc2(hidden_states)
        return hidden_states


class InternVisionEncoderLayer(nn.Module):
    def __init__(self, config: InternVisionConfig, drop_path_rate: float):
        super().__init__()
        self.embed_dim = config.hidden_size
        self.intermediate_size = config.intermediate_size

        self.attn = InternAttention(config)
        self.mlp = InternMLP(config)
        self.norm1 = InternRMSNorm(self.embed_dim, eps=config.layer_norm_eps)
        self.norm2 = InternRMSNorm(self.embed_dim, eps=config.layer_norm_eps)

        self.ls1 = nn.Parameter(config.initializer_factor * torch.ones(self.embed_dim))
        self.ls2 = nn.Parameter(config.initializer_factor * torch.ones(self.embed_dim))
        self.drop_path1 = DropPath(drop_path_rate) if drop_path_rate > 0. else nn.Identity()
        self.drop_path2 = DropPath(drop_path_rate) if drop_path_rate > 0. else nn.Identity()

    def forward(
            self,
            hidden_states: torch.Tensor,
    ) -> Tuple[torch.FloatTensor, Optional[torch.FloatTensor], Optional[Tuple[torch.FloatTensor]]]:
        """
        Args:
            hidden_states (`Tuple[torch.FloatTensor, Optional[torch.FloatTensor]]`): input to the layer of shape `(batch, seq_len, embed_dim)`
        """
        hidden_states = hidden_states + self.drop_path1(self.attn(self.norm1(hidden_states)) * self.ls1)

        hidden_states = hidden_states + self.drop_path2(self.mlp(self.norm2(hidden_states)) * self.ls2)

        return hidden_states


class InternVisionEncoder(nn.Module):
    """
    Transformer encoder consisting of `config.num_hidden_layers` self attention layers. Each layer is a
    [`InternEncoderLayer`].

    Args:
        config (`InternConfig`):
            The corresponding vision configuration for the `InternEncoder`.
    """

    def __init__(self, config: InternVisionConfig):
        super().__init__()
        self.config = config
        # stochastic depth decay rule
        dpr = [x.item() for x in torch.linspace(0, config.drop_path_rate, config.num_hidden_layers)]
        self.layers = nn.ModuleList([
            InternVisionEncoderLayer(config, dpr[idx]) for idx in range(config.num_hidden_layers)])
        self.gradient_checkpointing = True

    def forward(
            self,
            inputs_embeds,
            output_hidden_states: Optional[bool] = None,
            return_dict: Optional[bool] = None,
    ) -> Union[Tuple, BaseModelOutput]:
        r"""
        Args:
            inputs_embeds (`torch.FloatTensor` of shape `(batch_size, sequence_length, hidden_size)`):
                Embedded representation of the inputs. Should be float, not int tokens.
            output_hidden_states (`bool`, *optional*):
                Whether or not to return the hidden states of all layers. See `hidden_states` under returned tensors
                for more detail.
            return_dict (`bool`, *optional*):
                Whether or not to return a [`~utils.ModelOutput`] instead of a plain tuple.
        """
        output_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

        encoder_states = () if output_hidden_states else None
        hidden_states = inputs_embeds

        for idx, encoder_layer in enumerate(self.layers):
            if output_hidden_states:
                encoder_states = encoder_states + (hidden_states,)
            if self.gradient_checkpointing and self.training:
                layer_outputs = torch.utils.checkpoint.checkpoint(
                    encoder_layer,
                    hidden_states)
            else:
                layer_outputs = encoder_layer(
                    hidden_states,
                )
            hidden_states = layer_outputs

        if output_hidden_states:
            encoder_states = encoder_states + (hidden_states,)

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


class InternVisionModel(PreTrainedModel):
    main_input_name = 'pixel_values'
    config_class = InternVisionConfig
    _no_split_modules=["InternVisionEncoderLayer"]

    def __init__(self, config: InternVisionConfig):
        super().__init__(config)
        self.config = config

        self.embeddings = InternVisionEmbeddings(config)
        self.encoder = InternVisionEncoder(config)

    def resize_pos_embeddings(self, old_size, new_size, patch_size):
        pos_emb = self.embeddings.position_embedding
        _, num_positions, embed_dim = pos_emb.shape
        cls_emb = pos_emb[:, :1, :]
        pos_emb = pos_emb[:, 1:, :].reshape(1, old_size // patch_size, old_size // patch_size, -1).permute(0, 3, 1, 2)
        pos_emb = F.interpolate(pos_emb.float(), size=new_size // patch_size, mode='bicubic', align_corners=False)
        pos_emb = pos_emb.to(cls_emb.dtype).reshape(1, embed_dim, -1).permute(0, 2, 1)
        pos_emb = torch.cat([cls_emb, pos_emb], dim=1)
        self.embeddings.position_embedding = nn.Parameter(pos_emb)
        logger.info('Resized position embeddings from {} to {}'.format(old_size, new_size))

    def get_input_embeddings(self):
        return self.embeddings

    def forward(
            self,
            pixel_values: Optional[torch.FloatTensor] = None,
            output_hidden_states: Optional[bool] = None,
            return_dict: Optional[bool] = None,
            pixel_embeds: Optional[torch.FloatTensor] = None,
    ) -> Union[Tuple, BaseModelOutputWithPooling]:
        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

        if pixel_values is None and pixel_embeds is None:
            raise ValueError('You have to specify pixel_values or pixel_embeds')

        if pixel_embeds is not None:
            hidden_states = pixel_embeds
        else:
            if len(pixel_values.shape) == 4:
                hidden_states = self.embeddings(pixel_values)
            else:
                raise ValueError(f'wrong pixel_values size: {pixel_values.shape}')
        encoder_outputs = self.encoder(
            inputs_embeds=hidden_states,
            output_hidden_states=output_hidden_states,
            return_dict=return_dict,
        )
        last_hidden_state = encoder_outputs.last_hidden_state
        pooled_output = last_hidden_state[:, 0, :]

        if not return_dict:
            return (last_hidden_state, pooled_output) + encoder_outputs[1:]

        return BaseModelOutputWithPooling(
            last_hidden_state=last_hidden_state,
            pooler_output=pooled_output,
            hidden_states=encoder_outputs.hidden_states,
            attentions=encoder_outputs.attentions,
        )
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.
    """
    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(original_height * scale_factor)
        padding = (current_height - new_height) // 2
        unpadded_tensor = tensor[:, padding : current_height - padding, :]
    else:
        scale_factor = current_height / original_height
        new_width = int(original_width * scale_factor)
        padding = (current_width - new_width) // 2
        unpadded_tensor = tensor[:, :, padding : current_width - padding]

    return unpadded_tensor


@dataclass
# Copied from transformers.models.idefics.modeling_idefics.IdeficsCausalLMOutputWithPast with Idefics->OmChat
class OmChatCausalLMOutputWithPast(ModelOutput):
    """
    Base class for OmChat 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 (`tuple(torch.FloatTensor)`, *optional*):
            Tuple of `torch.FloatTensor` (one for the output of the image embeddings, `(batch_size, num_images,
            sequence_length, hidden_size)`.

            image_hidden_states of the model produced by the vision encoder, and optionally by the perceiver
    """

    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[Tuple[torch.FloatTensor]] = None


# Copied from transformers.models.llava.modeling_llava.LlavaMultiModalProjector with Llava->OmChat
class OmChatMultiModalProjector(nn.Module):
    def __init__(self, config: OmChatConfig):
        super().__init__()

        self.linear_1 = nn.Linear(config.vision_config.hidden_size, config.text_config.hidden_size, bias=True)
        self.act = nn.GELU()
        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

OMCHAT_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 ([`OmChatConfig`] or [`OmChatVisionConfig`]):
            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.",
    OMCHAT_START_DOCSTRING,
)
# Copied from transformers.models.llava.modeling_llava.LlavaPreTrainedModel with Llava->OmChat,llava->omchat
class OmChatPreTrainedModel(PreTrainedModel):
    config_class = OmChatConfig
    base_model_prefix = "model"
    supports_gradient_checkpointing = True
    _no_split_modules = ["OmChatVisionAttention"]
    _skip_keys_device_placement = "past_key_values"
    _supports_flash_attn_2 = True
    _supports_cache_class = True

    def _init_weights(self, module):
        # important: this ported version of OmChat 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/omchat 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_()

    @property
    def _supports_sdpa(self):
        """
        Retrieve language_model's attribute to check whether the model supports
        SDPA or not.
        """
        return self.language_model._supports_sdpa


OMCHAT_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 [`OmChatImageProcessor.__call__`] for details. [`LlavaProcessor`] uses
            [`OmChatImageProcessor`] 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.
"""


@add_start_docstrings(
    """The OmChat model which consists of a vision backbone and a language model.""",
    OMCHAT_START_DOCSTRING,
)
class OmChatForConditionalGeneration(OmChatPreTrainedModel):
    def __init__(self, config: OmChatConfig):
        super().__init__(config)
        self.vision_tower = InternVisionModel(InternVisionConfig())

        self.multi_modal_projector = OmChatMultiModalProjector(config)
        self.vocab_size = config.text_config.vocab_size
        self.language_model = Qwen2ForCausalLM._from_config(
            config.text_config, attn_implementation=config._attn_implementation
        )
        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.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

    # Copied from transformers.models.llava.modeling_llava.LlavaForConditionalGeneration.get_input_embeddings
    def get_input_embeddings(self):
        return self.language_model.get_input_embeddings()

    # Copied from transformers.models.llava.modeling_llava.LlavaForConditionalGeneration.set_input_embeddings
    def set_input_embeddings(self, value):
        self.language_model.set_input_embeddings(value)

    # Copied from transformers.models.llava.modeling_llava.LlavaForConditionalGeneration.get_output_embeddings
    def get_output_embeddings(self):
        return self.language_model.get_output_embeddings()

    # Copied from transformers.models.llava.modeling_llava.LlavaForConditionalGeneration.set_output_embeddings
    def set_output_embeddings(self, new_embeddings):
        self.language_model.set_output_embeddings(new_embeddings)

    # Copied from transformers.models.llava.modeling_llava.LlavaForConditionalGeneration.set_decoder
    def set_decoder(self, decoder):
        self.language_model.set_decoder(decoder)

    # Copied from transformers.models.llava.modeling_llava.LlavaForConditionalGeneration.get_decoder
    def get_decoder(self):
        return self.language_model.get_decoder()

    # Copied from transformers.models.llava.modeling_llava.LlavaForConditionalGeneration.tie_weights
    def tie_weights(self):
        return self.language_model.tie_weights()

    # Copied from transformers.models.llava.modeling_llava.LlavaForConditionalGeneration.resize_token_embeddings
    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 get_vision_tower(self):
        if isinstance(self.vision_tower, list):
            return self.vision_tower[0]
        return self.vision_tower
    
    def get_model(self):
        return self.language_model.model

    def encode_images(self, images):
        vision_tower = self.get_vision_tower()
        image_features = self.vision_tower_forward(images)
        return self.multi_modal_projector(image_features.to(torch.float16))

    def feature_select(self, image_forward_outs):
        image_features = image_forward_outs.hidden_states[-1]
        image_features = image_features[:, 1:]
        return image_features
    
    def vision_tower_forward(self, images):        
        if type(images) is list:
            image_features = []
            for image in images:
                image_forward_out = self.vision_tower(image.to(device=self.device, dtype=self.dtype).unsqueeze(0), output_hidden_states=True)
                image_feature = self.feature_select(image_forward_out).to(image.dtype)
                image_features.append(image_feature)
        else:
            image_forward_outs = self.vision_tower(images.to(device=self.device, dtype=torch.float16), output_hidden_states=True)
            #image_forward_outs = self.vision_tower(images, output_hidden_states=True)
            image_features = self.feature_select(image_forward_outs)

        return image_features

    def prepare_inputs_labels_for_multimodal(
        self, input_ids, position_ids, attention_mask, past_key_values, labels, images
    ):

        vision_tower = self.get_vision_tower()
        video_tower = self.get_vision_tower()
        if (vision_tower is None and video_tower is None) or images is None or input_ids.shape[1] == 1:
            if past_key_values is not None and (vision_tower is not None or video_tower is not None) and images is not None and input_ids.shape[1] == 1:
                target_shape = past_key_values[-1][-1].shape[-2] + 1
                attention_mask = torch.cat((attention_mask, torch.ones(
                    (attention_mask.shape[0], target_shape - attention_mask.shape[1]),
                    dtype=attention_mask.dtype,
                    device=attention_mask.device
                )), dim=1)
                position_ids = torch.sum(attention_mask, dim=1).unsqueeze(-1) - 1
            return input_ids, position_ids, attention_mask, past_key_values, None, labels

        image_idx = [idx for idx, img in enumerate(images) if img.ndim == 3]
        is_all_image = len(image_idx) == len(images)
        video_idx = [idx for idx, vid in enumerate(images) if vid.ndim == 4]
        images_minibatch = torch.stack([images[idx] for idx in image_idx]) if len(image_idx) > 0 else []  # mini_b c h w
        videos_minibatch = torch.stack([images[idx] for idx in video_idx]) if len(video_idx) > 0 else []  # mini_b c t h w

        tmp_image_features = [None] * (len(image_idx) + len(video_idx))
        if getattr(images_minibatch, 'ndim', 0) == 4:  # batch consists of images, [mini_b, c, h, w]
            if vision_tower is not None:
                image_features_minibatch = self.encode_images(images_minibatch)  # [mini_b, l, c]
            else:
                image_features_minibatch = torch.randn(1).to(self.device)  # dummy feature for video-only training under tuning
            for i, pos in enumerate(image_idx):
                tmp_image_features[pos] = image_features_minibatch[i]
        if getattr(videos_minibatch, 'ndim', 0) == 5:  # batch consists of videos, [mini_b, c, t, h, w]
            video_features_minibatch = self.encode_images(videos_minibatch)  # fake list [mini_b, t, l, c]
            for i, pos in enumerate(video_idx):
                tmp_image_features[pos] = video_features_minibatch[i]
        new_tmp = []
        for image in tmp_image_features:
            if isinstance(image, list):
                t = len(image)
                for i in range(t):
                    new_tmp.append(image[i])
            else:
                new_tmp.append(image)
        image_features = new_tmp

        if getattr(self.config, 'tune_mm_mlp_adapter', False) and getattr(self.config, 'mm_use_im_start_end', False):
            raise NotImplementedError

        _labels = labels
        _position_ids = position_ids
        _attention_mask = attention_mask
        if attention_mask is None:
            attention_mask = torch.ones_like(input_ids, dtype=torch.bool)
        else:
            attention_mask = attention_mask.bool()
        if position_ids is None:
            position_ids = torch.arange(0, input_ids.shape[1], dtype=torch.long, device=input_ids.device)
        if labels is None:
            labels = torch.full_like(input_ids, -100)

        input_ids = [cur_input_ids[cur_attention_mask] for cur_input_ids, cur_attention_mask in zip(input_ids, attention_mask)]
        labels = [cur_labels[cur_attention_mask] for cur_labels, cur_attention_mask in zip(labels, attention_mask)]
        new_input_embeds = []
        new_labels = []
        cur_image_idx = 0
        for batch_idx, cur_input_ids in enumerate(input_ids):
            num_images = (cur_input_ids == -200).sum()
            if num_images == 0:
                cur_image_features = image_features[cur_image_idx]
                cur_input_embeds_1 = self.get_model().embed_tokens(cur_input_ids)
                cur_input_embeds = torch.cat([cur_input_embeds_1, cur_image_features[0:0]], dim=0)
                new_input_embeds.append(cur_input_embeds)
                new_labels.append(labels[batch_idx])
                cur_image_idx += 1
                continue

            image_token_indices = [-1] + torch.where(cur_input_ids == -200)[0].tolist() + [cur_input_ids.shape[0]]
            cur_input_ids_noim = []
            cur_labels = labels[batch_idx]
            cur_labels_noim = []
            for i in range(len(image_token_indices) - 1):
                cur_input_ids_noim.append(cur_input_ids[image_token_indices[i]+1:image_token_indices[i+1]])
                cur_labels_noim.append(cur_labels[image_token_indices[i]+1:image_token_indices[i+1]])
            split_sizes = [x.shape[0] for x in cur_labels_noim]
            cur_input_embeds = self.get_model().embed_tokens(torch.cat(cur_input_ids_noim))
            cur_input_embeds_no_im = torch.split(cur_input_embeds, split_sizes, dim=0)

            cur_new_input_embeds = []
            cur_new_labels = []

            for i in range(num_images + 1):
                cur_new_input_embeds.append(cur_input_embeds_no_im[i])
                cur_new_labels.append(cur_labels_noim[i])
                if i < num_images:
                    cur_image_features = image_features[cur_image_idx].to(self.device)
                    cur_image_idx += 1
                    cur_new_input_embeds.append(cur_image_features)
                    cur_new_labels.append(torch.full((cur_image_features.shape[0],), -100, device=cur_labels.device, dtype=cur_labels.dtype))

            cur_new_input_embeds = torch.cat(cur_new_input_embeds)
            cur_new_labels = torch.cat(cur_new_labels)

            new_input_embeds.append(cur_new_input_embeds)
            new_labels.append(cur_new_labels)

        # Truncate sequences to max length as image embeddings can make the sequence longer
        tokenizer_model_max_length = getattr(self.config, 'tokenizer_model_max_length', None)
        if tokenizer_model_max_length is not None:
            new_input_embeds = [x[:tokenizer_model_max_length] for x in new_input_embeds]
            new_labels = [x[:tokenizer_model_max_length] for x in new_labels]

        max_len = max(x.shape[0] for x in new_input_embeds)
        batch_size = len(new_input_embeds)

        new_input_embeds_padded = []
        new_labels_padded = torch.full((batch_size, max_len), -100, dtype=new_labels[0].dtype, device=new_labels[0].device)
        attention_mask = torch.zeros((batch_size, max_len), dtype=attention_mask.dtype, device=attention_mask.device)
        position_ids = torch.zeros((batch_size, max_len), dtype=position_ids.dtype, device=position_ids.device)

        for i, (cur_new_embed, cur_new_labels) in enumerate(zip(new_input_embeds, new_labels)):
            cur_len = cur_new_embed.shape[0]
            if getattr(self.config, 'tokenizer_padding_side', 'right') == "left":
                new_input_embeds_padded.append(torch.cat((
                    torch.zeros((max_len - cur_len, cur_new_embed.shape[1]), dtype=cur_new_embed.dtype, device=cur_new_embed.device),
                    cur_new_embed
                ), dim=0))
                if cur_len > 0:
                    new_labels_padded[i, -cur_len:] = cur_new_labels
                    attention_mask[i, -cur_len:] = True
                    position_ids[i, -cur_len:] = torch.arange(0, cur_len, dtype=position_ids.dtype, device=position_ids.device)
            else:
                new_input_embeds_padded.append(torch.cat((
                    cur_new_embed,
                    torch.zeros((max_len - cur_len, cur_new_embed.shape[1]), dtype=cur_new_embed.dtype, device=cur_new_embed.device)
                ), dim=0))
                if cur_len > 0:
                    new_labels_padded[i, :cur_len] = cur_new_labels
                    attention_mask[i, :cur_len] = True
                    position_ids[i, :cur_len] = torch.arange(0, cur_len, dtype=position_ids.dtype, device=position_ids.device)

        new_input_embeds = torch.stack(new_input_embeds_padded, dim=0)
        if _labels is None:
            new_labels = None
        else:
            new_labels = new_labels_padded

        if _attention_mask is None:
            attention_mask = None
        else:
            attention_mask = attention_mask.to(dtype=_attention_mask.dtype)

        if _position_ids is None:
            position_ids = None

        return None, position_ids, attention_mask, past_key_values, new_input_embeds, new_labels


    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

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

            # 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, 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).
            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 height * width != 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

    @add_start_docstrings_to_model_forward(OMCHAT_INPUTS_DOCSTRING)
    @replace_return_docstrings(output_type=OmChatCausalLMOutputWithPast, config_class=_CONFIG_FOR_DOC)
    def forward(
        self,
        input_ids: 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,
        images: Optional[torch.FloatTensor] = None,
        return_dict: Optional[bool] = None,
    ) -> Union[Tuple, OmChatCausalLMOutputWithPast]:
        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]`.

        Returns:

        Example:

        ```python
        >>> from PIL import Image
        >>> import requests
        >>> from transformers import AutoProcessor, OmChatForConditionalGeneration

        >>> model = OmChatForConditionalGeneration.from_pretrained("llava-hf/llava-v1.6-mistral-7b-hf")
        >>> processor = AutoProcessor.from_pretrained("llava-hf/llava-v1.6-mistral-7b-hf")

        >>> prompt = "[INST] <image>\nWhat is shown in this image? [/INST]"
        >>> url = "https://www.ilankelman.org/stopsigns/australia.jpg"
        >>> image = Image.open(requests.get(url, stream=True).raw)

        >>> inputs = processor(text=prompt, images=image, return_tensors="pt")

        >>> # Generate
        >>> generate_ids = model.generate(**inputs, max_length=30)
        >>> processor.batch_decode(generate_ids, skip_special_tokens=True, clean_up_tokenization_spaces=False)[0]
        "[INST]  \nWhat is shown in this image? [/INST] The image appears to be a radar chart, which is a type of multi-dimensional plot (...)"
        ```"""

        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_layer = (
            vision_feature_layer if vision_feature_layer is not None else self.config.vision_feature_layer
        )
        vision_feature_select_strategy = (
            vision_feature_select_strategy
            if vision_feature_select_strategy is not None
            else self.config.vision_feature_select_strategy
        )
        if inputs_embeds is None:
            (
                input_ids,
                position_ids,
                attention_mask,
                past_key_values,
                inputs_embeds,
                labels
            ) = self.prepare_inputs_labels_for_multimodal(
                input_ids,
                position_ids,
                attention_mask,
                past_key_values,
                labels,
                images
            )
        outputs = self.language_model(
            input_ids=input_ids,
            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
        )
        return outputs
        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[..., 1:]
                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 OmChatCausalLMOutputWithPast(
            loss=loss,
            logits=logits,
            past_key_values=outputs.past_key_values,
            hidden_states=outputs.hidden_states,
            attentions=outputs.attentions,
        )

    def prepare_inputs_for_generation(
            self, input_ids, past_key_values=None, attention_mask=None, inputs_embeds=None, **kwargs
    ):
        if past_key_values:
            input_ids = input_ids[:, -1:]

        if inputs_embeds is not None and past_key_values is None:
            model_inputs = {"inputs_embeds": inputs_embeds}
        else:
            model_inputs = {"input_ids": input_ids}

        model_inputs.update(
            {
                "past_key_values": past_key_values,
                "use_cache": kwargs.get("use_cache"),
                "attention_mask": attention_mask,
                "images": kwargs.get("images", None),
            }
        )
        return model_inputs


    # Copied from transformers.models.llava.modeling_llava.LlavaForConditionalGeneration._reorder_cache
    def _reorder_cache(self, *args, **kwargs):
        return self.language_model._reorder_cache(*args, **kwargs)