modeling_centurio.py

# ADAPTED FROM https://raw.githubusercontent.com/huggingface/transformers/main/src/transformers/models/llava/modeling_llava.py
# coding=utf-8
# Copyright 2023 the HuggingFace Inc. team. All rights reserved.
#
# Licensed under the Apache License, Version 2.0 (the "License");
# you may not use this file except in compliance with the License.
# You may obtain a copy of the License at
#
#     http://www.apache.org/licenses/LICENSE-2.0
#
# Unless required by applicable law or agreed to in writing, software
# distributed under the License is distributed on an "AS IS" BASIS,
# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
# See the License for the specific language governing permissions and
# limitations under the License.
""" PyTorch Llava model."""
import math

import logging
from dataclasses import dataclass
from functools import partial
from typing import List, Optional, Tuple, Union

import timm
import torch
import torch.utils.checkpoint
from torch import nn
from transformers import LlavaConfig, PreTrainedModel, add_start_docstrings, AutoModel, AutoModelForCausalLM, Cache, \
    T5ForConditionalGeneration, HybridCache, Gemma2ForCausalLM
from transformers.utils import ModelOutput, add_start_docstrings_to_model_forward, replace_return_docstrings

from transformers import LlavaConfig
from transformers.activations import ACT2FN
import torch
from einops import rearrange, repeat
from torch import einsum, nn

from .configuration_centurio import CenturioConfig

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

        self.linear_1 = nn.Linear(config.image_hidden_size, config.text_config.hidden_size, bias=True)
        self.act = ACT2FN["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

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

        if config.adapter_type == "window-pool":
            self.adapter = WindowPoolProjector(config)
        elif config.adapter_type == "window-shuffel":
            self.adapter = WindowShuffelProjector(config)
        elif config.adapter_type == "multiscale-pool":
            self.adapter = MultiscalePoolProjector(config)
        elif config.adapter_type == "multiscale-shuffel":
            self.adapter = MultiscaleShuffleProjector(config)
        else:
            self.adapter = LlavaMLPProjector(config)

    def forward(self, image_features):
        return self.adapter(image_features)



class WindowMLPProjector(nn.Module):
    def __init__(self, config: LlavaConfig):
        super().__init__()
        self.multi_scale = config.adapter_config.get("multi_scale", 2) #config.adapter_config.get("multi_scale")
        self.linear_1 = nn.Linear(config.image_hidden_size, config.text_config.hidden_size, bias=True)
        self.act = ACT2FN["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)

        windows = 1 + self.multi_scale**2
        hidden_states = rearrange(hidden_states, "(b h) w d -> b (h w) d", h=windows)

        return hidden_states


class WindowPoolProjector(nn.Module):
    def __init__(self, config: LlavaConfig):
        super().__init__()
        self.multi_scale = config.adapter_config.get("multi_scale", 2) #config.adapter_config.get("multi_scale")
        self.pool = nn.AdaptiveAvgPool2d(getattr(config, "adapter_pool", 8))
        self.linear_1 = nn.Linear(config.image_hidden_size, config.text_config.hidden_size, bias=True)
        self.act = ACT2FN["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)

        b, num_tokens, c = hidden_states.shape
        h = int(math.sqrt(num_tokens))

        hidden_states = rearrange(hidden_states, "b (h w) d -> b d h w", h=h, w=h)
        hidden_states = self.pool(hidden_states)
        hidden_states = rearrange(hidden_states, "b d h w -> b (h w) d")

        windows = 1 + self.multi_scale**2
        hidden_states = rearrange(hidden_states, "(b h) w d -> b (h w) d", h=windows)
        return hidden_states


class WindowShuffelProjector(nn.Module):
    def __init__(self, config: LlavaConfig):
        super().__init__()
        self.multi_scale = config.adapter_config.get("multi_scale", 2) #config.adapter_config.get("multi_scale")
        self.scale_factor = getattr(config, "adapter_pool", 2)
        self.pixel_unshuffel = nn.PixelUnshuffle(self.scale_factor)
        self.linear_1 = nn.Linear(config.image_hidden_size*(self.scale_factor**2), config.text_config.hidden_size, bias=True)
        self.act = ACT2FN["gelu"]
        self.linear_2 = nn.Linear(config.text_config.hidden_size, config.text_config.hidden_size, bias=True)



    def forward(self, image_features):
        bsz, seq, embed_dim = image_features.size()
        height = width = int(seq ** 0.5)
        hidden_states = rearrange(image_features, "b (w h) d -> b d w h", w=width, h=height)
        hidden_states = self.pixel_unshuffel(hidden_states)
        hidden_states = rearrange(hidden_states, "b d w h -> b (w h) d")

        hidden_states = self.linear_1(hidden_states)
        hidden_states = self.act(hidden_states)
        hidden_states = self.linear_2(hidden_states)

        windows = 1 + self.multi_scale ** 2
        hidden_states = rearrange(hidden_states, "(b h) w d -> b (h w) d", h=windows)
        return hidden_states


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

        self.multi_scale = config.adapter_config.get("multi_scale", 2) #getattr(config.adapter_config, "adapter_multi_scale", 2)
        self.pool = nn.AvgPool2d(self.multi_scale)
        self.linear_1 = nn.Linear(config.image_hidden_size*2, config.text_config.hidden_size, bias=True)
        self.act = ACT2FN["gelu"]
        self.linear_2 = nn.Linear(config.text_config.hidden_size, config.text_config.hidden_size, bias=True)

    def forward(self, image_features):
        b, num_tokens, c = image_features.shape
        h = int(math.sqrt(num_tokens))
        assert h * h == num_tokens
        image_features = rearrange(image_features, "b (h w) d -> b d h w", h=h, w=h)

        steps = 1 + self.multi_scale**2
        low_res_features = image_features[::steps]
        high_res_features = image_features[[i for i in range(image_features.size(0)) if i%steps > 0]]

        merged_features = rearrange(high_res_features, "(b m) d h w -> b d h (m w)", m=self.multi_scale)
        merged_features = rearrange(merged_features, "(b m) d h w -> b d (m h) w", m=self.multi_scale)

        merged_features = self.pool(merged_features)

        concat_features = torch.cat([low_res_features, merged_features], dim=1)
        concat_features = rearrange(concat_features, "b d h w -> b (h w) d")

        hidden_states = self.linear_1(concat_features)
        hidden_states = self.act(hidden_states)
        hidden_states = self.linear_2(hidden_states)
        return hidden_states

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

        self.multi_scale = config.adapter_config.get("multi_scale", 2) #config.adapter_config.get("multi_scale")
        self.shuffle = nn.PixelUnshuffle(self.multi_scale)

        inc, ouc = config.image_hidden_size*(1+self.multi_scale**2), config.text_config.hidden_size
        #
        self.mlp = nn.Sequential(
            nn.Linear(inc, ouc), nn.GELU(), nn.Linear(ouc, ouc)
        )

        self.dwn = nn.AvgPool2d(2) #TokenDownLayer((12, 12))
        self.peg = nn.Conv2d(ouc, ouc, 3, 1, 1, bias=True, groups=ouc) #PosInjectLayer(ouc, ouc, stride=1)

    def forward(self, x):
        b, num_tokens, c = x.shape
        h = int(math.sqrt(num_tokens))
        assert h * h == num_tokens
        image_features = rearrange(x, "b (h w) d -> b d h w", h=h, w=h)

        steps = 1 + self.multi_scale ** 2
        low_res_features = image_features[::steps]
        high_res_features = image_features[[i for i in range(image_features.size(0)) if i % steps > 0]]

        merged_features = rearrange(high_res_features, "(b m) d h w -> b d h (m w)", m=self.multi_scale)
        merged_features = rearrange(merged_features, "(b m) d h w -> b d (m h) w", m=self.multi_scale)

        merged_features = self.shuffle(merged_features)

        concat_features = torch.cat([low_res_features, merged_features], dim=1)
        concat_features = rearrange(concat_features, "b d h w -> b (h w) d")

        x = self.mlp(concat_features)

        # x = self.dwn(x)
        b, num_tokens, c = x.shape
        h = int(math.sqrt(num_tokens))
        assert h * h == num_tokens
        x = rearrange(x, "b (h w) d -> b d h w", h=h, w=h) #x.permute(0, 2, 1).reshape(b, -1, h, h)
        x = self.dwn(x)
        x = self.peg(x) + x
        x = rearrange(x, "b d h w -> b (h w) d") #x.flatten(2).transpose(1, 2)

        return x
#

_CONFIG_FOR_DOC = "LlavaConfig"

LLAVA_PRETRAINED_MODEL_ARCHIVE_LIST = [
    "llava-hf/llava-1.5-7b-hf",
    "llava-hf/llava-1.5-13b-hf",
    "llava-hf/bakLlava-v1-hf",
    # See all Llava models at https://huggingface.co/models?filter=llava
]


@dataclass
# Copied from transformers.models.idefics.modeling_idefics.IdeficsCausalLMOutputWithPast with Idefics->Llava
class LlavaCausalLMOutputWithPast(ModelOutput):
    """
    Base class for Llava 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
    labels: Optional[torch.LongTensor] = None



LLAVA_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 ([`LlavaConfig`] or [`LlavaVisionConfig`]):
            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_START_DOCSTRING,
)
class LlavaPreTrainedModel(PreTrainedModel):
    config_class = LlavaConfig
    base_model_prefix = "model"
    supports_gradient_checkpointing = True
    _no_split_modules = ["LlavaVisionAttention"]
    _skip_keys_device_placement = "past_key_values"
    _supports_flash_attn_2 = True

    def _init_weights(self, module):
        # important: this ported version of Llava 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 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


LLAVA_INPUTS_DOCSTRING = r"""
    Args:
        input_ids (`torch.LongTensor` of shape `(batch_size, sequence_length)`):
            Indices of input sequence tokens in the vocabulary. Padding will be ignored by default should you provide
            it.

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

            [What are input IDs?](../glossary#input-ids)
        pixel_values (`torch.FloatTensor` of shape `(batch_size, num_channels, image_size, image_size)):
            The tensors corresponding to the input images. Pixel values can be obtained using
            [`AutoImageProcessor`]. See [`CLIPImageProcessor.__call__`] for details ([]`LlavaProcessor`] uses
            [`CLIPImageProcessor`] for processing images).
        attention_mask (`torch.Tensor` of shape `(batch_size, sequence_length)`, *optional*):
            Mask to avoid performing attention on padding token indices. Mask values selected in `[0, 1]`:

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

            [What are attention masks?](../glossary#attention-mask)

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

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

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

            - 1 indicates the head is **not masked**,
            - 0 indicates the head is **masked**.
        position_ids (`torch.LongTensor` of shape `(batch_size, sequence_length)`, *optional*):
            Indices of positions of each input sequence tokens in the position embeddings. Selected in the range `[0,
            config.n_positions - 1]`. [What are position IDs?](../glossary#position-ids)
        past_key_values (`tuple(tuple(torch.FloatTensor))`, *optional*, returned when `use_cache=True` is passed or when `config.use_cache=True`):
            Tuple of `tuple(torch.FloatTensor)` of length `config.n_layers`, with each tuple having 2 tensors of shape
            `(batch_size, num_heads, sequence_length, embed_size_per_head)`) and 2 additional tensors of shape
            `(batch_size, num_heads, encoder_sequence_length, embed_size_per_head)`.

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

            If `past_key_values` are used, the user can optionally input only the last `decoder_input_ids` (those that
            don't have their past key value states given to this model) of shape `(batch_size, 1)` instead of all
            `decoder_input_ids` of shape `(batch_size, sequence_length)`.
        inputs_embeds (`torch.FloatTensor` of shape `(batch_size, sequence_length, hidden_size)`, *optional*):
            Optionally, instead of passing `input_ids` you can choose to directly pass an embedded representation. This
            is useful if you want more control over how to convert `input_ids` indices into associated vectors than the
            model's internal embedding lookup matrix.
        use_cache (`bool`, *optional*):
            If set to `True`, `past_key_values` key value states are returned and can be used to speed up decoding (see
            `past_key_values`).
        output_attentions (`bool`, *optional*):
            Whether or not to return the attentions tensors of all attention layers. See `attentions` under returned
            tensors for more detail.
        output_hidden_states (`bool`, *optional*):
            Whether or not to return the hidden states of all layers. See `hidden_states` under returned tensors for
            more detail.
        return_dict (`bool`, *optional*):
            Whether or not to return a [`~utils.ModelOutput`] instead of a plain tuple.
"""


class CenturioForConditionalGeneration(LlavaPreTrainedModel):
    config_class = CenturioConfig
    _supports_cache_class = True
    _supports_quantized_cache = False
    _supports_static_cache = True

    def __init__(self, config: CenturioConfig):
        super().__init__(config)
        # self.vision_tower = AutoModel.from_config(config.vision_config)
        self.vision_tower = timm.create_model(
            config.timm_model,
            pretrained=False,
            num_classes=0,
        )
        # https://github.com/TRI-ML/prismatic-vlms/blob/main/prismatic/models/backbones/vision/base_vision.py#L125
        def unpack_tuple(fn):
            def wrapper(*args, **kwargs):
                result = fn(*args, **kwargs)
                return result[0] if isinstance(result, tuple) or isinstance(result, list) else result

            return wrapper
        self.vision_tower.forward = unpack_tuple(
            partial(
                self.vision_tower.get_intermediate_layers, n={len(self.vision_tower.blocks) - 2}
            )
        )

        config.image_hidden_size = self.vision_tower.embed_dim

        self.multi_modal_projector = LlavaMultiModalAdapter(config)
        self.vocab_size = config.text_config.vocab_size
        # if getattr(config, "delay_init", False):
        #     self.language_model = None
        # else:
        self.language_model = AutoModelForCausalLM.from_config(
            config.text_config, attn_implementation=config._attn_implementation, torch_dtype=config.torch_dtype,
            trust_remote_code = True
        )
        self.pad_token_id = self.config.pad_token_id if self.config.pad_token_id is not None else -1
        self.post_init()

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

    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 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.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, inputs_embeds, input_ids, attention_mask, labels):
        num_images, num_image_patches, embed_dim = image_features.shape
        batch_size, sequence_length = input_ids.shape
        left_padding = not torch.sum(input_ids[:, -1] == torch.tensor(self.pad_token_id))
        # 1. Create a mask to know where special image tokens are
        special_image_token_mask = input_ids == self.config.image_token_index
        num_special_image_tokens = torch.sum(special_image_token_mask, dim=-1)

        #check if preprocessing already expanded the number of <image_token> needed to directly replace them
        if torch.sum(special_image_token_mask) == image_features.shape[:-1].numel():
            new_inputs_embeds = inputs_embeds.clone()
            reshaped_image_hidden_states = image_features.view(-1, embed_dim)
            new_inputs_embeds[special_image_token_mask] = reshaped_image_hidden_states

            position_ids = (attention_mask.cumsum(-1) - 1).masked_fill_((attention_mask == 0), 1)

            return new_inputs_embeds, attention_mask, labels, position_ids


        # Compute the maximum embed dimension
        max_embed_dim = (num_special_image_tokens.max() * (num_image_patches - 1)) + sequence_length
        batch_indices, non_image_indices = torch.where(input_ids != self.config.image_token_index)

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

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

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

        # 5. Fill the embeddings corresponding to the images. Anything that is still zeros needs filling
        ## BUG: this does NOT work for models (Phi-3) that have set some embedding (padding) to be 0. Replaced with the below three lines.
        # image_to_overwrite = torch.all(final_embedding == 0, dim=-1)
        image_to_overwrite = torch.ones_like(final_attention_mask)
        image_to_overwrite[batch_indices, text_to_overwrite] = torch.zeros_like(attention_mask)[batch_indices, non_image_indices]
        image_to_overwrite = image_to_overwrite.bool()
        image_to_overwrite &= image_to_overwrite.cumsum(-1) - 1 >= nb_image_pad[:, None].to(target_device)

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

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

        if labels is None:
            final_labels = None

        return final_embedding, final_attention_mask, final_labels, position_ids

    @add_start_docstrings_to_model_forward(LLAVA_INPUTS_DOCSTRING)
    @replace_return_docstrings(output_type=LlavaCausalLMOutputWithPast, config_class=_CONFIG_FOR_DOC)
    def forward(
        self,
        input_ids: torch.LongTensor = None,
        pixel_values: torch.FloatTensor = None,
        attention_mask: Optional[torch.Tensor] = None,
        position_ids: Optional[torch.LongTensor] = None,
        past_key_values: Optional[List[torch.FloatTensor]] = None,
        inputs_embeds: Optional[torch.FloatTensor] = None,
        labels: Optional[torch.LongTensor] = None,
        use_cache: Optional[bool] = None,
        cache_position: Optional[torch.LongTensor] = None,
        output_attentions: Optional[bool] = None,
        output_hidden_states: Optional[bool] = None,
        return_dict: Optional[bool] = None,
        **kwargs
    ) -> Union[Tuple, LlavaCausalLMOutputWithPast]:
        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:

"""

        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

        if inputs_embeds is None:
            # 1. Extra the input embeddings
            inputs_embeds = self.get_input_embeddings()(input_ids)

            # 2. Merge text and images
            if pixel_values is not None and input_ids.shape[1] != 1:
                image_outputs = self.vision_tower(pixel_values)

                image_features = self.multi_modal_projector(image_outputs)
                image_features = image_features.to(inputs_embeds.dtype)
                inputs_embeds, attention_mask, labels, position_ids = self._merge_input_ids_with_image_features(
                    image_features, inputs_embeds, input_ids, attention_mask, labels
                )
                if labels is None:
                    labels = torch.full_like(attention_mask, self.config.ignore_index).to(torch.long)
            else:
                # In case input_ids.shape[1] == 1 & pixel_values==None & past_key_values != None, we are in the case of
                # generation with cache
                if past_key_values is not None and pixel_values is not None and input_ids.shape[1] == 1:
                    if isinstance(past_key_values, Cache):
                        first_layer_past_key_value = past_key_values.key_cache[0][:, :, :, 0]
                    else:
                        first_layer_past_key_value = past_key_values[0][0][:, :, :, 0]

                    target_seqlen = first_layer_past_key_value.shape[-1] + 1
                    extended_attention_mask = torch.ones(
                        (attention_mask.shape[0], target_seqlen - attention_mask.shape[1]),
                        dtype=attention_mask.dtype,
                        device=attention_mask.device,
                    )
                    attention_mask = torch.cat((attention_mask, extended_attention_mask), 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:
                    # ]

        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,
            # cache_position=cache_position,
            output_attentions=output_attentions,
            output_hidden_states=output_hidden_states,
            return_dict=return_dict,
        )

        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 LlavaCausalLMOutputWithPast(
            loss=loss,
            logits=logits,
            labels=labels,
            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,
        inputs_embeds=None,
        pixel_values=None,
        attention_mask=None,
        cache_position=None,
        use_cache=True,
        position_ids=None,
        **kwargs
    ):
        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,
            **kwargs,
        )
        #Ugly comparison. Should use a config var that knows how many image tokens we have like HF does.
        # But we are unlikely to use >30 images in one sample or use <=30 tokens per image.
        if cache_position[0] == 0:
            model_inputs["pixel_values"] = pixel_values
        # "legacy" mode
        if (input_ids == self.config.image_token_index).sum(1).max() < 30:
            if past_key_values is not None:
                if isinstance(past_key_values, Cache):
                    # branch for Gemma2 with hybrid cache
                    if past_key_values.seen_tokens is None:
                        past_length = cache_position[0] # torch.tensor(0, device=input_ids.device)
                        max_cache_length = (
                            torch.tensor(past_key_values.get_max_length(), device=input_ids.device)
                            if past_key_values.get_max_length() is not None
                            else None
                        )
                        cache_length = past_length if max_cache_length is None else torch.min(max_cache_length, past_length)
                    # old default branch
                    else:
                        cache_length = past_key_values.get_seq_length()
                        past_length = past_key_values.seen_tokens

                else:
                    cache_length = past_length = past_key_values[0][0].shape[2]

                # Keep only the unprocessed tokens:
                # 1 - If the length of the attention_mask exceeds the length of input_ids, then we are in a setting where
                # some of the inputs are exclusively passed as part of the cache (e.g. when passing input_embeds as
                # input)
                if attention_mask is not None and attention_mask.shape[1] > input_ids.shape[1]:
                    input_ids = input_ids[:, -(attention_mask.shape[1] - past_length) :]
                # 2 - If the past_length is smaller than input_ids', then input_ids holds all input tokens. We can discard
                # input_ids based on the past_length.
                elif past_length < input_ids.shape[1]:
                    input_ids = input_ids[:, past_length:]
                # 3 - Otherwise (past_length >= input_ids.shape[1]), let's assume input_ids only has unprocessed tokens.
                elif self.config.image_token_index in input_ids:
                    input_ids = input_ids[:, input_ids.shape[1] - 1 :]
                # If the cache has seen more tokens than it can hold, then the cache has a size limit. Let's discard the
                # older attention values, as their corresponding values are not part of the input.
                # if cache_length < past_length and attention_mask is not None:
                #     attention_mask = attention_mask[:, -(cache_length + input_ids.shape[1]) :]
            if attention_mask is not None and position_ids is None:
                # create position_ids on the fly for batch generation
                position_ids = attention_mask.long().cumsum(-1) - 1
                position_ids.masked_fill_(attention_mask == 0, 1)
                if past_key_values:
                    position_ids = position_ids[:, -input_ids.shape[1] :]

            # if `inputs_embeds` are passed, we only want to use them in the 1st generation step
            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}

            # if cache_position[0] == 0 or (input_ids == self.config.image_token_index).sum(1).max() > 0:
            model_inputs.update(
                {
                    "position_ids": position_ids,
                    "past_key_values": past_key_values,
                    "attention_mask": attention_mask,
                    "use_cache": use_cache,
                    "pixel_values": pixel_values,
                }
            )
        return model_inputs

    def _reorder_cache(self, *args, **kwargs):
        return self.language_model._reorder_cache(*args, **kwargs)