modeling_visual_language.py

import copy
import enum
import inspect
import logging
import math
import os
import warnings
from abc import abstractmethod
from dataclasses import dataclass
from pathlib import Path
from typing import TYPE_CHECKING, Any, Dict, List, Optional, Tuple, Union

import numpy as np
import openvino
import openvino as ov
import torch
from huggingface_hub import hf_hub_download
from huggingface_hub.constants import HUGGINGFACE_HUB_CACHE
from openvino._offline_transformations import apply_moc_transformations, compress_model_transformation
from transformers import (
    AutoConfig,
    AutoImageProcessor,
    GenerationConfig,
    GenerationMixin,
    PretrainedConfig,
    PreTrainedTokenizer,
)
from transformers.modeling_outputs import BaseModelOutputWithPooling
from transformers.models.qwen2_5_vl.modeling_qwen2_5_vl import Qwen2_5_VLModel
from transformers.models.qwen2_vl.modeling_qwen2_vl import Qwen2VLModel, VisionRotaryEmbedding
from transformers.models.qwen3_vl.modeling_qwen3_vl import (
    Qwen3VLModel,
    Qwen3VLVisionModel,
    Qwen3VLVisionRotaryEmbedding,
)
from transformers.utils import ModelOutput

from optimum.exporters.openvino import main_export
from optimum.exporters.openvino.stateful import ensure_stateful_is_available, model_has_input_output_name
from optimum.exporters.openvino.utils import save_config
from optimum.intel.openvino.configuration import OVConfig, OVQuantizationConfigBase, OVWeightQuantizationConfig
from optimum.intel.openvino.modeling_base import OVBaseModel, OVModelPart
from optimum.intel.openvino.modeling_decoder import CausalLMOutputWithPast, OVModelForCausalLM
from optimum.intel.openvino.utils import (
    OV_LANGUAGE_MODEL_NAME,
    OV_TEXT_EMBEDDINGS_MODEL_NAME,
    OV_VISION_EMBEDDINGS_MODEL_NAME,
    TemporaryDirectory,
    classproperty,
)
from optimum.intel.utils.import_utils import is_transformers_version


if is_transformers_version(">=", "4.46.0"):
    from transformers import AutoModelForImageTextToText

    transformers_auto_class = AutoModelForImageTextToText
else:
    from transformers import AutoModelForVision2Seq

    transformers_auto_class = AutoModelForVision2Seq


if TYPE_CHECKING:
    from PIL.Image import Image
    from transformers.image_utils import VideoInput

logger = logging.getLogger(__name__)

core = ov.Core()


class InputMode(enum.Enum):
    LANGUAGE = 0
    VISION = 1
    SPEECH = 2
    VISION_SPEECH = 3


class OVModelWithEmbedForCausalLM(OVModelForCausalLM):
    def __init__(
        self,
        model: ov.Model,
        text_embeds_model: ov.Model,
        config: PretrainedConfig = None,
        device: str = "CPU",
        dynamic_shapes: bool = None,
        ov_config: Optional[Dict[str, str]] = None,
        model_save_dir: Optional[Union[str, Path, TemporaryDirectory]] = None,
        quantization_config: Optional[Union[OVWeightQuantizationConfig, Dict]] = None,
        **kwargs,
    ):
        self.model = model
        self.text_emb_model = text_embeds_model
        self.request = None
        self.text_emb_request = None
        compile_only = kwargs.get("compile_only", False)
        if compile_only:
            self.text_emb_request = self.text_emb_model
            self.request = self.model.create_infer_request()

        super().__init__(
            model=model,
            config=config,
            device=device,
            dynamic_shapes=dynamic_shapes,
            ov_config=ov_config,
            model_save_dir=model_save_dir,
            quantization_config=quantization_config,
            **kwargs,
        )

    @property
    def _ov_model_names(self) -> List[str]:
        return ["model", "text_emb_model"]

    def compile(self):
        if self.request is None:
            logger.info(f"Compiling the Language model to {self._device} ...")
            super().compile()
        self._compile_text_emb()

    def _compile_text_emb(self):
        if self.text_emb_request is None:
            logger.info(f"Compiling the Text embeddings model to {self._device} ...")
            if self._compile_only:
                self.text_emb_request = self.text_emb_model
            else:
                logger.info(f"Compiling the Text embeddings model to {self._device} ...")
                self.text_emb_request = self._compile_model(
                    self.text_emb_model, self._device, self.ov_config, self.model_save_dir
                )

    def clear_requests(self):
        if self._compile_only:
            raise ValueError(
                "`clear_requests()` is not supported with `compile_only` mode, please initialize model without this option"
            )
        self.request = None
        self.text_emb_request = None

    def embed_tokens(self, input_ids: torch.LongTensor):
        self._compile_text_emb()
        res = self.text_emb_request(input_ids, share_inputs=True)
        return res[0]

    def prepare_inputs(
        self,
        input_ids: torch.LongTensor,
        attention_mask: Optional[torch.LongTensor] = None,
        past_key_values: Optional[Tuple[Tuple[torch.FloatTensor]]] = None,
        position_ids: Optional[torch.LongTensor] = None,
        inputs_embeds: Optional[torch.FloatTensor] = None,
        token_type_ids: Optional[torch.LongTensor] = None,
        visual_pos_masks: Optional[torch.FloatTensor] = None,
        deepstack_visual_embeds: Optional[torch.FloatTensor] = None,
        **kwargs,
    ):
        batch_size = input_ids.shape[0] if input_ids is not None else inputs_embeds.shape[0]

        inputs = {}
        # past_key_values are not used explicitly, instead they are handled inside the model
        if past_key_values is None:
            # This is the first iteration in a sequence, reset all states
            if self.request is not None:
                self.request.reset_state()
                # Set initial value for the next beam_idx input that will be used at the current iteration
                # and will be optionally updated by _reorder_cache at the next iterations if beam_search is used
                self.next_beam_idx = np.arange(batch_size, dtype=int)
                self._past_length = 0
        past_len = self._get_past_length(past_key_values)

        if inputs_embeds is None:
            inputs_embeds = self.embed_tokens(input_ids if past_key_values is None else input_ids[:, -1:])

            if hasattr(self.config, "scale_emb"):
                inputs_embeds = inputs_embeds * self.config.scale_emb
        inputs["inputs_embeds"] = inputs_embeds

        # Add the attention_mask inputs when needed
        if "attention_mask" in self.input_names or "position_ids" in self.input_names:
            if attention_mask is not None:
                attention_mask = attention_mask.cpu().numpy()
            else:
                attention_mask = np.ones((inputs_embeds.shape[0], inputs_embeds.shape[1] + past_len), dtype=int)

        if "attention_mask" in self.input_names:
            inputs["attention_mask"] = attention_mask

        if "position_ids" in self.input_names:
            if position_ids is not None:
                position_ids = position_ids.cpu().numpy()
            else:
                position_ids = np.cumsum(attention_mask, axis=1) - 1
                position_ids[attention_mask == 0] = 1
            if past_len:
                position_ids = position_ids[:, -inputs_embeds.shape[1] :]

            if (self.config.model_type in ["qwen2_vl", "qwen2_5_vl", "qwen3_vl"]) and position_ids.ndim != 3:
                position_ids = np.repeat(np.expand_dims(position_ids, 0), 3, axis=0)

            inputs["position_ids"] = position_ids

        if "visual_pos_masks" in self.input_names:
            if visual_pos_masks is not None:
                inputs["visual_pos_masks"] = visual_pos_masks
            else:
                inputs["visual_pos_masks"] = torch.zeros(1, 1, dtype=torch.bool)

        if "deepstack_visual_embeds" in self.input_names:
            if deepstack_visual_embeds is not None:
                inputs["deepstack_visual_embeds"] = torch.Tensor(deepstack_visual_embeds)
            else:
                num_layers = len(self.config.vision_config.deepstack_visual_indexes)
                emd_dim = self.config.text_config.hidden_size
                inputs["deepstack_visual_embeds"] = torch.zeros((num_layers, 1, emd_dim), dtype=torch.float32)

        if "token_type_ids" in self.input_names:
            if token_type_ids is None:
                token_type_ids = np.zeros(inputs_embeds.shape[:2], dtype=int)
            inputs["token_type_ids"] = token_type_ids

        if "beam_idx" in self.input_names:
            inputs["beam_idx"] = (
                self.next_beam_idx if self.next_beam_idx is not None else np.arange(batch_size, dtype=int)
            )
        return inputs

    def forward(
        self,
        input_ids: torch.LongTensor,
        attention_mask: Optional[torch.LongTensor] = None,
        past_key_values: Optional[Tuple[Tuple[torch.FloatTensor]]] = None,
        position_ids: Optional[torch.LongTensor] = None,
        inputs_embeds: Optional[torch.LongTensor] = None,
        visual_pos_masks: Optional[torch.FloatTensor] = None,
        deepstack_visual_embeds: Optional[torch.FloatTensor] = None,
        **kwargs,
    ):
        self.compile()
        inputs = self.prepare_inputs(
            input_ids=input_ids,
            attention_mask=attention_mask,
            past_key_values=past_key_values,
            position_ids=position_ids,
            inputs_embeds=inputs_embeds,
            visual_pos_masks=visual_pos_masks,
            deepstack_visual_embeds=deepstack_visual_embeds,
            **kwargs,
        )
        # Run inference
        self.request.start_async(inputs, share_inputs=True)
        self.request.wait()
        logits = self.request.get_tensor("logits").data
        logits = torch.from_numpy(logits).clone().to(self.device)
        past_key_values = ((),)
        self._past_length += inputs["inputs_embeds"].shape[1]

        return CausalLMOutputWithPast(logits=logits, past_key_values=past_key_values)


class OVVisionEmbedding(OVModelPart):
    _model_name = "vision_embeddings"

    def __init__(self, model: ov.Model, parent_model: OVBaseModel) -> None:
        super().__init__(model, parent_model, model_name=self._model_name)
        self.output_dtypes = {key.get_any_name(): key.get_element_type().get_type_name() for key in self.model.outputs}
        self.output_names = {key.get_any_name(): idx for idx, key in enumerate(self.model.outputs)}
        self.input_names = {key.get_any_name(): idx for idx, key in enumerate(self.model.inputs)}
        self.hidden_states_output_names = []
        if len(self.model.outputs) > 2:
            self.hidden_states_output_names = [
                key.get_any_name() for key in self.model.outputs[2:] if "hidden_states" in key.get_any_name()
            ]
        self.input_names = {key.get_any_name(): idx for idx, key in enumerate(self.model.inputs)}
        if model_has_input_output_name(self.model, "images"):
            self._main_input = "images"
        elif model_has_input_output_name(self.model, "hidden_states"):
            self._main_input = "hidden_states"
        else:
            self._main_input = "pixel_values"

    def forward(self, pixel_values, **kwargs):
        self.compile()
        inputs = {self._main_input: pixel_values}
        if len(self.input_names) > 1:
            for name in self.input_names:
                if name in kwargs:
                    inputs[name] = kwargs[name]
        result = self.request(inputs)
        last_hidden_state = result[0]
        hidden_states = None
        pooler_out = None
        if len(result) > 1:
            pooler_out = result[1]
            if self.hidden_states_output_names:
                hidden_states = []
                for out in self.hidden_states_output_names:
                    hidden_states.append(result[out])
        return BaseModelOutputWithPooling(
            pooler_output=pooler_out, last_hidden_state=last_hidden_state, hidden_states=hidden_states
        )


class OVResampler(OVModelPart):
    _model_name = "resampler"

    def __init__(self, model: ov.Model, parent_model: OVBaseModel) -> None:
        super().__init__(model, parent_model, model_name=self._model_name)
        self.output_dtypes = {key.get_any_name(): key.get_element_type().get_type_name() for key in self.model.outputs}
        self.output_names = {key.get_any_name(): idx for idx, key in enumerate(self.model.outputs)}

    def forward(self, image_feature, pos_embed, key_padding_mask):
        self.compile()
        result = self.request(
            {"image_feature": image_feature, "pos_embed": pos_embed, "key_padding_mask": key_padding_mask}
        )[0]
        return result


class OVVisionProjection(OVModelPart):
    _model_name = "vision_projection"

    def forward(self, img_features):
        self.compile()
        return self.request(img_features)[0]


class OVVisionResampler(OVVisionProjection):
    _model_name = "vision_resampler"


class OVMultiModalProjector(OVVisionProjection):
    _model_name = "multi_modal_projector"


class OVAudioEmbeddings(OVModelPart):
    _model_name = "audio_embeddings"

    def forward(self, audio_signal):
        self.compile()
        return self.request(audio_signal)[0]


class OVAudioEncoder(OVModelPart):
    _model_name = "audio_encoder"

    def forward(self, audio_feature, audio_mask):
        self.compile()
        return self.request({"audio_feature": audio_feature, "audio_mask": audio_mask})[0]


MODEL_PARTS_CLS_MAPPING = {
    "resampler": OVResampler,
    "language_model": OVModelWithEmbedForCausalLM,
    "vision_embeddings": OVVisionEmbedding,
    "vision_projection": OVVisionProjection,
    "vision_resampler": OVVisionResampler,
    "multi_modal_projector": OVMultiModalProjector,
    "vision_embeddings_merger": OVVisionEmbedding,
    "vision_embeddings_pos": OVVisionProjection,
    "audio_embeddings": OVAudioEmbeddings,
    "audio_forward_embeddings": OVAudioEmbeddings,
    "audio_encoder": OVAudioEncoder,
    "audio_vision_projection": OVAudioEmbeddings,
    "audio_speech_projection": OVAudioEmbeddings,
}


class OVModelForVisualCausalLM(OVBaseModel, GenerationMixin):
    export_feature = "image-text-to-text"
    additional_parts = []
    auto_model_class = transformers_auto_class

    @classproperty
    def _all_ov_model_paths(cls) -> Dict[str, str]:
        model_paths = {
            "lm_model": OV_LANGUAGE_MODEL_NAME,
            "text_embeddings_model": OV_TEXT_EMBEDDINGS_MODEL_NAME,
            "vision_embeddings_model": OV_VISION_EMBEDDINGS_MODEL_NAME,
        }
        for part in cls.additional_parts:
            model_paths[f"{part}_model"] = f"openvino_{part}_model.xml"
        return model_paths

    def __init__(
        self,
        language_model: ov.Model,
        text_embeddings: ov.Model,
        vision_embeddings: ov.Model,
        config: PretrainedConfig = None,
        device: str = "CPU",
        dynamic_shapes: bool = None,
        ov_config: Optional[Dict[str, str]] = None,
        model_save_dir: Optional[Union[str, Path, TemporaryDirectory]] = None,
        quantization_config: Union[OVWeightQuantizationConfig, Dict] = None,
        **kwargs,
    ):
        if dynamic_shapes is not None:
            logger.warning(
                f"`dynamic_shapes` was set to {dynamic_shapes}, but this value will be ignored as only dynamic shapes are supported."
            )

        self.is_dynamic = True
        self.config = config
        self.use_cache = kwargs.get("use_cache", True)
        self.model_save_dir = model_save_dir
        self._device = device.upper()
        self.ov_config = {} if ov_config is None else {**ov_config}
        self.preprocessors = kwargs.get("preprocessors", [])
        self._supports_cache_class = False
        self.main_input_name = "input_ids"
        self._compile_only = kwargs.get("compile_only", False)

        for part in self.additional_parts:
            setattr(self, f"{part}_model", kwargs.get(part))

        enable_compilation = kwargs.get("compile", True)
        self.generation_config = kwargs.get("generation_config", GenerationConfig.from_model_config(config))
        self._openvino_config = None
        if quantization_config:
            self._openvino_config = OVConfig(quantization_config=quantization_config)
        self._set_ov_config_parameters()
        self.language_model = OVModelWithEmbedForCausalLM(
            language_model,
            text_embeddings,
            config=config,
            device=device,
            ov_config=ov_config,
            model_save_dir=model_save_dir,
            quantization_config=quantization_config,
            compile=self._compile_only or enable_compilation,
            compile_only=self._compile_only,
        )
        self.vision_embeddings = OVVisionEmbedding(vision_embeddings, self)
        for part in self.additional_parts:
            model_part = getattr(self, f"{part}_model", None)
            if model_part is not None:
                model_part = MODEL_PARTS_CLS_MAPPING[part](model_part, self)
            setattr(self, part, model_part)

        if enable_compilation and not self._compile_only:
            self.compile()

        # Avoid warnings when creating a transformers pipeline
        AutoConfig.register(self.base_model_prefix, AutoConfig)
        self.auto_model_class.register(AutoConfig, self.__class__)

    def clear_requests(self):
        if self._compile_only:
            raise ValueError(
                "`clear_requests()` is not supported with `compile_only` mode, please initialize model without this option"
            )

        for component in self.components.values():
            component.clear_requests()

    def compile(self):
        for component in self.components.values():
            component.compile()

    def _save_config(self, save_directory):
        """
        Saves a model configuration into a directory, so that it can be re-loaded using the
        [`from_pretrained`] class method.
        """
        save_config(self.config, save_directory)

    @classmethod
    def _from_pretrained(
        cls,
        model_id: Union[str, Path],
        config: PretrainedConfig,
        token: Optional[Union[bool, str]] = None,
        revision: Optional[str] = None,
        force_download: bool = False,
        cache_dir: str = HUGGINGFACE_HUB_CACHE,
        local_files_only: bool = False,
        load_in_8bit: bool = False,
        quantization_config: Union[OVWeightQuantizationConfig, Dict] = None,
        trust_remote_code: bool = False,
        **kwargs,
    ):
        """
        Loads a model and its configuration file from a directory or the HF Hub.

        Arguments:
            model_id (`str` or `Path`):
                The directory from which to load the model.
                Can be either:
                    - The model id of a pretrained model hosted inside a model repo on huggingface.co.
                    - The path to a directory containing the model weights.
            token (Optional[Union[bool, str]], defaults to `None`):
                The token to use as HTTP bearer authorization for remote files. If `True`, will use the token generated
                when running `huggingface-cli login` (stored in `~/.huggingface`).
            revision (`str`):
                The specific model version to use. It can be a branch name, a tag name, or a commit id.
            force_download (`bool`, *optional*, defaults to `False`):
                Whether or not to force the (re-)download of the model weights and configuration files, overriding the
                cached versions if they exist.
            cache_dir (`Union[str, Path]`, *optional*):
                The path to a directory in which a downloaded pretrained model configuration should be cached if the
                standard cache should not be used.
            encoder_file_name(`str`, *optional*):
                The encoder model file name. Overwrites the default file name openvino_encoder_model.xml and allows one to
                load the encoder model with a different name.
            decoder_file_name(`str`, *optional*):
                The decoder model file name. Overwrites the default file name openvino_decoder_model.xml and allows one to
                load the decoder model with a different name.
            decoder_with_past_file_name(`str`, *optional*):
                The decoder with past key values model file name overwriting the default file name
                openvino_decoder_with_past_model.xml, allowing to load the decoder model with a different name.
            local_files_only(`bool`, *optional*, defaults to `False`):
                Whether or not to only look at local files (i.e., do not try to download the model).
            trust_remote_code (`bool`, *optional*, defaults to `False`):
                Whether to trust remote code when loading model tokenizer/processor during quantization.
        """
        model_cls = MODEL_TYPE_TO_CLS_MAPPING[config.model_type]
        model_file_names = model_cls._all_ov_model_paths.copy()
        for k in tuple(model_file_names):
            model_file_names[f"{k}_bin"] = model_file_names[k].replace(".xml", ".bin")
        compile_only = kwargs.get("compile_only", False)
        if os.path.isdir(model_id):
            # Load model from a local directory
            model_save_dir = Path(model_id)
            file_names = {k: os.path.join(model_id, model_file_names[k]) for k in model_file_names}
        else:
            file_names = {}
            for name, file_name in model_file_names.items():
                model_cache_path = hf_hub_download(
                    repo_id=model_id,
                    filename=file_name,
                    token=token,
                    revision=revision,
                    cache_dir=cache_dir,
                    force_download=force_download,
                    local_files_only=local_files_only,
                )
                file_names[name] = model_cache_path
            model_save_dir = Path(model_cache_path).parent
        if not compile_only:
            language_model = model_cls.load_model(file_names["lm_model"])
            text_embeddings = model_cls.load_model(file_names["text_embeddings_model"])
            vision_embeddings = model_cls.load_model(file_names["vision_embeddings_model"])
            for part in model_cls.additional_parts:
                kwargs[part] = model_cls.load_model(file_names[f"{part}_model"])
        else:
            language_model = model_cls._compile_model(
                file_names["lm_model"],
                kwargs.get("device", "CPU"),
                kwargs.get("ov_config"),
                model_save_dir,
            )
            text_embeddings = model_cls._compile_model(
                file_names["text_embeddings_model"],
                kwargs.get("device", "CPU"),
                kwargs.get("ov_config"),
                model_save_dir,
            )
            vision_embeddings = model_cls._compile_model(
                file_names["vision_embeddings_model"],
                kwargs.get("device", "CPU"),
                kwargs.get("ov_config"),
                model_save_dir,
            )
            for part in model_cls.additional_parts:
                kwargs[part] = model_cls._compile_model(
                    file_names[f"{part}_model"],
                    kwargs.get("device", "CPU"),
                    kwargs.get("ov_config"),
                    model_save_dir,
                )
        try:
            generation_config = GenerationConfig.from_pretrained(
                model_id,
                token=token,
                revision=revision,
                cache_dir=cache_dir,
                force_download=force_download,
                local_files_only=local_files_only,
            )
            kwargs["generation_config"] = generation_config
        except Exception:
            pass

        quantization_config = quantization_config or (OVWeightQuantizationConfig(bits=8) if load_in_8bit else None)
        compile_model = kwargs.pop("compile", True)
        model = model_cls(
            language_model=language_model,
            text_embeddings=text_embeddings,
            vision_embeddings=vision_embeddings,
            config=config,
            model_save_dir=model_save_dir,
            quantization_config=quantization_config,
            compile=compile_model and not quantization_config,
            **kwargs,
        )

        if quantization_config:
            if hasattr(config, "name_or_path"):
                model_id = config.name_or_path
            else:
                logger.warning(
                    "`model_id` could not be determined from the config. In the case there are default quantization "
                    "configurations for this model, they will not be applied."
                )
            quantization_config = cls._resolve_default_quantization_config(model_id, quantization_config)
            model._apply_quantization(quantization_config, compile_only, compile_model, model_id, trust_remote_code)

        return model

    @classmethod
    def _export(
        cls,
        model_id: str,
        config: PretrainedConfig,
        token: Optional[Union[bool, str]] = None,
        revision: Optional[str] = None,
        force_download: bool = False,
        cache_dir: str = HUGGINGFACE_HUB_CACHE,
        subfolder: str = "",
        local_files_only: bool = False,
        task: Optional[str] = None,
        use_cache: bool = True,
        trust_remote_code: bool = False,
        load_in_8bit: Optional[bool] = None,
        quantization_config: Union[OVWeightQuantizationConfig, Dict] = None,
        **kwargs,
    ):
        compile_only = kwargs.pop("compile_only", False)
        if compile_only:
            logger.warning(
                "`compile_only` mode will be disabled because it does not support model export."
                "Please provide openvino model obtained using optimum-cli or saved on disk using `save_pretrained`"
            )
            compile_only = False
        save_dir = TemporaryDirectory()
        save_dir_path = Path(save_dir.name)

        # This attribute is needed to keep one reference on the temporary directory, since garbage collecting
        # would end-up removing the directory containing the underlying OpenVINO model
        cls._model_save_dir_tempdirectory_instance = save_dir

        if task is None:
            task = cls.export_feature

        # If load_in_8bit and quantization_config not specified then ov_config is set to None and will be set by default in convert depending on the model size
        if load_in_8bit is None and not quantization_config:
            ov_config = None
        else:
            # Export in fp32 if compression won't be applied later
            ov_config = OVConfig(dtype="fp32" if load_in_8bit is False else "auto")

        stateful = kwargs.pop("stateful", ensure_stateful_is_available(warn=False) and use_cache)
        variant = kwargs.pop("variant", None)

        main_export(
            model_name_or_path=model_id,
            output=save_dir_path,
            task=task,
            subfolder=subfolder,
            revision=revision,
            cache_dir=cache_dir,
            token=token,
            local_files_only=local_files_only,
            force_download=force_download,
            trust_remote_code=trust_remote_code,
            ov_config=ov_config,
            stateful=stateful,
            variant=variant,
        )
        name_or_path = config.name_or_path
        config = AutoConfig.from_pretrained(save_dir_path, trust_remote_code=trust_remote_code)
        # Keep the original name_or_path to be able to resolve default quantization config later
        config.name_or_path = name_or_path
        return cls._from_pretrained(
            model_id=save_dir_path,
            config=config,
            use_cache=use_cache,
            load_in_8bit=load_in_8bit,
            quantization_config=quantization_config,
            trust_remote_code=trust_remote_code,
            **kwargs,
        )

    @property
    def _component_names(self) -> List[str]:
        base_components = ["language_model", "vision_embeddings"]
        additional_components = [part for part in self.additional_parts if hasattr(self, part)]
        return base_components + additional_components

    @property
    def _ov_model_names(self):
        # TODO (nikita.savelyevv): Consider deprecating `lm_model` in favor of `language_model`
        model_names = ["lm_model", "text_embeddings_model", "vision_embeddings_model"]
        for part in self.additional_parts:
            if hasattr(self, part):
                model_names.append(part + "_model")
        return model_names

    @property
    def ov_models(self) -> Dict[str, Union[openvino.Model, openvino.CompiledModel]]:
        ov_models = {}
        for ov_model_name in self._ov_model_names:
            if ov_model_name == "lm_model":
                ov_model = self.language_model.model
            elif ov_model_name == "text_embeddings_model":
                ov_model = self.language_model.text_emb_model
            else:
                ov_model = getattr(self, ov_model_name.replace("_model", "")).model
            ov_models[ov_model_name] = ov_model
        return ov_models

    def reshape(self, batch_size: int, sequence_length: int):
        logger.warning("Static shapes are not supported for causal language model.")
        return self

    def half(self):
        """
        Converts all the model weights to FP16 for more efficient inference on GPU.
        """
        for ov_model in self.ov_models.values():
            apply_moc_transformations(ov_model, cf=False)
            compress_model_transformation(ov_model)
        return self

    def to(self, device):
        self.language_model.to(device)
        super().to(device)
        return self

    def forward(
        self,
        input_ids,
        pixel_values=None,
        past_key_values=None,
        inputs_embeds=None,
        image_sizes=None,
        attention_mask=None,
        position_ids=None,
        image_bound=None,
        tgt_sizes=None,
        pixel_values_videos=None,
        image_grid_thw=None,
        video_grid_thw=None,
        rope_deltas=None,
        images=None,
        second_per_grid_ts=None,
        token_type_ids=None,
        pixel_attention_mask=None,
        input_image_embeds: Optional[torch.FloatTensor] = None,
        image_pixel_values: Optional[torch.FloatTensor] = None,
        image_attention_mask=None,
        audio_input_features: Optional[torch.FloatTensor] = None,
        input_audio_embeds: Optional[torch.FloatTensor] = None,
        audio_embed_sizes=None,
        audio_attention_mask=None,
        input_mode=None,
        **kwargs,
    ):
        if pixel_values is None:
            pixel_values = images if images is not None else image_pixel_values
        inputs_embeds, attention_mask, position_ids, *extra_outputs = self.get_multimodal_embeddings(
            input_ids,
            pixel_values,
            inputs_embeds=inputs_embeds,
            image_sizes=image_sizes,
            attention_mask=attention_mask,
            position_ids=position_ids,
            past_key_values=past_key_values,
            image_bound=image_bound,
            tgt_sizes=tgt_sizes,
            pixel_values_videos=pixel_values_videos,
            image_grid_thw=image_grid_thw,
            video_grid_thw=video_grid_thw,
            rope_deltas=rope_deltas,
            second_per_grid_ts=second_per_grid_ts,
            pixel_attention_mask=pixel_attention_mask,
            input_image_embeds=input_image_embeds,
            image_attention_mask=image_attention_mask,
            input_audio_embeds=input_audio_embeds if input_audio_embeds is not None else audio_input_features,
            audio_embed_sizes=audio_embed_sizes,
            audio_attention_mask=audio_attention_mask,
            input_mode=input_mode,
            **kwargs,
        )

        # Prepare additional kwargs for qwen3_vl models
        additional_kwargs = {}
        if self.config.model_type in ("qwen3_vl",) and extra_outputs:
            additional_kwargs["visual_pos_masks"] = extra_outputs[0]
            additional_kwargs["deepstack_visual_embeds"] = extra_outputs[1]

        return self.language_model.forward(
            input_ids=None,
            inputs_embeds=inputs_embeds,
            attention_mask=attention_mask,
            position_ids=position_ids,
            token_type_ids=token_type_ids,
            past_key_values=past_key_values,
            **additional_kwargs,
            **kwargs,
        )

    def _reorder_cache(self, past_key_values, beam_idx):
        return self.language_model._reorder_cache(past_key_values, beam_idx)

    def get_vision_embeddings(self, pixel_values, **kwargs):
        raise NotImplementedError

    def get_text_embeddings(self, input_ids, **kwargs):
        return self.language_model.embed_tokens(input_ids)

    def merge_vision_text_embeddings(
        self, vision_embeds, inputs_embeds, input_ids=None, attention_mask=None, position_ids=None, **kwargs
    ):
        raise NotImplementedError

    def get_multimodal_embeddings(
        self, input_ids, pixel_values=None, attention_mask=None, position_ids=None, **kwargs
    ):
        embeds_from_args = kwargs.pop("inputs_embeds", None)
        inputs_embeds = (
            embeds_from_args if embeds_from_args is not None else self.get_text_embeddings(input_ids, **kwargs)
        )
        if pixel_values is not None:
            vision_embeds = self.get_vision_embeddings(pixel_values, input_ids=input_ids, **kwargs)
            if vision_embeds is not None:
                inputs_embeds, attention_mask, position_ids = self.merge_vision_text_embeddings(
                    vision_embeds,
                    inputs_embeds,
                    input_ids=input_ids,
                    attention_mask=attention_mask,
                    position_ids=position_ids,
                    **kwargs,
                )
        return inputs_embeds, attention_mask, position_ids

    # Adopted from https://github.com/huggingface/transformers/blob/v4.44.2/src/transformers/models/llava/modeling_llava.py#L521
    def prepare_inputs_for_generation(
        self,
        input_ids,
        past_key_values=None,
        inputs_embeds=None,
        pixel_values=None,
        image_sizes=None,
        attention_mask=None,
        **kwargs,
    ):
        if past_key_values is not None:
            past_length = self.language_model._get_past_length(past_key_values)

            # 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 past_length + 1 > input_ids.shape[1]:
                input_discount = max(attention_mask.shape[1] - past_length, 1)
                input_ids = input_ids[:, -input_discount:]
            # 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.llava
            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 getattr(self.config, "image_token_index", -1) in input_ids:
                input_ids = input_ids[:, input_ids.shape[1] - 1 :]

        position_ids = kwargs.get("position_ids", None)
        if attention_mask is not None and position_ids is None:
            position_ids = attention_mask.long().cumsum(-1) - 1
            position_ids.masked_fill_(attention_mask == 0, 1)

            # position_ids in Gemma3 are 1-indexed
            if self.config.model_type == "gemma3":
                position_ids += 1

            if past_key_values is not None:
                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 pixel_values is None:
            pixel_values = kwargs.get("input_image_embeds", kwargs.get("images", kwargs.get("image_pixel_values")))

        model_inputs.update(
            {
                "position_ids": position_ids,
                "past_key_values": past_key_values,
                "use_cache": kwargs.get("use_cache"),
                "attention_mask": attention_mask,
                "pixel_values": pixel_values,
                "image_sizes": image_sizes,
                "image_bound": kwargs.get("image_bound"),
                "tgt_sizes": kwargs.get("tgt_sizes"),
                "pixel_values_videos": kwargs.get("pixel_values_videos"),
                "image_grid_thw": kwargs.get("image_grid_thw"),
                "video_grid_thw": kwargs.get("video_grid_thw"),
                "token_type_ids": kwargs.get("token_type_ids"),
                "pixel_attention_mask": kwargs.get("pixel_attention_mask"),
                "image_attention_mask": kwargs.get("image_attention_mask"),
                "input_audio_embeds": kwargs.get("input_audio_embeds", kwargs.get("audio_input_features")),
                "audio_embed_sizes": kwargs.get("audio_embed_sizes"),
                "input_mode": kwargs.get("input_mode"),
            }
        )
        return model_inputs

    def can_generate(self):
        """Returns True to validate the check that the model using `GenerationMixin.generate()` can indeed generate."""
        return True

    @staticmethod
    @abstractmethod
    def preprocess_inputs(
        text: str,
        image: Optional["Image"] = None,
        processor: Optional[AutoImageProcessor] = None,
        tokenizer: Optional[PreTrainedTokenizer] = None,
        config: Optional[PretrainedConfig] = None,
        video: Optional["VideoInput"] = None,
        audio: Optional[np.ndarray] = None,
    ):
        """
        Preprocess input instruction and an image.
        """

    # modified from https://github.com/huggingface/transformers/blob/v4.55.0/src/transformers/generation/utils.py#L1992
    def _prepare_cache_for_generation(self, *args, **kwargs):
        """
        This function is used to prepare the cache : when calling `generate` before the first inference, an instance of `DynamicCache` will be created.
        For OVModel, we don't want model_kwargs to be updated before generation.
        """
        return

    def _preprocess_quantization_config(
        self,
        quantization_config: OVQuantizationConfigBase,
        model_name_or_path: str,
    ) -> OVQuantizationConfigBase:
        if quantization_config.processor is None or quantization_config.tokenizer is None:
            quantization_config = quantization_config.clone()
            if quantization_config.processor is None:
                potential_processor_id = (
                    self.config.mm_vision_tower if isinstance(self, _OVNanoLlavaForCausalLM) else model_name_or_path
                )
                quantization_config.processor = potential_processor_id
            if quantization_config.tokenizer is None:
                quantization_config.tokenizer = model_name_or_path
        return quantization_config


class _OVLlavaForCausalLM(OVModelForVisualCausalLM):
    def __init__(
        self,
        language_model: ov.Model,
        text_embeddings: ov.Model,
        vision_embeddings: ov.Model,
        config: PretrainedConfig = None,
        device: str = "CPU",
        dynamic_shapes: bool = None,
        ov_config: Optional[Dict[str, str]] = None,
        model_save_dir: Optional[Union[str, Path, TemporaryDirectory]] = None,
        quantization_config: Union[OVWeightQuantizationConfig, Dict] = None,
        **kwargs,
    ):
        super().__init__(
            language_model=language_model,
            text_embeddings=text_embeddings,
            vision_embeddings=vision_embeddings,
            config=config,
            device=device,
            dynamic_shapes=dynamic_shapes,
            ov_config=ov_config,
            model_save_dir=model_save_dir,
            quantization_config=quantization_config,
            **kwargs,
        )
        self._support_new_processing = hasattr(self.config, "image_seq_length")

    def get_vision_embeddings(self, pixel_values, input_ids=None, **kwargs):
        if input_ids is not None and input_ids.shape[1] == 1:
            return None
        if not isinstance(pixel_values, list):
            image_features = self.vision_embeddings(pixel_values).last_hidden_state
        else:
            image_features = []
            for patch in pixel_values:
                if isinstance(patch, list):
                    patch_feats = []
                    for patch_value in patch:
                        patch_feats.append(self.vision_embeddings(np.expand_dims(patch_value, 0)).last_hidden_state)
                    patch_feats = np.concatenate(patch_feats, axis=1)
                else:
                    patch_feats = self.vision_embeddings(patch).last_hidden_state
                image_features.append(patch_feats)
            image_features = np.concatenate(image_features, 0)

        return image_features

    # Adopted from https://github.com/huggingface/transformers/blob/d7950bff82b18c823193d17d72188c5e46d06c83/src/transformers/models/llava/modeling_llava.py#L297C9-L297C45
    def merge_vision_text_embeddings(
        self,
        vision_embeds,
        inputs_embeds,
        input_ids,
        attention_mask,
        position_ids=None,
        legacy_processing=False,
        **kwargs,
    ):
        image_features = torch.from_numpy(vision_embeds) if isinstance(vision_embeds, np.ndarray) else vision_embeds
        inputs_embeds = torch.from_numpy(inputs_embeds) if isinstance(inputs_embeds, np.ndarray) else inputs_embeds

        if legacy_processing:
            pad_token_id = self.config.pad_token_id if self.config.pad_token_id is not None else -1

            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(pad_token_id))
            # 1. Create a mask to know where special image tokens are
            special_image_token_mask = input_ids == self.config.image_token_index
            num_special_image_tokens = torch.sum(special_image_token_mask, dim=-1)
            # Compute the maximum embed dimension
            max_embed_dim = (num_special_image_tokens.max() * (num_image_patches - 1)) + sequence_length
            batch_indices, non_image_indices = torch.where(input_ids != self.config.image_token_index)

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

            # 3. Create the full embedding, already padded to the maximum position
            final_embedding = torch.zeros(
                batch_size, max_embed_dim, embed_dim, dtype=inputs_embeds.dtype, device=inputs_embeds.device
            )
            final_attention_mask = torch.zeros(
                batch_size, max_embed_dim, dtype=attention_mask.dtype, device=inputs_embeds.device
            )

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

            if image_to_overwrite.sum() != image_features.shape[:-1].numel():
                raise ValueError(
                    f"The input provided to the model a/pre-releasesre 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)
            final_attention_mask |= image_to_overwrite
            position_ids = (final_attention_mask.cumsum(-1) - 1).masked_fill_((final_attention_mask == 0), 1)

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

            final_embedding[batch_indices, indices_to_mask] = 0
        else:
            special_image_mask = (input_ids == self.config.image_token_index).unsqueeze(-1).expand_as(inputs_embeds)
            image_features = image_features.to(inputs_embeds.dtype)
            final_embedding = inputs_embeds.masked_scatter(special_image_mask, image_features)
            final_attention_mask = attention_mask

        return final_embedding, final_attention_mask, position_ids

    def get_multimodal_embeddings(
        self, input_ids, pixel_values=None, attention_mask=None, position_ids=None, past_key_values=None, **kwargs
    ):
        if pixel_values is not None and self._support_new_processing and past_key_values is None:
            legacy_processing = (input_ids == self.config.image_token_index).sum(
                1
            ).max() < self.config.image_seq_length
        else:
            legacy_processing = True
        inputs_embeds, attention_mask, position_ids = super().get_multimodal_embeddings(
            input_ids, pixel_values, attention_mask, position_ids, legacy_processing=legacy_processing, **kwargs
        )

        if legacy_processing and pixel_values is not None and past_key_values is not None:
            attention_mask, position_ids = self._filter_unattended_tokens(input_ids, attention_mask, past_key_values)

        return inputs_embeds, attention_mask, position_ids

    def _filter_unattended_tokens(self, input_ids, attention_mask, past_key_values):
        # Get the target length
        target_length = input_ids.shape[1]
        past_length = self.language_model._get_past_length(past_key_values)

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

        attention_mask = torch.cat((extended_attention_mask, attention_mask[:, -target_length:]), dim=1)
        position_ids = torch.cumsum(attention_mask, axis=1) - 1
        position_ids[attention_mask == 0] = 1
        return attention_mask, position_ids

    @staticmethod
    def preprocess_inputs(
        text: str,
        image: Optional["Image"] = None,
        processor: Optional[AutoImageProcessor] = None,
        tokenizer: Optional[PreTrainedTokenizer] = None,
        config: Optional[PretrainedConfig] = None,
        video: Optional["VideoInput"] = None,
        audio: Optional[np.ndarray] = None,
    ):
        if processor is None:
            raise ValueError("Processor is required.")
        if video is not None:
            raise ValueError("Video input is not supported")
        if audio is not None:
            raise ValueError("Audio input is not supported")
        if getattr(processor, "chat_template", None) is not None:
            chat_prompt = [{"role": "user", "content": [{"type": "text", "text": text}]}]
            if image is not None:
                chat_prompt[0]["content"].append({"type": "image"})
            prompt = processor.apply_chat_template(chat_prompt, add_generation_prompt=True, tokenize=False)
        else:
            if image is not None and "<image>" not in text:
                prompt = "<image>\n" + text
            else:
                prompt = text

        if is_transformers_version(">", "4.47.99") and getattr(processor, "patch_size", None) is None:
            if (
                getattr(config, "vision_config", None) is not None
                and getattr(config.vision_config, "patch_size", None) is not None
            ):
                processor.patch_size = config.vision_config.patch_size
            else:
                raise ValueError(
                    "Processor does not have `patch_size` attribute. Please fix the processor or provide `patch_size` in the config."
                )

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


class _OVLlavaNextForCausalLM(_OVLlavaForCausalLM):
    # Adopted from https://github.com/huggingface/transformers/blob/main/src/transformers/models/llava_next/modeling_llava_next.py#L655
    def pack_image_features(self, image_features, image_sizes, image_newline=None):
        from transformers.models.llava_next.modeling_llava_next import get_anyres_image_grid_shape, unpad_image

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

    def add_image_features(
        self,
        input_ids,
        inputs_embeds,
        pixel_values,
        attention_mask,
        position_ids,
        image_sizes,
        legacy_processing,
        **kwargs,
    ):
        from transformers.models.llava_next.modeling_llava_next import image_size_to_num_patches

        # ! infer image_num_patches from image_sizes
        image_num_patches = [
            image_size_to_num_patches(
                image_size=imsize,
                grid_pinpoints=self.config.image_grid_pinpoints,
                patch_size=self.config.vision_config.image_size,
            )
            for imsize in image_sizes
        ]
        # figure out if pixel_values is concatenated or stacked
        if pixel_values.dim() == 5:
            # stacking when input is (batch_size, num_patches, num_channels, height, width)
            _pixel_values_list = [pix_val[:num_patch] for pix_val, num_patch in zip(pixel_values, image_num_patches)]
            pixel_values = torch.cat(_pixel_values_list, dim=0)
        elif pixel_values.dim() != 4:
            # otherwise has to be stacked from list of (num_patches, num_channels, height, width)
            raise ValueError(f"pixel_values of shape {pixel_values.shape}, expect to be of 4 or 5 dimensions")
        vision_embeds = self.get_vision_embeddings(pixel_values, input_ids=input_ids, **kwargs)
        if vision_embeds is not None:
            image_newline = torch.tensor(self.config.image_newline)
            image_features = torch.split(torch.from_numpy(vision_embeds), image_num_patches, dim=0)
            image_features, feature_lens = self.pack_image_features(
                image_features,
                image_sizes,
                image_newline=image_newline,
            )
            inputs_embeds, attention_mask, position_ids = self.merge_vision_text_embeddings(
                image_features,
                inputs_embeds,
                feature_lens=feature_lens,
                input_ids=input_ids,
                attention_mask=attention_mask,
                position_ids=position_ids,
                legacy_processing=legacy_processing,
                **kwargs,
            )
        return inputs_embeds, attention_mask, position_ids

    # Adopted from https://github.com/huggingface/transformers/blob/main/src/transformers/models/llava_next/modeling_llava_next.py#L416
    def get_multimodal_embeddings(
        self,
        input_ids,
        pixel_values=None,
        attention_mask=None,
        position_ids=None,
        past_key_values=None,
        image_sizes=None,
        **kwargs,
    ):
        embeds_from_args = kwargs.pop("inputs_embeds", None)
        inputs_embeds = (
            embeds_from_args if embeds_from_args is not None else self.get_text_embeddings(input_ids, **kwargs)
        )

        if pixel_values is not None and self._support_new_processing and past_key_values is None:
            legacy_processing = (input_ids == self.config.image_token_index).sum(
                1
            ).max() < self.config.image_seq_length
        else:
            legacy_processing = True

        if pixel_values is not None and pixel_values.size(0) > 0:
            inputs_embeds, attention_mask, position_ids = self.add_image_features(
                input_ids,
                inputs_embeds,
                pixel_values,
                attention_mask,
                position_ids,
                image_sizes,
                legacy_processing,
                **kwargs,
            )

        if legacy_processing and pixel_values is not None and past_key_values is not None and input_ids.shape[1] == 1:
            attention_mask, position_ids = self._filter_unattended_tokens(input_ids, attention_mask, past_key_values)

        return inputs_embeds, attention_mask, position_ids

    def merge_vision_text_embeddings(
        self,
        vision_embeds,
        inputs_embeds,
        feature_lens,
        input_ids,
        attention_mask,
        position_ids=None,
        legacy_processing=False,
        image_token_index=None,
        **kwargs,
    ):
        image_token_index = self.config.image_token_index if image_token_index is None else image_token_index
        image_features = torch.from_numpy(vision_embeds) if isinstance(vision_embeds, np.ndarray) else vision_embeds
        inputs_embeds = torch.from_numpy(inputs_embeds) if isinstance(inputs_embeds, np.ndarray) else inputs_embeds

        if legacy_processing:
            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 batch_size > 1:
                    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:
                        left_padding = True
                    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
            )
            # 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]

            # 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)
        else:
            special_image_mask = (input_ids == image_token_index).unsqueeze(-1).expand_as(inputs_embeds)
            image_features = image_features.to(inputs_embeds.dtype)
            final_embedding = inputs_embeds.masked_scatter(special_image_mask, image_features)
            final_attention_mask = attention_mask

        return final_embedding, final_attention_mask, position_ids

    def get_text_embeddings(self, input_ids, **kwargs):
        for_inputs_embeds_ids = input_ids.clone()
        for_inputs_embeds_ids[(input_ids == self.config.image_token_index)] = 0
        return super().get_text_embeddings(for_inputs_embeds_ids, **kwargs)


class _OVLlavaNextVideoForCausalLM(_OVLlavaNextForCausalLM):
    additional_parts = ["vision_resampler", "multi_modal_projector"]

    def get_vision_embeddings(self, pixel_values, input_ids=None, **kwargs):
        if input_ids is not None and input_ids.shape[1] == 1:
            return None
        image_features = self.vision_embeddings(pixel_values).last_hidden_state
        image_features = self.multi_modal_projector(image_features)
        return image_features

    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).
            vision_feature_select_strategy (`str`)
                The feature selection strategy used to select the vision feature from the vision backbone.
            image_newline (`torch.Tensor` of shape `(embed_dim)`)
                New line embedding vector.
        Returns:
            image_features (`torch.Tensor` of shape `(all_feat_len, embed_dim)`)
            feature_lens (`List[int]`)
                token length of each image in image_features
        """
        from transformers.models.llava_next_video.modeling_llava_next_video import (
            get_anyres_image_grid_shape,
            unpad_image,
        )

        new_image_features = []
        feature_lens = []
        vision_feature_select_strategy = self.config.vision_feature_select_strategy
        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

                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,
                )

                if (
                    np.prod(image_feature.shape) % (num_patch_height * num_patch_width * height * width) != 0
                    and vision_feature_select_strategy == "default"
                ):
                    logger.warning_once(
                        "Image feature shape does not line up with the provided patch size. "
                        "You may be using the `default` vision_feature_select_strategy with a"
                        " visual encoder that does not have CLS."
                    )

                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.device, 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

    @staticmethod
    def preprocess_inputs(
        text: str,
        image: Optional["Image"] = None,
        processor: Optional[AutoImageProcessor] = None,
        tokenizer: Optional[PreTrainedTokenizer] = None,
        config: Optional[PretrainedConfig] = None,
        video: Optional["VideoInput"] = None,
        audio: Optional[np.ndarray] = None,
    ):
        if processor is None:
            raise ValueError("Processor is required.")
        if audio is not None:
            raise ValueError("Audio input is not supported")
        if getattr(processor, "chat_template", None) is not None:
            chat_prompt = [{"role": "user", "content": [{"type": "text", "text": text}]}]
            if image is not None:
                chat_prompt[0]["content"].append({"type": "image"})
            if video is not None:
                chat_prompt[0]["content"].append({"type": "video"})
            prompt = processor.apply_chat_template(chat_prompt, add_generation_prompt=True, tokenize=False)
        else:
            prompt = text
            if image is not None and "<image>" not in prompt:
                prompt = "<image>\n" + prompt
            if video is not None and "<video>" not in prompt:
                prompt = "<video>\n" + prompt

        if is_transformers_version(">", "4.47.99") and getattr(processor, "patch_size", None) is None:
            if (
                getattr(config, "vision_config", None) is not None
                and getattr(config.vision_config, "patch_size", None) is not None
            ):
                processor.patch_size = config.vision_config.patch_size
            else:
                raise ValueError(
                    "Processor does not have `patch_size` attribute. Please fix the processor or provide `patch_size` in the config."
                )

        inputs = processor(images=image, text=prompt, videos=video, return_tensors="pt")
        return inputs

    def get_multimodal_embeddings(
        self,
        input_ids,
        pixel_values=None,
        attention_mask=None,
        position_ids=None,
        past_key_values=None,
        image_sizes=None,
        pixel_values_videos=None,
        **kwargs,
    ):
        embeds_from_args = kwargs.pop("inputs_embeds", None)
        inputs_embeds = (
            embeds_from_args if embeds_from_args is not None else self.get_text_embeddings(input_ids, **kwargs)
        )

        if (
            pixel_values is not None
            and pixel_values.size(0) > 0
            and self._support_new_processing
            and past_key_values is None
        ):
            legacy_processing = (
                (input_ids == self.config.image_token_index).sum(1).max() < self.config.image_seq_length
            ).item()
        elif (
            pixel_values_videos is not None
            and pixel_values_videos.size(0) > 0
            and self._support_new_processing
            and past_key_values is None
        ):
            legacy_processing = (
                (input_ids == self.config.video_token_index).sum(1).max() < self.config.video_seq_length
            ).item()
        else:
            legacy_processing = True

        legacy_processing = (
            legacy_processing.item() if isinstance(legacy_processing, torch.Tensor) else legacy_processing
        )

        if pixel_values is not None and pixel_values.size(0) > 0:
            inputs_embeds, attention_mask, position_ids = self.add_image_features(
                input_ids,
                inputs_embeds,
                pixel_values,
                attention_mask,
                position_ids,
                image_sizes,
                legacy_processing,
                **kwargs,
            )

        if pixel_values_videos is not None and pixel_values_videos.size(0) > 0:
            inputs_embeds, attention_mask, position_ids = self.add_video_features(
                input_ids,
                inputs_embeds,
                pixel_values_videos,
                attention_mask,
                position_ids,
                legacy_processing=legacy_processing,
                **kwargs,
            )

        if legacy_processing and pixel_values is not None and past_key_values is not None and input_ids.shape[1] == 1:
            attention_mask, position_ids = self._filter_unattended_tokens(input_ids, attention_mask, past_key_values)

        return inputs_embeds, attention_mask, position_ids

    def add_video_features(
        self,
        input_ids,
        inputs_embeds,
        pixel_values_videos,
        attention_mask,
        position_ids,
        legacy_processing,
        **kwargs,
    ):
        # Adopted from https://github.com/huggingface/transformers/blob/v4.49.0/src/transformers/models/llava_next_video/modeling_llava_next_video.py#L732-L751
        video_features = self.get_video_features(pixel_values_videos, input_ids)
        if video_features is not None and len(video_features) != 0:
            video_features = [feature.flatten(0, 1) for feature in video_features]
            video_feature_lens = [feature.size(0) for feature in video_features]
            video_features = torch.cat(video_features, dim=0)
            video_feature_lens = torch.tensor(video_feature_lens, dtype=torch.long, device=video_features.device)

            if legacy_processing:
                inputs_embeds, attention_mask, position_ids = self.merge_vision_text_embeddings(
                    video_features,
                    inputs_embeds,
                    video_feature_lens,
                    input_ids,
                    attention_mask,
                    position_ids,
                    legacy_processing,
                    self.config.video_token_index,
                )
            else:
                inputs_embeds = (
                    torch.from_numpy(inputs_embeds) if isinstance(inputs_embeds, np.ndarray) else inputs_embeds
                )
                special_image_mask = (input_ids == self.config.video_token_index).unsqueeze(-1)
                special_image_mask = special_image_mask.expand_as(inputs_embeds)
                if inputs_embeds[special_image_mask].numel() != video_features.numel():
                    n_video_tokens = (input_ids == self.config.video_token_index).sum().item()
                    n_video_features = video_features.shape[0]
                    raise ValueError(
                        f"Video features and video tokens do not match: tokens: {n_video_tokens}, features {n_video_features}"
                    )
                inputs_embeds = inputs_embeds.masked_scatter(special_image_mask, video_features)
        return inputs_embeds, attention_mask, position_ids

    def get_video_features(self, pixel_values, input_ids=None, **kwargs):
        # Adopted from https://github.com/huggingface/transformers/blob/v4.49.0/src/transformers/models/llava_next_video/modeling_llava_next_video.py#L835
        if input_ids is not None and input_ids.shape[1] == 1:
            return None
        batch_size, frames, channels, height, width = pixel_values.shape
        pixel_values = pixel_values.reshape(batch_size * frames, channels, height, width)
        selected_video_features = self.vision_embeddings(pixel_values).last_hidden_state
        video_features = self.vision_resampler(selected_video_features)
        video_features = self.multi_modal_projector(video_features)
        video_features = torch.split(torch.from_numpy(video_features), frames, dim=0)
        return video_features


class _OVInternVLForCausalLM(OVModelForVisualCausalLM):
    def get_vision_embeddings(self, pixel_values, input_ids=None, **kwargs):
        if input_ids is not None and input_ids.shape[1] == 1:
            return None
        image_features = self.vision_embeddings(pixel_values, **kwargs).last_hidden_state
        return image_features

    def merge_vision_text_embeddings(
        self, vision_embeds, input_embeds, input_ids, attention_mask, position_ids=None, **kwargs
    ):
        input_embeds = torch.from_numpy(input_embeds) if isinstance(input_embeds, np.ndarray) else input_embeds
        vision_embeds = torch.from_numpy(vision_embeds) if isinstance(vision_embeds, np.ndarray) else vision_embeds
        B, N, C = input_embeds.shape
        input_embeds = input_embeds.reshape(B * N, C)

        input_ids = input_ids.reshape(B * N)
        selected = input_ids == self.config.img_context_token_id
        assert selected.sum() != 0
        input_embeds[selected] = vision_embeds.reshape(-1, C)

        input_embeds = input_embeds.reshape(B, N, C)
        return input_embeds, attention_mask, position_ids

    @staticmethod
    def preprocess_inputs(
        text: str,
        image: Optional["Image"] = None,
        processor: Optional[AutoImageProcessor] = None,
        tokenizer: Optional[PreTrainedTokenizer] = None,
        config: Optional[PretrainedConfig] = None,
        video: Optional["VideoInput"] = None,
        audio: Optional[np.ndarray] = None,
    ):
        if tokenizer is None:
            raise ValueError("Tokenizer is required.")
        if video is not None:
            raise ValueError("Video input is not supported")
        if audio is not None:
            raise ValueError("Audio input is not supported")
        import torchvision.transforms as T
        from torchvision.transforms.functional import InterpolationMode

        IMG_START_TOKEN = "<img>"
        IMG_END_TOKEN = "</img>"
        IMG_CONTEXT_TOKEN = "<IMG_CONTEXT>"

        IMAGENET_MEAN = (0.485, 0.456, 0.406)
        IMAGENET_STD = (0.229, 0.224, 0.225)

        def build_transform(input_size):
            MEAN, STD = IMAGENET_MEAN, IMAGENET_STD
            transform = T.Compose(
                [
                    T.Lambda(lambda img: img.convert("RGB") if img.mode != "RGB" else img),
                    T.Resize((input_size, input_size), interpolation=InterpolationMode.BICUBIC),
                    T.ToTensor(),
                    T.Normalize(mean=MEAN, std=STD),
                ]
            )
            return transform

        def find_closest_aspect_ratio(aspect_ratio, target_ratios, width, height, image_size):
            best_ratio_diff = float("inf")
            best_ratio = (1, 1)
            area = width * height
            for ratio in target_ratios:
                target_aspect_ratio = ratio[0] / ratio[1]
                ratio_diff = abs(aspect_ratio - target_aspect_ratio)
                if ratio_diff < best_ratio_diff:
                    best_ratio_diff = ratio_diff
                    best_ratio = ratio
                elif ratio_diff == best_ratio_diff:
                    if area > 0.5 * image_size * image_size * ratio[0] * ratio[1]:
                        best_ratio = ratio
            return best_ratio

        def dynamic_preprocess(image, min_num=1, max_num=12, image_size=28, use_thumbnail=False):
            orig_width, orig_height = image.size
            aspect_ratio = orig_width / orig_height

            # calculate the existing image aspect ratio
            target_ratios = {
                (i, j)
                for n in range(min_num, max_num + 1)
                for i in range(1, n + 1)
                for j in range(1, n + 1)
                if i * j <= max_num and i * j >= min_num
            }
            target_ratios = sorted(target_ratios, key=lambda x: x[0] * x[1])

            # find the closest aspect ratio to the target
            target_aspect_ratio = find_closest_aspect_ratio(
                aspect_ratio, target_ratios, orig_width, orig_height, image_size
            )

            # calculate the target width and height
            target_width = image_size * target_aspect_ratio[0]
            target_height = image_size * target_aspect_ratio[1]
            blocks = target_aspect_ratio[0] * target_aspect_ratio[1]

            # resize the image
            resized_img = image.resize((target_width, target_height))
            processed_images = []
            for i in range(blocks):
                box = (
                    (i % (target_width // image_size)) * image_size,
                    (i // (target_width // image_size)) * image_size,
                    ((i % (target_width // image_size)) + 1) * image_size,
                    ((i // (target_width // image_size)) + 1) * image_size,
                )
                # split the image
                split_img = resized_img.crop(box)
                processed_images.append(split_img)
            assert len(processed_images) == blocks
            if use_thumbnail and len(processed_images) != 1:
                thumbnail_img = image.resize((image_size, image_size))
                processed_images.append(thumbnail_img)
            return processed_images

        def load_image(image, input_size=448, max_num=12):
            transform = build_transform(input_size=input_size)
            images = dynamic_preprocess(image, image_size=input_size, use_thumbnail=True, max_num=max_num)
            pixel_values = [transform(image) for image in images]
            pixel_values = torch.stack(pixel_values)
            return pixel_values

        if image is not None and "<image>" not in text:
            text = "<image>\n" + text

        if tokenizer.chat_template is not None:
            text = tokenizer.apply_chat_template(
                [{"role": "user", "content": text}], add_generation_prompt=True, tokenize=False
            )

        inputs = {}

        if image is not None:
            if config is None:
                raise ValueError("Config is required.")
            pixel_values = load_image(image, input_size=config.vision_config.image_size)
            num_patches = pixel_values.shape[0]
            num_image_token = int(
                (config.vision_config.image_size // config.vision_config.patch_size) ** 2
                * (config.downsample_ratio**2)
            )
            image_tokens = IMG_START_TOKEN + IMG_CONTEXT_TOKEN * num_image_token * num_patches + IMG_END_TOKEN
            text = text.replace("<image>", image_tokens, 1)
            inputs.update({"pixel_values": pixel_values})
        inputs.update(tokenizer(text, return_tensors="pt"))
        return inputs

    # internvl has issue with check  _get_non_default_parameters, as workaround override _prepare_generation_config
    def _prepare_generation_config(
        self, generation_config: Optional[GenerationConfig], use_model_defaults: Optional[bool] = None, **kwargs: Dict
    ) -> Tuple[GenerationConfig, Dict]:
        using_model_generation_config = False
        if generation_config is None:
            if (
                self.generation_config._from_model_config  # 1)
                and self.generation_config._original_object_hash == hash(self.generation_config)  # 2)
            ):
                new_generation_config = GenerationConfig.from_model_config(self.config)
                if new_generation_config != self.generation_config:  # 4)
                    warnings.warn(
                        "You have modified the pretrained model configuration to control generation. This is a"
                        " deprecated strategy to control generation and will be removed in v5."
                        " Please use and modify the model generation configuration (see"
                        " https://huggingface.co/docs/transformers/generation_strategies#default-text-generation-configuration )",
                        UserWarning,
                    )
                    self.generation_config = new_generation_config

            generation_config = self.generation_config
            using_model_generation_config = True

        generation_config = copy.deepcopy(generation_config)
        model_kwargs = generation_config.update(**kwargs)
        # If `generation_config` is provided, let's fallback ALL special tokens to the default values for the model
        if not using_model_generation_config:
            if generation_config.bos_token_id is None:
                generation_config.bos_token_id = self.generation_config.bos_token_id
            if generation_config.eos_token_id is None:
                generation_config.eos_token_id = self.generation_config.eos_token_id
            if generation_config.pad_token_id is None:
                generation_config.pad_token_id = self.generation_config.pad_token_id
            if generation_config.decoder_start_token_id is None:
                generation_config.decoder_start_token_id = self.generation_config.decoder_start_token_id

        return generation_config, model_kwargs


class _OVMiniCPMVForCausalLM(OVModelForVisualCausalLM):
    additional_parts = ["resampler"]

    def __init__(
        self,
        language_model: ov.Model,
        text_embeddings: ov.Model,
        vision_embeddings: ov.Model,
        config: PretrainedConfig = None,
        device: str = "CPU",
        dynamic_shapes: bool = None,
        ov_config: Optional[Dict[str, str]] = None,
        model_save_dir: Optional[Union[str, Path, TemporaryDirectory]] = None,
        quantization_config: Union[OVWeightQuantizationConfig, Dict] = None,
        **kwargs,
    ):
        super().__init__(
            language_model=language_model,
            text_embeddings=text_embeddings,
            vision_embeddings=vision_embeddings,
            config=config,
            device=device,
            dynamic_shapes=dynamic_shapes,
            ov_config=ov_config,
            model_save_dir=model_save_dir,
            quantization_config=quantization_config,
            **kwargs,
        )

        self.embed_dim = self.language_model.config.hidden_size
        max_size = self.config.vision_config.image_size // self.config.vision_config.patch_size
        self._pos_embeds = torch.from_numpy(self._get_2d_sincos_pos_embed(self.embed_dim, max_size)).float()
        self.max_size = (max_size, max_size)

    def get_vision_embeddings(self, pixel_values, input_ids=None, **kwargs):
        if input_ids is not None and input_ids.shape[1] == 1:
            return None
        tgt_sizes = kwargs["tgt_sizes"]
        pixel_values_list = pixel_values
        vision_hidden_states = []
        all_pixel_values = []
        img_cnt = []
        for pixel_value in pixel_values_list:
            img_cnt.append(len(pixel_value))
            all_pixel_values.extend([i.flatten(end_dim=1).permute(1, 0) for i in pixel_value])

        vision_embedding = None
        # exist image
        if all_pixel_values:
            tgt_sizes = [tgt_size for tgt_size in tgt_sizes if isinstance(tgt_size, torch.Tensor)]
            tgt_sizes = torch.vstack(tgt_sizes).type(torch.int32)

            max_patches = torch.max(tgt_sizes[:, 0] * tgt_sizes[:, 1])

            all_pixel_values = torch.nn.utils.rnn.pad_sequence(all_pixel_values, batch_first=True, padding_value=0.0)
            B, L, _ = all_pixel_values.shape
            all_pixel_values = all_pixel_values.permute(0, 2, 1).reshape(B, 3, -1, L)

            patch_attn_mask = torch.zeros((B, 1, max_patches), dtype=torch.bool)
            for i in range(B):
                patch_attn_mask[i, 0, : tgt_sizes[i][0] * tgt_sizes[i][1]] = True
            position_ids = self._prepare_vis_position_ids(
                all_pixel_values,
                patch_attn_mask,
                tgt_sizes,
                self.config.vision_config.patch_size,
                self.config.vision_config.image_size // self.config.patch_size,
            )
            vision_embedding = torch.from_numpy(
                self.vision_embeddings(
                    pixel_values=all_pixel_values, patch_attention_mask=patch_attn_mask, position_ids=position_ids
                )[0]
            )
            vision_embedding = self.resampling(vision_embedding, tgt_sizes)

            start = 0
            for pixel_value in pixel_values_list:
                img_cnt = len(pixel_value)
                if img_cnt > 0:
                    vision_hidden_states.append(vision_embedding[start : start + img_cnt])
                    start += img_cnt
                else:
                    vision_hidden_states.append([])
        else:  # no image
            dummy_feature = []
            for _ in range(len(pixel_values_list)):
                vision_hidden_states.append(dummy_feature)
        return vision_hidden_states

    def resampling(self, x, tgt_sizes):
        bs = x.shape[0]

        patch_len = tgt_sizes[:, 0] * tgt_sizes[:, 1]

        self._adjust_pos_cache(tgt_sizes)

        max_patch_len = torch.max(patch_len)
        key_padding_mask = torch.zeros((bs, max_patch_len), dtype=torch.bool)

        pos_embed = []
        for i in range(bs):
            tgt_h, tgt_w = tgt_sizes[i]
            pos_embed.append(self._pos_embeds[:tgt_h, :tgt_w, :].reshape((tgt_h * tgt_w, -1)))  # patches * D
            key_padding_mask[i, patch_len[i] :] = True

        pos_embed = torch.nn.utils.rnn.pad_sequence(pos_embed, batch_first=True, padding_value=0.0).permute(
            1, 0, 2
        )  # BLD => L * B * D
        res = torch.from_numpy(self.resampler(image_feature=x, pos_embed=pos_embed, key_padding_mask=key_padding_mask))
        return res

    def _set_2d_pos_cache(self, max_size):
        pos_embed = torch.from_numpy(self._get_2d_sincos_pos_embed(self.embed_dim, max_size)).float()
        self._pos_embeds = pos_embed

    def _adjust_pos_cache(self, tgt_sizes):
        max_h = torch.max(tgt_sizes[:, 0])
        max_w = torch.max(tgt_sizes[:, 1])
        if max_h > self.max_size[0] or max_w > self.max_size[1]:
            self.max_size = [max(max_h, self.max_size[0]), max(max_w, self.max_size[1])]
            self._set_2d_pos_cache(self.max_size)

    def _get_2d_sincos_pos_embed(self, embed_dim, image_size):
        """
        image_size: image_size or (image_height, image_width)
        return:
        pos_embed: [image_height, image_width, embed_dim]
        """
        if isinstance(image_size, int):
            grid_h_size, grid_w_size = image_size, image_size
        else:
            grid_h_size, grid_w_size = image_size[0], image_size[1]

        grid_h = np.arange(grid_h_size, dtype=np.float32)
        grid_w = np.arange(grid_w_size, dtype=np.float32)
        grid = np.meshgrid(grid_w, grid_h)  # here w goes first
        grid = np.stack(grid, axis=0)

        pos_embed = self._get_2d_sincos_pos_embed_from_grid(embed_dim, grid)
        return pos_embed

    def _get_2d_sincos_pos_embed_from_grid(self, embed_dim, grid):
        assert embed_dim % 2 == 0

        # use half of dimensions to encode grid_h
        emb_h = self._get_1d_sincos_pos_embed_from_grid_new(embed_dim // 2, grid[0])  # (H, W, D/2)
        emb_w = self._get_1d_sincos_pos_embed_from_grid_new(embed_dim // 2, grid[1])  # (H, W, D/2)

        emb = np.concatenate([emb_h, emb_w], axis=-1)  # (H, W, D)
        return emb

    def _get_1d_sincos_pos_embed_from_grid_new(self, embed_dim, pos):
        """
        embed_dim: output dimension for each position
        pos: a list of positions to be encoded: size (H, W)
        out: (H, W, D)
        """
        assert embed_dim % 2 == 0
        omega = np.arange(embed_dim // 2, dtype=np.float32)
        omega /= embed_dim / 2.0
        omega = 1.0 / 10000**omega  # (D/2,)

        out = np.einsum("hw,d->hwd", pos, omega)  # (H, W, D/2), outer product

        emb_sin = np.sin(out)  # (H, W, D/2)
        emb_cos = np.cos(out)  # (H, W, D/2)

        emb = np.concatenate([emb_sin, emb_cos], axis=-1)  # (H, W, D)
        return emb

    def _prepare_vis_position_ids(
        self, pixel_values, patch_attention_mask, tgt_sizes, patch_size, num_patches_per_side
    ):
        batch_size = pixel_values.size(0)
        max_im_h, max_im_w = pixel_values.size(2), pixel_values.size(3)
        max_nb_patches_h, max_nb_patches_w = max_im_h // patch_size, max_im_w // patch_size
        boundaries = torch.arange(1 / num_patches_per_side, 1.0, 1 / num_patches_per_side)
        position_ids = torch.full(size=(batch_size, max_nb_patches_h * max_nb_patches_w), fill_value=0)

        for batch_idx, p_attn_mask in enumerate(patch_attention_mask):
            if tgt_sizes is not None:
                nb_patches_h = tgt_sizes[batch_idx][0]
                nb_patches_w = tgt_sizes[batch_idx][1]
            else:
                nb_patches_h = p_attn_mask[:, 0].sum()
                nb_patches_w = p_attn_mask[0].sum()

            fractional_coords_h = torch.arange(0, 1 - 1e-6, 1 / nb_patches_h)
            fractional_coords_w = torch.arange(0, 1 - 1e-6, 1 / nb_patches_w)

            bucket_coords_h = torch.bucketize(fractional_coords_h, boundaries, right=True)
            bucket_coords_w = torch.bucketize(fractional_coords_w, boundaries, right=True)

            pos_ids = (bucket_coords_h[:, None] * num_patches_per_side + bucket_coords_w).flatten()
            position_ids[batch_idx][p_attn_mask.view(-1).cpu()] = pos_ids

        return position_ids

    def merge_vision_text_embeddings(
        self, vision_embeds, input_embeds, input_ids, attention_mask, position_ids=None, **kwargs
    ):
        bs = input_ids.shape[0]
        image_bound = kwargs["image_bound"]
        vllm_embedding = torch.from_numpy(input_embeds)
        for i in range(bs):
            cur_vs_hs = vision_embeds[i]
            if len(cur_vs_hs) > 0:
                cur_vllm_emb = vllm_embedding[i]
                cur_image_bound = image_bound[i]
                if len(cur_image_bound) > 0:
                    image_indices = torch.stack([torch.arange(r[0], r[1], dtype=torch.long) for r in cur_image_bound])

                    cur_vllm_emb.scatter_(
                        0,
                        image_indices.view(-1, 1).repeat(1, cur_vllm_emb.shape[-1]),
                        cur_vs_hs.view(-1, cur_vs_hs.shape[-1]),
                    )
        return vllm_embedding, attention_mask, position_ids

    @staticmethod
    def preprocess_inputs(
        text: str,
        image: Optional["Image"] = None,
        processor: Optional[AutoImageProcessor] = None,
        tokenizer: Optional[PreTrainedTokenizer] = None,
        config: Optional[PretrainedConfig] = None,
        video: Optional["VideoInput"] = None,
        audio: Optional[np.ndarray] = None,
    ):
        if processor is None:
            raise ValueError("Processor is required.")
        if video is not None:
            raise ValueError("Video input is not supported")
        if audio is not None:
            raise ValueError("Audio input is not supported")
        if getattr(processor, "chat_template", None) is not None:
            messages = [{"role": "user", "content": text if image is None else "(<image>./</image>)\n" + text}]
            prompt = processor.apply_chat_template(messages, tokenize=False, add_generation_prompt=True)
        else:
            prompt = (
                f"<|im_start|>user\n(<image>./</image>)\n{text}<|im_end|>\n<|im_start|>assistant\n"
                if image is not None
                else text
            )
        inputs = processor([prompt], [image], return_tensors="pt")
        inputs.pop("image_sizes", None)
        return inputs


class _OVMiniCPMOForCausalLM(_OVMiniCPMVForCausalLM):
    def prepare_inputs_for_generation(
        self,
        input_ids,
        past_key_values=None,
        inputs_embeds=None,
        pixel_values=None,
        image_sizes=None,
        attention_mask=None,
        audio_bounds=None,
        spk_bounds=None,
        audio_features=None,
        audio_feature_lens=None,
        **kwargs,
    ):
        # Audio modality is not supported for MiniCPMO
        if audio_features is not None and len(audio_features) > 0:
            raise ValueError("Audio input is not supported for MiniCPMO")

        return super().prepare_inputs_for_generation(
            input_ids=input_ids,
            past_key_values=past_key_values,
            inputs_embeds=inputs_embeds,
            pixel_values=pixel_values,
            image_sizes=image_sizes,
            attention_mask=attention_mask,
            **kwargs,
        )

    @staticmethod
    def preprocess_inputs(
        text: str,
        image: Optional["Image"] = None,
        processor: Optional[AutoImageProcessor] = None,
        tokenizer: Optional[PreTrainedTokenizer] = None,
        config: Optional[PretrainedConfig] = None,
        video: Optional["VideoInput"] = None,
        audio: Optional[np.ndarray] = None,
    ):
        if processor is None:
            raise ValueError("Processor is required.")
        if video is not None:
            raise ValueError("Video input is not supported")
        if audio is not None:
            raise ValueError("Audio input is not supported")
        messages = [{"role": "user", "content": text if image is None else "(<image>./</image>)\n" + text}]
        prompt = processor.tokenizer.apply_chat_template(messages, tokenize=False, add_generation_prompt=True)
        inputs = processor([prompt], [image], return_tensors="pt")
        inputs.pop("image_sizes", None)
        return inputs


class _OVNanoLlavaForCausalLM(OVModelForVisualCausalLM):
    def get_vision_embeddings(self, pixel_values, input_ids=None, **kwargs):
        if input_ids is not None and input_ids.shape[1] == 1:
            return None
        if isinstance(pixel_values, list) or pixel_values.ndim == 5:
            concat_images = torch.cat(pixel_values, dim=0) if isinstance(pixel_values, list) else pixel_values
            image_features = torch.from_numpy(self.vision_embeddings(concat_images).last_hidden_state)
            split_sizes = [image.shape[0] for image in pixel_values]
            image_features = torch.split(image_features, split_sizes, dim=0)
            image_features = [x.flatten(0, 1).to(self.device) for x in image_features]
        else:
            image_features = self.vision_embeddings(pixel_values).last_hidden_state

        return image_features

    def get_multimodal_embeddings(
        self, input_ids, pixel_values=None, attention_mask=None, position_ids=None, **kwargs
    ):
        vision_embeds = None
        IGNORE_INDEX = -100
        IMAGE_TOKEN_INDEX = -200
        if pixel_values is None and "images" in kwargs:
            pixel_values = kwargs["images"]
        if pixel_values is not None:
            vision_embeds = self.get_vision_embeddings(pixel_values, input_ids=input_ids, **kwargs)
        if vision_embeds is None:
            inputs_embeds = torch.from_numpy(self.get_text_embeddings(input_ids))
            past_len = self.language_model._get_past_length(kwargs.get("past_key_values"))
            if attention_mask is not None and attention_mask.shape[1] < past_len + input_ids.shape[1]:
                attention_mask = torch.cat(
                    [
                        attention_mask,
                        torch.ones(attention_mask.shape[0], past_len + input_ids.shape[1] - attention_mask.shape[1]),
                    ],
                    dim=1,
                )
                position_ids = None
            return inputs_embeds, attention_mask, position_ids

        vision_embeds = torch.from_numpy(vision_embeds) if isinstance(vision_embeds, np.ndarray) else vision_embeds

        if attention_mask is None:
            attention_mask = torch.ones_like(input_ids, dtype=torch.long)
        if position_ids is None:
            position_ids = torch.arange(0, input_ids.shape[1], dtype=torch.long, device=input_ids.device)
        labels = torch.full_like(input_ids, IGNORE_INDEX)

        # remove the padding using attention_mask -- TODO: double check
        input_ids = [
            cur_input_ids[cur_attention_mask]
            for cur_input_ids, cur_attention_mask in zip(input_ids, attention_mask.bool())
        ]
        labels = [
            cur_labels[cur_attention_mask] for cur_labels, cur_attention_mask in zip(labels, attention_mask.bool())
        ]

        new_input_embeds = []
        new_labels = []
        cur_image_idx = 0
        for batch_idx, cur_input_ids in enumerate(input_ids):
            num_images = (cur_input_ids == IMAGE_TOKEN_INDEX).sum()
            if num_images == 0:
                cur_image_features = vision_embeds[cur_image_idx]
                cur_input_embeds_1 = torch.from_numpy(self.get_text_embeddings(cur_input_ids.unsqueeze(0))[0])
                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 == IMAGE_TOKEN_INDEX)[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 = torch.from_numpy(
                self.get_text_embeddings(torch.cat(cur_input_ids_noim).unsqueeze(0))[0]
            )
            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 = vision_embeds[cur_image_idx]
                    cur_image_idx += 1
                    cur_new_input_embeds.append(cur_image_features)
                    cur_new_labels.append(
                        torch.full(
                            (cur_image_features.shape[0],),
                            IGNORE_INDEX,
                            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]

        # Combine them
        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), IGNORE_INDEX, 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)

        return new_input_embeds, attention_mask, position_ids

    @staticmethod
    def preprocess_inputs(
        text: str,
        image: Optional["Image"] = None,
        processor: Optional[AutoImageProcessor] = None,
        tokenizer: Optional[PreTrainedTokenizer] = None,
        config: Optional[PretrainedConfig] = None,
        video: Optional["VideoInput"] = None,
        audio: Optional[np.ndarray] = None,
    ):
        if tokenizer is None:
            raise ValueError("Tokenizer is required.")
        if video is not None:
            raise ValueError("Video input is not supported")
        if audio is not None:
            raise ValueError("Audio input is not supported")
        if image is not None and processor is None:
            raise ValueError("Processor is required.")
        text = f"<image>\n{text}" if image is not None else text
        messages = [{"role": "user", "content": text}]
        if tokenizer.chat_template is not None:
            text = tokenizer.apply_chat_template(messages, tokenize=False, add_generation_prompt=True)
        if image is not None:
            text_chunks = [tokenizer(chunk).input_ids for chunk in text.split("<image>")]
            input_ids = torch.tensor(text_chunks[0] + [-200] + text_chunks[1], dtype=torch.long).unsqueeze(0)
        else:
            input_ids = tokenizer(text, return_tensors="pt").input_ids
        attention_mask = torch.ones_like(input_ids, dtype=torch.int64)
        result = {"input_ids": input_ids, "attention_mask": attention_mask}

        if image is not None:
            result["images"] = processor(images=[image], return_tensors="pt")["pixel_values"]
        return result


class _OVPhi3VisionForCausalLM(OVModelForVisualCausalLM):
    additional_parts = ["vision_projection"]

    def __init__(
        self,
        language_model: ov.Model,
        text_embeddings: ov.Model,
        vision_embeddings: ov.Model,
        config: PretrainedConfig = None,
        device: str = "CPU",
        dynamic_shapes: bool = None,
        ov_config: Optional[Dict[str, str]] = None,
        model_save_dir: Optional[Union[str, Path, TemporaryDirectory]] = None,
        quantization_config: Union[OVWeightQuantizationConfig, Dict] = None,
        **kwargs,
    ):
        super().__init__(
            language_model=language_model,
            text_embeddings=text_embeddings,
            vision_embeddings=vision_embeddings,
            config=config,
            device=device,
            dynamic_shapes=dynamic_shapes,
            ov_config=ov_config,
            model_save_dir=model_save_dir,
            quantization_config=quantization_config,
            **kwargs,
        )

        self.sub_GN = torch.tensor(self.config.sub_GN)
        self.glb_GN = torch.tensor(self.config.glb_GN)
        self.image_dim_out = self.config.img_processor["image_dim_out"]

    def get_vision_embeddings(self, pixel_values, image_sizes, **kwargs):
        num_images, num_crops, c, h, w = pixel_values.shape
        img_features = self.vision_embeddings(pixel_values.flatten(0, 1)).last_hidden_state.reshape(
            num_images, num_crops, -1, self.image_dim_out
        )
        image_features_proj = self.hd_feature_transform(img_features, image_sizes)
        return image_features_proj

    def hd_feature_transform(self, image_features, image_sizes):
        """
        image_features: (num_images, num_crops+1, 24*24, 1024)
        """

        image_features = torch.from_numpy(image_features)
        global_image_features = image_features[:, 0]  # (num_images, 24*24, 1024)
        # global feature can be viewed as a special HD case with num_crops 1x1
        global_image_features_hd = self.reshape_hd_patches_2x2merge(global_image_features, 1, 1)
        global_image_features_hd_newline = self.add_image_newline(global_image_features_hd)

        all_image_embeddings = []
        # need a for loop to process each image because of different image sizes
        # (patch arrangement is different for each image)
        for i, img_size in enumerate(image_sizes):
            h, w = img_size
            h_crop = h // 336
            w_crop = w // 336
            num_crops = h_crop * w_crop

            # NOTE: real num_crops is padded
            # (num_crops, 24*24, 1024)
            sub_image_features = image_features[i, 1 : 1 + num_crops]
            sub_image_features_hd = self.reshape_hd_patches_2x2merge(sub_image_features, h_crop, w_crop)
            sub_image_features_hd_newline = self.add_image_newline(sub_image_features_hd)

            # [sub features, separator, global features]
            all_image_embeddings.extend(
                [
                    sub_image_features_hd_newline.squeeze(0),  # (h_crop*12*(w_crop*12+1), 4096)
                    self.glb_GN.squeeze(0),
                    global_image_features_hd_newline[i],
                ]
            )
        image_features_proj = self.vision_projection(torch.cat(all_image_embeddings, dim=0).unsqueeze(0))[0]

        return image_features_proj

    def reshape_hd_patches_2x2merge(self, image_features, h_crop, w_crop):
        """
        image_features: (num_images*num_crops, 24*24, 1024)
        output: (num_images, h_crop*12, w_crop*12, 4096), h_crop*w_crop == num_crops
        """
        N, L, C = image_features.shape
        assert L == 24 * 24 and C == 1024 and N % (h_crop * w_crop) == 0
        num_images = N // (h_crop * w_crop)
        H = int(L**0.5)
        image_features_hd = (
            image_features.reshape(N, H, H, C)  # N, 24, 24, 1024
            .reshape(N, H // 2, 2, H // 2, 2, C)  # N, 12, 2, 12, 2, 1024
            .permute(0, 1, 3, 2, 4, 5)  # N, 12, 12, 2, 2, 1024
            .reshape(N, -1, 4 * C)  # N, 144, 4096
            .reshape(num_images, h_crop, w_crop, H // 2, H // 2, -1)  # n_img, h_crop, w_crop, 12, 12, 4096
            .permute(0, 1, 3, 2, 4, 5)  # n_img, h_crop, 12, w_crop, 12, 4096
            .reshape(num_images, h_crop * H // 2, w_crop * H // 2, 4 * C)  # n_img, h_crop*12, w_crop*12, 4096
        )

        return image_features_hd

    def add_image_newline(self, image_features_hd):
        """
        image_features_hd: (num_images, h_crop*12, w_crop*12, 4096)
        output: (num_images, (h_crop*12) * (w_crop*12+1), 4096)
        """
        num_images, h, w, hid_dim = image_features_hd.shape
        # add the newline token to the HD image feature patches
        newline_embeddings = self.sub_GN.expand(num_images, h, -1, -1)  # (n_img, h, 1, hid_dim)
        image_features_hd_newline = torch.cat([image_features_hd, newline_embeddings], dim=2).reshape(
            num_images, -1, hid_dim
        )
        return image_features_hd_newline

    def get_multimodal_embeddings(
        self, input_ids, pixel_values=None, attention_mask=None, position_ids=None, image_sizes=None, **kwargs
    ):
        MAX_INPUT_ID = int(1e9)
        input_shape = input_ids.size()
        input_ids = input_ids.view(-1, input_shape[-1])

        # positions for image tokens
        positions = torch.nonzero((input_ids < 0) & (input_ids > -MAX_INPUT_ID), as_tuple=True)
        has_image = len(positions[0].tolist()) > 0
        input_ids = input_ids.clamp_min(0).clamp_max(self.config.vocab_size)
        inputs_embeds = torch.from_numpy(self.get_text_embeddings(input_ids, **kwargs))
        if has_image:
            vision_embeds = self.get_vision_embeddings(
                pixel_values, input_ids=input_ids, image_sizes=image_sizes, **kwargs
            )
            image_features_proj = torch.from_numpy(vision_embeds)
            inputs_embeds = inputs_embeds.index_put(positions, image_features_proj, accumulate=False)

        return inputs_embeds, attention_mask, position_ids

    @staticmethod
    def preprocess_inputs(
        text: str,
        image: Optional["Image"] = None,
        processor: Optional[AutoImageProcessor] = None,
        tokenizer: Optional[PreTrainedTokenizer] = None,
        config: Optional[PretrainedConfig] = None,
        video: Optional["VideoInput"] = None,
        audio: Optional[np.ndarray] = None,
    ):
        if processor is None:
            raise ValueError("Processor is required.")
        if video is not None:
            raise ValueError("Video input is not supported")
        if audio is not None:
            raise ValueError("Audio input is not supported")
        if image is not None and "<|image_1|>" not in text:
            text = "<|image_1|>\n" + text
        if getattr(processor.tokenizer, "chat_template", None) is not None:
            chat_prompt = [{"role": "user", "content": text}]
            text = processor.tokenizer.apply_chat_template(chat_prompt, add_generation_prompt=True, tokenize=False)
        inputs = processor(images=image, text=text, return_tensors="pt")
        return inputs


@dataclass
class QWen2VLModelOutputWithPast(ModelOutput):
    loss: Optional[torch.FloatTensor] = None
    logits: torch.FloatTensor = None
    past_key_values: Optional[Tuple[Tuple[torch.FloatTensor]]] = None
    hidden_states: Optional[Tuple[torch.FloatTensor, ...]] = None
    attentions: Optional[Tuple[torch.FloatTensor, ...]] = None
    rope_deltas: Optional[torch.FloatTensor] = None
    second_per_grid_ts: Optional[torch.FloatTensor] = None


class _OVQwen2VLForCausalLM(OVModelForVisualCausalLM):
    get_rope_index = Qwen2VLModel.get_rope_index
    get_vision_position_ids = getattr(Qwen2_5_VLModel, "get_vision_position_ids", None)
    additional_parts = ["vision_embeddings_merger"]

    def __init__(
        self,
        language_model: ov.Model,
        text_embeddings: ov.Model,
        vision_embeddings: ov.Model,
        config: PretrainedConfig = None,
        device: str = "CPU",
        dynamic_shapes: bool = None,
        ov_config: Optional[Dict[str, str]] = None,
        model_save_dir: Optional[Union[str, Path, TemporaryDirectory]] = None,
        quantization_config: Union[OVWeightQuantizationConfig, Dict] = None,
        **kwargs,
    ):
        super().__init__(
            language_model=language_model,
            text_embeddings=text_embeddings,
            vision_embeddings=vision_embeddings,
            config=config,
            device=device,
            dynamic_shapes=dynamic_shapes,
            ov_config=ov_config,
            model_save_dir=model_save_dir,
            quantization_config=quantization_config,
            **kwargs,
        )
        self.rope_deltas = None  # cache rope_deltas here
        self._rotary_pos_emb = VisionRotaryEmbedding(
            self.config.vision_config.embed_dim // self.config.vision_config.num_heads // 2
        )

    def prepare_inputs_for_generation(
        self,
        input_ids,
        past_key_values=None,
        attention_mask=None,
        inputs_embeds=None,
        cache_position=None,
        position_ids=None,
        use_cache=True,
        pixel_values=None,
        pixel_values_videos=None,
        image_grid_thw=None,
        video_grid_thw=None,
        mm_token_type_ids=None,
        is_first_iteration=False,
        next_sequence_length=None,
        **kwargs,
    ):
        # reconstruct cache_position as partially removed in v5.3 and totally removed in v5.5
        if is_transformers_version(">=", "5.3") and (
            cache_position is None or (not is_first_iteration and cache_position[0] == 0)
        ):
            if next_sequence_length is not None:
                past_len = input_ids.shape[1] - next_sequence_length
                cache_position = torch.arange(past_len, past_len + next_sequence_length, device=input_ids.device)
            elif not is_first_iteration and attention_mask is not None:
                # v5.3 decode step: input_ids is already sliced to 1 token, use attention_mask length
                past_len = attention_mask.shape[1] - 1
                cache_position = torch.tensor([past_len], device=input_ids.device)
            else:
                cache_position = torch.arange(input_ids.shape[1], device=input_ids.device)

        # Overwritten -- in specific circumstances we don't want to forward image inputs to the model
        if past_key_values is not None:
            if inputs_embeds is not None and input_ids.shape[1] == 0:  # Exception 4
                inputs_embeds = inputs_embeds[:, -cache_position.shape[0] :]
            elif inputs_embeds is not None:
                input_ids = input_ids[:, -cache_position.shape[0] :]
            elif input_ids.shape[1] != cache_position.shape[0]:  # Default case (the "else", a no op, is Exception 2)
                input_ids = input_ids[:, cache_position]

        if cache_position[0] != 0:
            pixel_values = None
            pixel_values_videos = None

        # if `inputs_embeds` are passed, we only want to use them in the 1st generation step
        if inputs_embeds is not None and len(cache_position) == inputs_embeds.shape[1]:
            model_inputs = {"inputs_embeds": inputs_embeds, "input_ids": None}
        else:
            model_inputs = {"input_ids": input_ids, "inputs_embeds": None}

        model_inputs.update(
            {
                "position_ids": position_ids,
                "past_key_values": past_key_values,
                "use_cache": use_cache,
                "attention_mask": attention_mask,
                "pixel_values": pixel_values,
                "pixel_values_videos": pixel_values_videos,
                "image_grid_thw": image_grid_thw,
                "video_grid_thw": video_grid_thw,
                "cache_position": cache_position,
                "mm_token_type_ids": mm_token_type_ids,
            }
        )
        return model_inputs

    # Copied from https://github.com/huggingface/transformers/blob/v4.45.2/src/transformers/models/qwen2_vl/modeling_qwen2_vl.py#L1602
    def _update_model_kwargs_for_generation(
        self,
        outputs: ModelOutput,
        model_kwargs: Dict[str, Any],
        is_encoder_decoder: bool = False,
        num_new_tokens: int = 1,
    ) -> Dict[str, Any]:
        model_kwargs = super()._update_model_kwargs_for_generation(
            outputs=outputs,
            model_kwargs=model_kwargs,
            is_encoder_decoder=is_encoder_decoder,
            num_new_tokens=num_new_tokens,
        )

        if getattr(outputs, "rope_deltas", None) is not None:
            model_kwargs["rope_deltas"] = outputs.rope_deltas

        return model_kwargs

    def get_vision_embeddings(self, pixel_values, grid_thw, **kwargs):
        hidden_states = self.vision_embeddings(pixel_values)[0]
        rotary_pos_emb = self.rot_pos_emb(grid_thw)
        cu_seqlens = torch.repeat_interleave(grid_thw[:, 1] * grid_thw[:, 2], grid_thw[:, 0]).cumsum(
            dim=0, dtype=torch.int32
        )
        cu_seqlens = torch.nn.functional.pad(cu_seqlens, (1, 0), value=0)
        attention_mask = torch.zeros((1, hidden_states.shape[0], hidden_states.shape[0]), dtype=torch.bool)
        causal_mask = torch.zeros_like(attention_mask, dtype=torch.float32)
        for i in range(1, len(cu_seqlens)):
            attention_mask[..., cu_seqlens[i - 1] : cu_seqlens[i], cu_seqlens[i - 1] : cu_seqlens[i]] = True

        causal_mask.masked_fill_(torch.logical_not(attention_mask), float("-inf"))

        res = self.vision_embeddings_merger(
            pixel_values=hidden_states, attention_mask=causal_mask, rotary_pos_emb=rotary_pos_emb
        )[0]
        return res

    # Adopted from https://github.com/huggingface/transformers/blob/v4.45.2/src/transformers/models/qwen2_vl/modeling_qwen2_vl.py#L1089
    # Use config values instead of model attributes, replace self.rotary_pos_emb -> self._rotary_pos_emb
    def rot_pos_emb(self, grid_thw):
        pos_ids = []
        for t, h, w in grid_thw:
            hpos_ids = torch.arange(h).unsqueeze(1).expand(-1, w)
            hpos_ids = hpos_ids.reshape(
                h // self.config.vision_config.spatial_merge_size,
                self.config.vision_config.spatial_merge_size,
                w // self.config.vision_config.spatial_merge_size,
                self.config.vision_config.spatial_merge_size,
            )
            hpos_ids = hpos_ids.permute(0, 2, 1, 3)
            hpos_ids = hpos_ids.flatten()

            wpos_ids = torch.arange(w).unsqueeze(0).expand(h, -1)
            wpos_ids = wpos_ids.reshape(
                h // self.config.vision_config.spatial_merge_size,
                self.config.vision_config.spatial_merge_size,
                w // self.config.vision_config.spatial_merge_size,
                self.config.vision_config.spatial_merge_size,
            )
            wpos_ids = wpos_ids.permute(0, 2, 1, 3)
            wpos_ids = wpos_ids.flatten()
            pos_ids.append(torch.stack([hpos_ids, wpos_ids], dim=-1).repeat(t, 1))
        pos_ids = torch.cat(pos_ids, dim=0)
        max_grid_size = grid_thw[:, 1:].max()
        rotary_pos_emb_full = self._rotary_pos_emb(max_grid_size)
        rotary_pos_emb = rotary_pos_emb_full[pos_ids].flatten(1)
        return rotary_pos_emb

    def get_multimodal_embeddings(
        self,
        input_ids,
        pixel_values=None,
        attention_mask=None,
        position_ids=None,
        pixel_values_videos=None,
        image_grid_thw=None,
        video_grid_thw=None,
        cache_position=None,
        **kwargs,
    ):
        inputs_embeds = torch.from_numpy(self.get_text_embeddings(input_ids))
        if pixel_values is not None and input_ids.shape[1] != 1:
            image_embeds = torch.from_numpy(self.get_vision_embeddings(pixel_values, image_grid_thw))
            image_mask = input_ids == self.config.image_token_id
            inputs_embeds[image_mask] = image_embeds
        if pixel_values_videos is not None and input_ids.shape[1] != 1:
            pixel_values_videos = pixel_values_videos
            video_embeds = torch.from_numpy(self.get_vision_embeddings(pixel_values_videos, video_grid_thw))
            video_mask = input_ids == self.config.video_token_id
            inputs_embeds[video_mask] = video_embeds

        # if we get 4D attention mask we cannot calculate rope deltas anymore.
        if (
            (position_ids is None or position_ids.ndim < 3)
            and input_ids is not None
            and (attention_mask is None or attention_mask.ndim == 2)
        ):
            # calculate RoPE index once per generation in the pre-fill stage only
            if (cache_position is not None and cache_position[0] == 0) or self.rope_deltas is None:
                if is_transformers_version(">=", "5.3.0"):
                    # since transformers v5.3, get_rope_index requires mm_token_type_ids
                    mm_token_type_ids = kwargs.get("mm_token_type_ids")
                    if mm_token_type_ids is None:
                        mm_token_type_ids = torch.zeros_like(input_ids, dtype=torch.int32)
                        mm_token_type_ids[input_ids == self.config.image_token_id] = 1
                        mm_token_type_ids[input_ids == self.config.video_token_id] = 2
                    position_ids, rope_deltas = self.get_rope_index(
                        input_ids,
                        mm_token_type_ids,
                        image_grid_thw,
                        video_grid_thw,
                        attention_mask,
                    )
                else:
                    position_ids, rope_deltas = self.get_rope_index(
                        input_ids,
                        image_grid_thw,
                        video_grid_thw,
                        attention_mask,
                    )

                self.rope_deltas = rope_deltas
            # then use the prev pre-calculated rope-deltas to get the correct position ids
            else:
                batch_size, seq_length, _ = inputs_embeds.shape
                delta = cache_position[0] + self.rope_deltas if cache_position is not None else 0
                position_ids = torch.arange(seq_length, device=inputs_embeds.device)
                position_ids = position_ids.view(1, -1).expand(batch_size, -1)
                if cache_position is not None:  # otherwise `deltas` is an int `0`
                    delta = delta.repeat_interleave(batch_size // delta.shape[0], dim=0)
                position_ids = position_ids.add(delta)
                position_ids = position_ids.unsqueeze(0).expand(3, -1, -1)

        return inputs_embeds, attention_mask, position_ids

    def forward(
        self,
        input_ids,
        pixel_values=None,
        past_key_values=None,
        inputs_embeds=None,
        image_sizes=None,
        attention_mask=None,
        position_ids=None,
        image_bound=None,
        tgt_sizes=None,
        pixel_values_videos=None,
        image_grid_thw=None,
        video_grid_thw=None,
        rope_deltas=None,
        **kwargs,
    ):
        result = super().forward(
            input_ids,
            pixel_values,
            past_key_values,
            inputs_embeds,
            image_sizes,
            attention_mask,
            position_ids,
            image_bound,
            tgt_sizes,
            pixel_values_videos,
            image_grid_thw,
            video_grid_thw,
            rope_deltas,
            **kwargs,
        )
        final_result = QWen2VLModelOutputWithPast(
            logits=result.logits, past_key_values=result.past_key_values, rope_deltas=rope_deltas
        )
        return final_result

    @staticmethod
    def preprocess_inputs(
        text: str,
        image: Optional["Image"] = None,
        processor: Optional[AutoImageProcessor] = None,
        tokenizer: Optional[PreTrainedTokenizer] = None,
        config: Optional[PretrainedConfig] = None,
        video: Optional["VideoInput"] = None,
        audio: Optional[np.ndarray] = None,
    ):
        if processor is None:
            raise ValueError("Processor is required.")
        if audio is not None:
            raise ValueError("Audio input is not supported")
        conversation = [
            {
                "role": "user",
                "content": [
                    {"type": "text", "text": text},
                ],
            }
        ]
        if image is not None:
            conversation[0]["content"].insert(0, {"type": "image"})
        if video is not None:
            conversation[0]["content"].insert(0, {"type": "video"})

        text_prompt = processor.apply_chat_template(conversation, add_generation_prompt=True)

        inputs = processor(images=image, text=text_prompt, videos=video, return_tensors="pt")
        return inputs


class _OVQwen2_5_VLForCausalLM(OVModelForVisualCausalLM):
    get_rope_index = Qwen2_5_VLModel.get_rope_index
    get_vision_position_ids = getattr(Qwen2_5_VLModel, "get_vision_position_ids", None)

    additional_parts = ["vision_embeddings_merger"]

    def __init__(
        self,
        language_model: ov.Model,
        text_embeddings: ov.Model,
        vision_embeddings: ov.Model,
        config: PretrainedConfig = None,
        device: str = "CPU",
        dynamic_shapes: bool = None,
        ov_config: Optional[Dict[str, str]] = None,
        model_save_dir: Optional[Union[str, Path, TemporaryDirectory]] = None,
        quantization_config: Union[OVWeightQuantizationConfig, Dict] = None,
        **kwargs,
    ):
        super().__init__(
            language_model=language_model,
            text_embeddings=text_embeddings,
            vision_embeddings=vision_embeddings,
            config=config,
            device=device,
            dynamic_shapes=dynamic_shapes,
            ov_config=ov_config,
            model_save_dir=model_save_dir,
            quantization_config=quantization_config,
            **kwargs,
        )
        self.rope_deltas = None  # cache rope_deltas here

        class Qwen2_5_VisionRotaryEmbedding(torch.nn.Module):
            def __init__(self, dim: int, theta: float = 10000.0) -> None:
                super().__init__()
                inv_freq = 1.0 / (theta ** (torch.arange(0, dim, 2, dtype=torch.float) / dim))
                self.register_buffer("inv_freq", inv_freq, persistent=False)

            def forward(self, seqlen: int) -> torch.Tensor:
                seq = torch.arange(seqlen, device=self.inv_freq.device, dtype=self.inv_freq.dtype)
                freqs = torch.outer(seq, self.inv_freq)
                return freqs

        head_dim = config.vision_config.hidden_size // config.vision_config.num_heads
        self._rotary_pos_emb = Qwen2_5_VisionRotaryEmbedding(head_dim // 2)

    def prepare_inputs_for_generation(
        self,
        input_ids,
        past_key_values=None,
        attention_mask=None,
        inputs_embeds=None,
        cache_position=None,
        position_ids=None,
        use_cache=True,
        pixel_values=None,
        pixel_values_videos=None,
        image_grid_thw=None,
        video_grid_thw=None,
        second_per_grid_ts=None,
        mm_token_type_ids=None,
        is_first_iteration=False,
        next_sequence_length=None,
        **kwargs,
    ):
        # reconstruct cache_position as partially removed in v5.3 and totally removed in v5.5
        if is_transformers_version(">=", "5.3") and (
            cache_position is None or (not is_first_iteration and cache_position[0] == 0)
        ):
            if next_sequence_length is not None:
                past_len = input_ids.shape[1] - next_sequence_length
                cache_position = torch.arange(past_len, past_len + next_sequence_length, device=input_ids.device)
            elif not is_first_iteration and attention_mask is not None:
                # v5.3 decode step: input_ids is already sliced to 1 token, use attention_mask length
                past_len = attention_mask.shape[1] - 1
                cache_position = torch.tensor([past_len], device=input_ids.device)
            else:
                cache_position = torch.arange(input_ids.shape[1], device=input_ids.device)

        if past_key_values is not None:
            if inputs_embeds is not None and input_ids.shape[1] == 0:
                inputs_embeds = inputs_embeds[:, -cache_position.shape[0] :]
            elif inputs_embeds is not None:
                input_ids = input_ids[:, -cache_position.shape[0] :]
            elif input_ids.shape[1] != cache_position.shape[0]:  # Default case (the "else", a no op, is Exception 2)
                input_ids = input_ids[:, cache_position]

        if cache_position[0] != 0:
            pixel_values = None
            pixel_values_videos = None

        # if `inputs_embeds` are passed, we only want to use them in the 1st generation step
        if inputs_embeds is not None and len(cache_position) == inputs_embeds.shape[1]:
            model_inputs = {"inputs_embeds": inputs_embeds, "input_ids": None}
        else:
            model_inputs = {"input_ids": input_ids, "inputs_embeds": None}

        model_inputs.update(
            {
                "position_ids": position_ids,
                "past_key_values": past_key_values,
                "use_cache": use_cache,
                "attention_mask": attention_mask,
                "pixel_values": pixel_values,
                "pixel_values_videos": pixel_values_videos,
                "image_grid_thw": image_grid_thw,
                "video_grid_thw": video_grid_thw,
                "cache_position": cache_position,
                "second_per_grid_ts": second_per_grid_ts,
                "mm_token_type_ids": mm_token_type_ids,
            }
        )
        return model_inputs

    def rot_pos_emb(self, grid_thw):
        pos_ids = []
        for t, h, w in grid_thw:
            hpos_ids = torch.arange(h).unsqueeze(1).expand(-1, w)
            hpos_ids = hpos_ids.reshape(
                h // self.config.vision_config.spatial_merge_size,
                self.config.vision_config.spatial_merge_size,
                w // self.config.vision_config.spatial_merge_size,
                self.config.vision_config.spatial_merge_size,
            )
            hpos_ids = hpos_ids.permute(0, 2, 1, 3)
            hpos_ids = hpos_ids.flatten()

            wpos_ids = torch.arange(w).unsqueeze(0).expand(h, -1)
            wpos_ids = wpos_ids.reshape(
                h // self.config.vision_config.spatial_merge_size,
                self.config.vision_config.spatial_merge_size,
                w // self.config.vision_config.spatial_merge_size,
                self.config.vision_config.spatial_merge_size,
            )
            wpos_ids = wpos_ids.permute(0, 2, 1, 3)
            wpos_ids = wpos_ids.flatten()
            pos_ids.append(torch.stack([hpos_ids, wpos_ids], dim=-1).repeat(t, 1))
        pos_ids = torch.cat(pos_ids, dim=0)
        max_grid_size = grid_thw[:, 1:].max()
        rotary_pos_emb_full = self._rotary_pos_emb(max_grid_size)
        rotary_pos_emb = rotary_pos_emb_full[pos_ids].flatten(1)
        return rotary_pos_emb

    def get_multimodal_embeddings(
        self,
        input_ids,
        pixel_values=None,
        attention_mask=None,
        position_ids=None,
        pixel_values_videos=None,
        image_grid_thw=None,
        video_grid_thw=None,
        cache_position=None,
        second_per_grid_ts: Optional[torch.Tensor] = None,
        **kwargs,
    ):
        # Adopted from https://github.com/huggingface/transformers/blob/v4.49.0/src/transformers/models/qwen2_5_vl/modeling_qwen2_5_vl.py#L1791-L1861
        inputs_embeds = torch.from_numpy(self.get_text_embeddings(input_ids))
        if pixel_values is not None and input_ids.shape[1] != 1:
            image_embeds = torch.from_numpy(self.get_vision_embeddings(pixel_values, image_grid_thw))
            n_image_tokens = (input_ids == self.config.image_token_id).sum().item()
            n_image_features = image_embeds.shape[0]
            if n_image_tokens != n_image_features:
                raise ValueError(
                    f"Image features and image tokens do not match: tokens: {n_image_tokens}, features {n_image_features}"
                )

            mask = input_ids == self.config.image_token_id
            mask_unsqueezed = mask.unsqueeze(-1)
            mask_expanded = mask_unsqueezed.expand_as(inputs_embeds)
            image_mask = mask_expanded.to(inputs_embeds.device)

            image_embeds = image_embeds.to(inputs_embeds.device, inputs_embeds.dtype)
            inputs_embeds = inputs_embeds.masked_scatter(image_mask, image_embeds)
        if pixel_values_videos is not None and input_ids.shape[1] != 1:
            video_embeds = torch.from_numpy(self.get_vision_embeddings(pixel_values_videos, video_grid_thw))
            n_video_tokens = (input_ids == self.config.video_token_id).sum().item()
            n_video_features = video_embeds.shape[0]
            if n_video_tokens != n_video_features:
                raise ValueError(
                    f"Video features and video tokens do not match: tokens: {n_video_tokens}, features {n_video_features}"
                )

            mask = input_ids == self.config.video_token_id
            mask_unsqueezed = mask.unsqueeze(-1)
            mask_expanded = mask_unsqueezed.expand_as(inputs_embeds)
            video_mask = mask_expanded.to(inputs_embeds.device)

            video_embeds = video_embeds.to(inputs_embeds.device, inputs_embeds.dtype)
            inputs_embeds = inputs_embeds.masked_scatter(video_mask, video_embeds)

        # if we get 4D attention mask we cannot calculate rope deltas anymore.
        if (
            (position_ids is None or position_ids.ndim < 3)
            and input_ids is not None
            and (attention_mask is None or attention_mask.ndim == 2)
        ):
            # calculate RoPE index once per generation in the pre-fill stage only
            if (cache_position is not None and cache_position[0] == 0) or self.rope_deltas is None:
                if is_transformers_version(">=", "5.3.0"):
                    # since transformers v5.3, get_rope_index requires mm_token_type_ids
                    mm_token_type_ids = kwargs.get("mm_token_type_ids")
                    if mm_token_type_ids is None:
                        mm_token_type_ids = torch.zeros_like(input_ids, dtype=torch.int32)
                        mm_token_type_ids[input_ids == self.config.image_token_id] = 1
                        mm_token_type_ids[input_ids == self.config.video_token_id] = 2
                    position_ids, rope_deltas = self.get_rope_index(
                        input_ids,
                        mm_token_type_ids,
                        image_grid_thw,
                        video_grid_thw,
                        second_per_grid_ts,
                        attention_mask,
                    )
                else:
                    position_ids, rope_deltas = self.get_rope_index(
                        input_ids,
                        image_grid_thw,
                        video_grid_thw,
                        second_per_grid_ts,
                        attention_mask,
                    )
                self.rope_deltas = rope_deltas
            # then use the prev pre-calculated rope-deltas to get the correct position ids
            else:
                batch_size, seq_length, _ = inputs_embeds.shape
                delta = cache_position[0] + self.rope_deltas if cache_position is not None else 0
                position_ids = torch.arange(seq_length, device=inputs_embeds.device)
                position_ids = position_ids.view(1, -1).expand(batch_size, -1)
                if cache_position is not None:  # otherwise `deltas` is an int `0`
                    delta = delta.repeat_interleave(batch_size // delta.shape[0], dim=0)
                position_ids = position_ids.add(delta)
                position_ids = position_ids.unsqueeze(0).expand(3, -1, -1)

        return inputs_embeds, attention_mask, position_ids

    def get_vision_embeddings(self, pixel_values, grid_thw, **kwargs):
        hidden_states = self.vision_embeddings(pixel_values)[0]
        rotary_pos_emb = self.rot_pos_emb(grid_thw)
        window_index, cu_window_seqlens = self.get_window_index(grid_thw)
        cu_window_seqlens = torch.tensor(
            cu_window_seqlens,
            dtype=torch.int32,
        )
        cu_window_seqlens = torch.unique_consecutive(cu_window_seqlens)
        cu_seqlens = torch.repeat_interleave(grid_thw[:, 1] * grid_thw[:, 2], grid_thw[:, 0]).cumsum(
            dim=0, dtype=torch.int32
        )
        cu_seqlens = torch.nn.functional.pad(cu_seqlens, (1, 0), value=0)
        attention_mask = torch.zeros((1, hidden_states.shape[0], hidden_states.shape[0]), dtype=torch.bool)
        causal_mask = torch.zeros_like(attention_mask, dtype=torch.float32)
        for i in range(1, len(cu_seqlens)):
            attention_mask[..., cu_seqlens[i - 1] : cu_seqlens[i], cu_seqlens[i - 1] : cu_seqlens[i]] = True

        causal_mask.masked_fill_(torch.logical_not(attention_mask), float("-inf"))

        window_attention_mask = torch.zeros((1, hidden_states.shape[0], hidden_states.shape[0]), dtype=torch.bool)
        window_causal_mask = torch.zeros_like(attention_mask, dtype=torch.float32)
        for i in range(1, len(cu_window_seqlens)):
            window_attention_mask[
                ..., cu_window_seqlens[i - 1] : cu_window_seqlens[i], cu_window_seqlens[i - 1] : cu_window_seqlens[i]
            ] = True

        window_causal_mask.masked_fill_(torch.logical_not(window_attention_mask), float("-inf"))

        res = self.vision_embeddings_merger(
            pixel_values=hidden_states,
            attention_mask=causal_mask,
            window_attention_mask=window_causal_mask,
            window_index=window_index,
            rotary_pos_emb=rotary_pos_emb,
        )[0]
        return res

    def get_window_index(self, grid_thw):
        window_index: list = []
        cu_window_seqlens: list = [0]
        window_index_id = 0
        vit_merger_window_size = (
            self.config.vision_config.window_size
            // self.config.vision_config.spatial_merge_size
            // self.config.vision_config.patch_size
        )
        spatial_merge_unit = (
            self.config.vision_config.spatial_merge_size * self.config.vision_config.spatial_merge_size
        )

        for grid_t, grid_h, grid_w in grid_thw:
            llm_grid_h, llm_grid_w = (
                grid_h // self.config.vision_config.spatial_merge_size,
                grid_w // self.config.vision_config.spatial_merge_size,
            )
            index = torch.arange(grid_t * llm_grid_h * llm_grid_w).reshape(grid_t, llm_grid_h, llm_grid_w)
            pad_h = vit_merger_window_size - llm_grid_h % vit_merger_window_size
            pad_w = vit_merger_window_size - llm_grid_w % vit_merger_window_size
            num_windows_h = (llm_grid_h + pad_h) // vit_merger_window_size
            num_windows_w = (llm_grid_w + pad_w) // vit_merger_window_size
            index_padded = torch.nn.functional.pad(index, (0, pad_w, 0, pad_h), "constant", -100)
            index_padded = index_padded.reshape(
                grid_t,
                num_windows_h,
                vit_merger_window_size,
                num_windows_w,
                vit_merger_window_size,
            )
            index_padded = index_padded.permute(0, 1, 3, 2, 4).reshape(
                grid_t,
                num_windows_h * num_windows_w,
                vit_merger_window_size,
                vit_merger_window_size,
            )
            seqlens = (index_padded != -100).sum([2, 3]).reshape(-1)
            index_padded = index_padded.reshape(-1)
            index_new = index_padded[index_padded != -100]
            window_index.append(index_new + window_index_id)
            cu_seqlens_tmp = seqlens.cumsum(0) * spatial_merge_unit + cu_window_seqlens[-1]
            cu_window_seqlens.extend(cu_seqlens_tmp.tolist())
            window_index_id += (grid_t * llm_grid_h * llm_grid_w).item()
        window_index = torch.cat(window_index, dim=0)

        return window_index, cu_window_seqlens

    @staticmethod
    def preprocess_inputs(
        text: str,
        image: Optional["Image"] = None,
        processor: Optional[AutoImageProcessor] = None,
        tokenizer: Optional[PreTrainedTokenizer] = None,
        config: Optional[PretrainedConfig] = None,
        video: Optional["VideoInput"] = None,
        audio: Optional[np.ndarray] = None,
    ):
        if processor is None:
            raise ValueError("Processor is required.")
        if audio is not None:
            raise ValueError("Audio input is not supported")
        conversation = [
            {
                "role": "user",
                "content": [
                    {"type": "text", "text": text},
                ],
            }
        ]
        if image is not None:
            conversation[0]["content"].insert(0, {"type": "image"})
        if video is not None:
            conversation[0]["content"].insert(0, {"type": "video"})

        text_prompt = processor.apply_chat_template(conversation, add_generation_prompt=True)

        inputs = processor(images=image, text=text_prompt, videos=video, return_tensors="pt")

        # since transformers v5.3, get_rope_index consumes one second_per_grid_ts entry per
        # non-text token group (images and videos), but the processor only populates it for videos
        if is_transformers_version(">=", "5.3.0") and "second_per_grid_ts" in inputs and "mm_token_type_ids" in inputs:
            non_text_groups = inputs["mm_token_type_ids"][0].unique_consecutive()
            non_text_groups = non_text_groups[non_text_groups != 0].tolist()
            video_ts_iter = iter(inputs["second_per_grid_ts"].tolist())
            full_ts = [next(video_ts_iter, 1.0) if g == 2 else 1.0 for g in non_text_groups]
            inputs["second_per_grid_ts"] = torch.tensor(full_ts, dtype=inputs["second_per_grid_ts"].dtype)

        return inputs


# The inheritance from Qwen3VLModel is needed to get access to methods:
# get_placeholder_mask(): https://github.com/huggingface/transformers/blob/v4.57.6/src/transformers/models/qwen3_vl/modeling_qwen3_vl.py#L1066
# get_rope_index(): https://github.com/huggingface/transformers/blob/v4.57.6/src/transformers/models/qwen3_vl/modeling_qwen3_vl.py#L916
# get_video_features(): https://github.com/huggingface/transformers/blob/v4.57.6/src/transformers/models/qwen3_vl/modeling_qwen3_vl.py#L1035
#
# and inheritance from Qwen3VLVisionModel is needed for accessing the following method:
# rot_pos_emb(): https://github.com/huggingface/transformers/blob/v4.57.6/src/transformers/models/qwen3_vl/modeling_qwen3_vl.py#L603
class _OVQwen3VLForCausalLM(OVModelForVisualCausalLM, Qwen3VLModel, Qwen3VLVisionModel):
    additional_parts = ["vision_embeddings_merger", "vision_embeddings_pos"]

    def __init__(
        self,
        language_model: ov.Model,
        text_embeddings: ov.Model,
        vision_embeddings: ov.Model,
        config: PretrainedConfig = None,
        device: str = "CPU",
        dynamic_shapes: bool = None,
        ov_config: Optional[Dict[str, str]] = None,
        model_save_dir: Optional[Union[str, Path, TemporaryDirectory]] = None,
        quantization_config: Union[OVWeightQuantizationConfig, Dict] = None,
        **kwargs,
    ):
        if is_transformers_version("<", "4.57.0"):
            raise Exception("Qwen3VL is not supported in transformers versions earlier than 4.57.0.")

        super().__init__(
            language_model=language_model,
            text_embeddings=text_embeddings,
            vision_embeddings=vision_embeddings,
            config=config,
            device=device,
            dynamic_shapes=dynamic_shapes,
            ov_config=ov_config,
            model_save_dir=model_save_dir,
            quantization_config=quantization_config,
            **kwargs,
        )
        self.rope_deltas = None  # cache rope_deltas here

        self._rotary_pos_emb = Qwen3VLVisionRotaryEmbedding(
            self.config.vision_config.hidden_size // self.config.vision_config.num_heads // 2
        )
        self.num_grid_per_side = int(config.vision_config.num_position_embeddings**0.5)
        self.spatial_merge_size = config.vision_config.spatial_merge_size
        head_dim = config.vision_config.hidden_size // config.vision_config.num_heads
        self.rotary_pos_emb = Qwen3VLVisionRotaryEmbedding(head_dim // 2)

    def __setattr__(self, name, value):
        OVModelForVisualCausalLM.__setattr__(self, name, value)

    # Adapted from https://github.com/huggingface/transformers/blob/v5.0.0/src/transformers/models/qwen3_vl/modeling_qwen3_vl.py#L1471-1537
    def prepare_inputs_for_generation(
        self,
        input_ids,
        past_key_values=None,
        attention_mask=None,
        inputs_embeds=None,
        cache_position=None,
        position_ids=None,
        use_cache=True,
        pixel_values=None,
        pixel_values_videos=None,
        image_grid_thw=None,
        video_grid_thw=None,
        mm_token_type_ids=None,
        is_first_iteration=False,
        next_sequence_length=None,
        **kwargs,
    ):
        # reconstruct cache_position as partially removed in v5.3 and totally removed in v5.5
        if is_transformers_version(">=", "5.3") and (
            cache_position is None or (not is_first_iteration and cache_position[0] == 0)
        ):
            if next_sequence_length is not None:
                past_len = input_ids.shape[1] - next_sequence_length
                cache_position = torch.arange(past_len, past_len + next_sequence_length, device=input_ids.device)
            elif not is_first_iteration and attention_mask is not None:
                # v5.3 decode step: input_ids is already sliced to 1 token, use attention_mask length
                past_len = attention_mask.shape[1] - 1
                cache_position = torch.tensor([past_len], device=input_ids.device)
            else:
                cache_position = torch.arange(input_ids.shape[1], device=input_ids.device)

        # Overwritten -- in specific circumstances we don't want to forward image inputs to the model
        if past_key_values is not None:
            if inputs_embeds is not None and input_ids.shape[1] == 0:  # Exception 4
                inputs_embeds = inputs_embeds[:, -cache_position.shape[0] :]
            elif inputs_embeds is not None:
                input_ids = input_ids[:, -cache_position.shape[0] :]
            elif input_ids.shape[1] != cache_position.shape[0]:  # Default case (the "else", a no op, is Exception 2)
                input_ids = input_ids[:, cache_position]

        if cache_position[0] != 0:
            pixel_values = None
            pixel_values_videos = None

        # if `inputs_embeds` are passed, we only want to use them in the 1st generation step
        if inputs_embeds is not None and len(cache_position) == inputs_embeds.shape[1]:
            model_inputs = {"inputs_embeds": inputs_embeds, "input_ids": None}
        else:
            model_inputs = {"input_ids": input_ids, "inputs_embeds": None}

        model_inputs.update(
            {
                "position_ids": position_ids,
                "past_key_values": past_key_values,
                "use_cache": use_cache,
                "attention_mask": attention_mask,
                "pixel_values": pixel_values,
                "pixel_values_videos": pixel_values_videos,
                "image_grid_thw": image_grid_thw,
                "video_grid_thw": video_grid_thw,
                "cache_position": cache_position,
                "mm_token_type_ids": mm_token_type_ids,
            }
        )
        return model_inputs

    # Adapted from https://github.com/huggingface/transformers/blob/8ac2b916b042b1f78b75c9eb941c0f5d2cdd8e10/src/transformers/models/qwen3_vl/modeling_qwen3_vl.py#L642
    # This method needs to be changed, as instead of running self.pos_embed of type nn.Embedding, openvino model needs to be inferred (self.vision_embeddings_pos)
    # and self.vision_embeddings_pos does not have some attributes like weights or device, which are used in the original method
    def fast_pos_embed_interpolate(self, grid_thw):
        grid_ts, grid_hs, grid_ws = grid_thw[:, 0], grid_thw[:, 1], grid_thw[:, 2]

        idx_list = [[] for _ in range(4)]
        weight_list = [[] for _ in range(4)]

        for t, h, w in zip(grid_ts, grid_hs, grid_ws):
            h_idxs = torch.linspace(0, self.num_grid_per_side - 1, h)
            w_idxs = torch.linspace(0, self.num_grid_per_side - 1, w)

            h_idxs_floor = h_idxs.int()
            w_idxs_floor = w_idxs.int()
            h_idxs_ceil = (h_idxs.int() + 1).clip(max=self.num_grid_per_side - 1)
            w_idxs_ceil = (w_idxs.int() + 1).clip(max=self.num_grid_per_side - 1)

            dh = h_idxs - h_idxs_floor
            dw = w_idxs - w_idxs_floor

            base_h = h_idxs_floor * self.num_grid_per_side
            base_h_ceil = h_idxs_ceil * self.num_grid_per_side

            indices = [
                (base_h[None].T + w_idxs_floor[None]).flatten(),
                (base_h[None].T + w_idxs_ceil[None]).flatten(),
                (base_h_ceil[None].T + w_idxs_floor[None]).flatten(),
                (base_h_ceil[None].T + w_idxs_ceil[None]).flatten(),
            ]

            weights = [
                ((1 - dh)[None].T * (1 - dw)[None]).flatten(),
                ((1 - dh)[None].T * dw[None]).flatten(),
                (dh[None].T * (1 - dw)[None]).flatten(),
                (dh[None].T * dw[None]).flatten(),
            ]

            for i in range(4):
                idx_list[i].extend(indices[i].tolist())
                weight_list[i].extend(weights[i].tolist())

        idx_tensor = torch.tensor(idx_list)
        weight_tensor = torch.tensor(weight_list)
        pos_embeds = torch.from_numpy(self.vision_embeddings_pos(idx_tensor)) * weight_tensor[:, :, None]
        patch_pos_embeds = pos_embeds[0] + pos_embeds[1] + pos_embeds[2] + pos_embeds[3]

        patch_pos_embeds = patch_pos_embeds.split([h * w for h, w in zip(grid_hs, grid_ws)])

        patch_pos_embeds_permute = []
        merge_size = self.config.vision_config.spatial_merge_size
        for pos_embed, t, h, w in zip(patch_pos_embeds, grid_ts, grid_hs, grid_ws):
            pos_embed = pos_embed.repeat(t, 1)
            pos_embed = (
                pos_embed.view(t, h // merge_size, merge_size, w // merge_size, merge_size, -1)
                .permute(0, 1, 3, 2, 4, 5)
                .flatten(0, 4)
            )
            patch_pos_embeds_permute.append(pos_embed)
        patch_pos_embeds = torch.cat(patch_pos_embeds_permute)
        return patch_pos_embeds

    # Adapted from https://github.com/huggingface/transformers/blob/v5.0.0/src/transformers/models/qwen3_vl/modeling_qwen3_vl.py#L758-814
    def get_vision_embeddings(self, pixel_values, grid_thw, **kwargs):
        hidden_states = torch.from_numpy(self.vision_embeddings(pixel_values)[0])

        pos_embeds = self.fast_pos_embed_interpolate(grid_thw)
        hidden_states = hidden_states + pos_embeds

        rotary_pos_emb = self.rot_pos_emb(grid_thw)
        seq_len, _ = hidden_states.size()
        hidden_states = hidden_states.reshape(seq_len, -1)
        rotary_pos_emb = rotary_pos_emb.reshape(seq_len, -1)
        cu_seqlens = torch.repeat_interleave(grid_thw[:, 1] * grid_thw[:, 2], grid_thw[:, 0]).cumsum(
            dim=0, dtype=torch.int32
        )
        cu_seqlens = torch.nn.functional.pad(cu_seqlens, (1, 0), value=0)
        attention_mask = torch.zeros((1, hidden_states.shape[0], hidden_states.shape[0]), dtype=torch.bool)
        causal_mask = torch.zeros_like(attention_mask, dtype=torch.float32)
        for i in range(1, len(cu_seqlens)):
            attention_mask[..., cu_seqlens[i - 1] : cu_seqlens[i], cu_seqlens[i - 1] : cu_seqlens[i]] = True

        causal_mask.masked_fill_(torch.logical_not(attention_mask), float("-inf"))

        res = self.vision_embeddings_merger(
            pixel_values=hidden_states, attention_mask=causal_mask, rotary_pos_emb=rotary_pos_emb
        )
        return res[0], res[1]

    # Adapted from https://github.com/huggingface/transformers/blob/08810b1e278938278c50153ee1edfd7a20a759da/src/transformers/models/qwen3_vl/modeling_qwen3_vl.py#L1110-1131
    def get_image_features(self, pixel_values: torch.FloatTensor, image_grid_thw: Optional[torch.LongTensor] = None):
        """
        Encodes images into continuous embeddings that can be forwarded to the language model. The deepstack visual features are also returned.

        Args:
            pixel_values (`torch.FloatTensor` of shape `(batch_size, num_channels, image_size, image_size)`):
                The tensors corresponding to the input images.
            image_grid_thw (`torch.LongTensor` of shape `(num_images, 3)`, *optional*):
                The temporal, height and width of feature shape of each image in LLM.
        """
        image_embeds, deepstack_image_embeds = self.get_vision_embeddings(pixel_values, image_grid_thw)
        image_embeds, deepstack_image_embeds = (
            torch.from_numpy(image_embeds),
            torch.from_numpy(deepstack_image_embeds),
        )
        deepstack_image_embeds = deepstack_image_embeds.tolist()
        split_sizes = (image_grid_thw.prod(-1) // self.spatial_merge_size**2).tolist()
        image_embeds = torch.split(image_embeds, split_sizes)
        return image_embeds, deepstack_image_embeds

    # Adapted from https://github.com/huggingface/transformers/blob/v5.0.0/src/transformers/models/qwen3_vl/modeling_qwen3_vl.py#L1176-1290
    def get_multimodal_embeddings(
        self,
        input_ids,
        pixel_values=None,
        attention_mask=None,
        position_ids=None,
        pixel_values_videos=None,
        image_grid_thw=None,
        video_grid_thw=None,
        cache_position=None,
        **kwargs,
    ):
        image_mask = None
        video_mask = None
        inputs_embeds = torch.from_numpy(self.get_text_embeddings(input_ids))

        if pixel_values is not None:
            image_embeds, deepstack_image_embeds = self.get_image_features(pixel_values, image_grid_thw)
            image_embeds = torch.cat(image_embeds, dim=0)
            image_mask, _ = self.get_placeholder_mask(
                input_ids, inputs_embeds=inputs_embeds, image_features=image_embeds
            )
            inputs_embeds = inputs_embeds.masked_scatter(image_mask, image_embeds)

        if pixel_values_videos is not None:
            video_embeds, deepstack_video_embeds = self.get_video_features(pixel_values_videos, video_grid_thw)
            video_embeds = torch.cat(video_embeds, dim=0)
            _, video_mask = self.get_placeholder_mask(
                input_ids, inputs_embeds=inputs_embeds, video_features=video_embeds
            )
            inputs_embeds = inputs_embeds.masked_scatter(video_mask, video_embeds)

        visual_pos_masks = None
        deepstack_visual_embeds = None
        if image_mask is not None and video_mask is not None:
            # aggregate visual_pos_masks and deepstack_visual_embeds
            image_mask = image_mask[..., 0]
            video_mask = video_mask[..., 0]
            visual_pos_masks = image_mask | video_mask
            deepstack_visual_embeds = []
            image_mask_joint = image_mask[visual_pos_masks]
            video_mask_joint = video_mask[visual_pos_masks]
            for img_embed, vid_embed in zip(deepstack_image_embeds, deepstack_video_embeds):
                embed_joint = np.zeros((visual_pos_masks.sum(), len(img_embed[0])))
                embed_joint[image_mask_joint, :] = img_embed
                embed_joint[video_mask_joint, :] = vid_embed
                deepstack_visual_embeds.append(embed_joint)
        elif image_mask is not None:
            image_mask = image_mask[..., 0]
            visual_pos_masks = image_mask
            deepstack_visual_embeds = deepstack_image_embeds
        elif video_mask is not None:
            video_mask = video_mask[..., 0]
            visual_pos_masks = video_mask
            deepstack_visual_embeds = deepstack_video_embeds

        if position_ids is None or position_ids.ndim < 3:
            attention_mask_tensor = (
                attention_mask if not isinstance(attention_mask, dict) else attention_mask["full_attention"]
            )
            if attention_mask_tensor is not None and attention_mask_tensor.ndim == 4:
                attention_mask_tensor = torch.diagonal(attention_mask_tensor[:, 0], dim1=1, dim2=2)
                # Only apply conversion for floating point tensors (inverted masks)
                if attention_mask_tensor.dtype.is_floating_point:
                    attention_mask_tensor = attention_mask_tensor / torch.finfo(attention_mask_tensor.dtype).min
                    attention_mask_tensor = (1.0 - attention_mask_tensor).int()

            # Calculate RoPE index once per generation in the pre-fill stage only.
            # When compiling, we can't check tensor values thus we check only input length
            # It is safe to assume that `length!=1` means we're in pre-fill because compiled
            # models currently cannot do asssisted decoding
            if self.rope_deltas is None:
                if is_transformers_version(">=", "5.3.0"):
                    # since transformers v5.3, get_rope_index requires mm_token_type_ids
                    mm_token_type_ids = kwargs.get("mm_token_type_ids")
                    if mm_token_type_ids is None:
                        mm_token_type_ids = torch.zeros_like(input_ids, dtype=torch.int32)
                        mm_token_type_ids[input_ids == self.config.image_token_id] = 1
                        mm_token_type_ids[input_ids == self.config.video_token_id] = 2
                    position_ids, rope_deltas = self.get_rope_index(
                        input_ids,
                        mm_token_type_ids,
                        image_grid_thw,
                        video_grid_thw,
                        attention_mask=attention_mask_tensor,
                    )
                else:
                    position_ids, rope_deltas = self.get_rope_index(
                        input_ids,
                        image_grid_thw,
                        video_grid_thw,
                        attention_mask=attention_mask_tensor,
                    )
                self.rope_deltas = rope_deltas
            # then use the prev pre-calculated rope-deltas to get the correct position ids
            else:
                batch_size, seq_length, _ = inputs_embeds.shape
                delta = (
                    (cache_position[0] + self.rope_deltas).to(inputs_embeds.device)
                    if cache_position is not None
                    else 0
                )
                position_ids = torch.arange(seq_length, device=inputs_embeds.device)
                position_ids = position_ids.view(1, -1).expand(batch_size, -1)
                if cache_position is not None:  # otherwise `deltas` is an int `0`
                    delta = delta.repeat_interleave(batch_size // delta.shape[0], dim=0)
                position_ids = position_ids.add(delta)
                position_ids = position_ids.unsqueeze(0).expand(3, -1, -1)
        return inputs_embeds, attention_mask, position_ids, visual_pos_masks, deepstack_visual_embeds

    @staticmethod
    def preprocess_inputs(
        text: str,
        image: Optional["Image"] = None,
        processor: Optional[AutoImageProcessor] = None,
        tokenizer: Optional[PreTrainedTokenizer] = None,
        config: Optional[PretrainedConfig] = None,
        video: Optional["VideoInput"] = None,
        audio: Optional[np.ndarray] = None,
    ):
        if processor is None:
            raise ValueError("Processor is required.")
        if audio is not None:
            raise ValueError("Audio input is not supported")
        conversation = [
            {
                "role": "user",
                "content": [
                    {"type": "text", "text": text},
                ],
            }
        ]
        if image is not None:
            conversation[0]["content"].insert(0, {"type": "image"})
        if video is not None:
            conversation[0]["content"].insert(0, {"type": "video"})

        text_prompt = processor.apply_chat_template(conversation, add_generation_prompt=True)

        inputs = processor(images=image, text=text_prompt, videos=video, return_tensors="pt")
        return inputs

    def forward(
        self,
        input_ids,
        pixel_values=None,
        past_key_values=None,
        inputs_embeds=None,
        image_sizes=None,
        attention_mask=None,
        position_ids=None,
        image_bound=None,
        tgt_sizes=None,
        pixel_values_videos=None,
        image_grid_thw=None,
        video_grid_thw=None,
        rope_deltas=None,
        **kwargs,
    ):
        result = super().forward(
            input_ids,
            pixel_values,
            past_key_values,
            inputs_embeds,
            image_sizes,
            attention_mask,
            position_ids,
            image_bound,
            tgt_sizes,
            pixel_values_videos,
            image_grid_thw,
            video_grid_thw,
            rope_deltas,
            **kwargs,
        )
        final_result = QWen2VLModelOutputWithPast(
            logits=result.logits, past_key_values=result.past_key_values, rope_deltas=rope_deltas
        )
        return final_result

    def generate(self, *args, **kwargs):
        # Clear cached rope delta from previous generations
        self.rope_deltas = None

        return super().generate(*args, **kwargs)


class _OVMaira2ForCausalLM(_OVLlavaForCausalLM):
    @staticmethod
    def preprocess_inputs(
        text: str,
        image: Optional["Image"] = None,
        processor: Optional[AutoImageProcessor] = None,
        tokenizer: Optional[PreTrainedTokenizer] = None,
        config: Optional[PretrainedConfig] = None,
        video: Optional["VideoInput"] = None,
        audio: Optional[np.ndarray] = None,
    ):
        if processor is None:
            raise ValueError("processor is required")
        if video is not None:
            raise ValueError("Video input is not supported")
        if audio is not None:
            raise ValueError("Audio input is not supported")
        if image is None:
            return processor(text=text, return_tensors="pt")
        processed_inputs = processor.format_and_preprocess_phrase_grounding_input(
            frontal_image=image,
            phrase=text,
            return_tensors="pt",
        )
        return processed_inputs


class _OVGemma3ForCausalLM(OVModelForVisualCausalLM):
    def get_vision_embeddings(self, pixel_values, input_ids=None, **kwargs):
        if input_ids is not None and input_ids.shape[1] == 1:
            return None
        return self.vision_embeddings(pixel_values).last_hidden_state

    def merge_vision_text_embeddings(
        self, vision_embeds, inputs_embeds, input_ids=None, attention_mask=None, position_ids=None, **kwargs
    ):
        # Adopted from https://github.com/huggingface/transformers/blob/v4.49.0-Gemma-3/src/transformers/models/gemma3/modeling_gemma3.py#L1323-L1339
        image_features = torch.from_numpy(vision_embeds) if isinstance(vision_embeds, np.ndarray) else vision_embeds
        inputs_embeds = torch.from_numpy(inputs_embeds) if isinstance(inputs_embeds, np.ndarray) else inputs_embeds
        if input_ids is None:
            special_image_mask = inputs_embeds == torch.from_numpy(
                self.get_text_embeddings(torch.tensor([[self.config.image_token_index]], dtype=torch.long))[0]
            )
        else:
            special_image_mask = (input_ids == self.config.image_token_index).unsqueeze(-1)
            special_image_mask = special_image_mask.expand_as(inputs_embeds)

            image_features = image_features.to(inputs_embeds.dtype)
            inputs_embeds = inputs_embeds.masked_scatter(special_image_mask, image_features)

        return inputs_embeds, attention_mask, position_ids

    @staticmethod
    def preprocess_inputs(
        text: str,
        image: Optional["Image"] = None,
        processor: Optional[AutoImageProcessor] = None,
        tokenizer: Optional[PreTrainedTokenizer] = None,
        config: Optional[PretrainedConfig] = None,
        video: Optional["VideoInput"] = None,
        audio: Optional[np.ndarray] = None,
    ):
        if processor is None:
            raise ValueError("Processor is required.")
        if video is not None:
            raise ValueError("Video input is not supported")
        if audio is not None:
            raise ValueError("Audio input is not supported")
        conversation = [
            {
                "role": "user",
                "content": [
                    {"type": "text", "text": text},
                ],
            }
        ]
        if image is not None:
            conversation[0]["content"].insert(0, {"type": "image"})

        text_prompt = processor.apply_chat_template(conversation, add_generation_prompt=True, tokenize=False)

        # switch off add_bos_token if chat template already includes it
        orig_add_bos_token = processor.tokenizer.add_bos_token
        if "bos_token" in processor.tokenizer.chat_template:
            processor.tokenizer.add_bos_token = False

        inputs = processor(images=image, text=text_prompt, videos=video, return_tensors="pt")

        # recover add_bos_token flag in tokenizer
        processor.tokenizer.add_bos_token = orig_add_bos_token

        return inputs

    def _update_model_kwargs_for_generation(
        self,
        outputs: ModelOutput,
        model_kwargs: Dict[str, Any],
        is_encoder_decoder: bool = False,
        num_new_tokens: int = 1,
    ) -> Dict[str, Any]:
        model_kwargs = super()._update_model_kwargs_for_generation(
            outputs=outputs,
            model_kwargs=model_kwargs,
            is_encoder_decoder=is_encoder_decoder,
            num_new_tokens=num_new_tokens,
        )

        # Token type ids used only for first inference mask generation
        model_kwargs.pop("token_type_ids", None)

        return model_kwargs


class _OVGotOCR2ForCausalLM(OVModelForVisualCausalLM):
    def get_vision_embeddings(self, pixel_values, input_ids, **kwargs):
        if input_ids is not None and input_ids.shape[1] == 1 and kwargs.get("past_key_values") is not None:
            return None
        return self.vision_embeddings(pixel_values).last_hidden_state

    def merge_vision_text_embeddings(
        self, vision_embeds, inputs_embeds, input_ids=None, attention_mask=None, position_ids=None, **kwargs
    ):
        # Adopted from https://github.com/huggingface/transformers/blob/v4.49.0/src/transformers/models/got_ocr2/modeling_got_ocr2.py#L836-L845
        image_features = torch.from_numpy(vision_embeds) if isinstance(vision_embeds, np.ndarray) else vision_embeds
        inputs_embeds = torch.from_numpy(inputs_embeds) if isinstance(inputs_embeds, np.ndarray) else inputs_embeds
        n_image_tokens = (input_ids == self.config.image_token_index).sum()
        n_image_features = image_features.shape[0] * image_features.shape[1]
        if n_image_tokens != n_image_features:
            raise ValueError(
                f"Image features and image tokens do not match: tokens: {n_image_tokens}, features {n_image_features}"
            )
        special_image_mask = (input_ids == self.config.image_token_index).unsqueeze(-1)
        special_image_mask = special_image_mask.expand_as(inputs_embeds).to(inputs_embeds.device)
        image_features = image_features.to(inputs_embeds.device, inputs_embeds.dtype)
        inputs_embeds = inputs_embeds.masked_scatter(special_image_mask, image_features)

        return inputs_embeds, attention_mask, position_ids

    @staticmethod
    def preprocess_inputs(
        text: Optional[str] = None,
        image: Optional["Image"] = None,
        processor: Optional[AutoImageProcessor] = None,
        tokenizer: Optional[PreTrainedTokenizer] = None,
        config: Optional[PretrainedConfig] = None,
        video: Optional["VideoInput"] = None,
        audio: Optional[np.ndarray] = None,
    ):
        if processor is None:
            raise ValueError("processor is required")
        if video is not None:
            raise ValueError("Video input is not supported")
        if audio is not None:
            raise ValueError("Audio input is not supported")
        if image is None:
            raise ValueError("Image is required")
        processed_inputs = processor(image, return_tensors="pt")
        return processed_inputs


class _OVIdefics3ForCausalLM(OVModelForVisualCausalLM):
    def get_vision_embeddings(self, pixel_values, input_ids, **kwargs):
        # Adopted from https://github.com/huggingface/transformers/blob/v4.49.0-SmolVLM-2/src/transformers/models/smolvlm/modeling_smolvlm.py#L899-L942
        if input_ids is not None and input_ids.shape[1] == 1 and kwargs.get("past_key_values") is not None:
            return None
        batch_size, num_images, num_channels, height, width = pixel_values.shape
        pixel_values = pixel_values
        pixel_values = pixel_values.view(batch_size * num_images, *pixel_values.shape[2:])

        # Remove padding images - padding images are full 0.
        nb_values_per_image = pixel_values.shape[1:].numel()
        real_images_inds = (pixel_values == 0.0).sum(dim=(-1, -2, -3)) != nb_values_per_image

        if not any(real_images_inds):
            # no images, leave one empty image.
            real_images_inds[0] = True

        pixel_values = pixel_values[real_images_inds].contiguous()
        pixel_attention_mask = kwargs.get("pixel_attention_mask")

        # Handle the vision attention mask
        if pixel_attention_mask is None:
            pixel_attention_mask = torch.ones(
                size=[pixel_values.shape[i] for i in (0, 2, 3)],
                dtype=torch.bool,
                device=pixel_values.device,
            )
        else:
            # Remove padding images from the mask
            pixel_attention_mask = pixel_attention_mask.view(batch_size * num_images, *pixel_attention_mask.shape[2:])
            pixel_attention_mask = pixel_attention_mask[real_images_inds].contiguous()

        patch_size = self.config.vision_config.patch_size
        num_patches_per_side = self.config.vision_config.image_size // patch_size
        patches_subgrid = pixel_attention_mask.unfold(dimension=1, size=patch_size, step=patch_size)
        patches_subgrid = patches_subgrid.unfold(dimension=2, size=patch_size, step=patch_size)
        patch_attention_mask = (patches_subgrid.sum(dim=(-1, -2)) > 0).bool()
        batch_size_, _, max_im_h, max_im_w = pixel_values.shape
        max_nb_patches_h, max_nb_patches_w = max_im_h // patch_size, max_im_w // patch_size
        boundaries = torch.arange(1 / num_patches_per_side, 1.0, 1 / num_patches_per_side)
        position_ids = torch.full(size=(batch_size_, max_nb_patches_h * max_nb_patches_w), fill_value=0)

        for batch_idx, p_attn_mask in enumerate(patch_attention_mask):
            nb_patches_h = p_attn_mask[:, 0].sum()
            nb_patches_w = p_attn_mask[0].sum()
            if is_transformers_version("<", "4.55"):
                fractional_coords_h = torch.arange(0, 1 - 1e-6, 1 / nb_patches_h)
                fractional_coords_w = torch.arange(0, 1 - 1e-6, 1 / nb_patches_w)
            else:
                h_indices = torch.arange(nb_patches_h, device=pixel_values.device, dtype=pixel_values.dtype)
                w_indices = torch.arange(nb_patches_w, device=pixel_values.device, dtype=pixel_values.dtype)
                fractional_coords_h = h_indices / nb_patches_h * (1 - 1e-6)
                fractional_coords_w = w_indices / nb_patches_w * (1 - 1e-6)

            bucket_coords_h = torch.bucketize(fractional_coords_h, boundaries, right=True)
            bucket_coords_w = torch.bucketize(fractional_coords_w, boundaries, right=True)

            pos_ids = (bucket_coords_h[:, None] * num_patches_per_side + bucket_coords_w).flatten()
            position_ids[batch_idx][p_attn_mask.view(-1).cpu()] = pos_ids
        return self.vision_embeddings(
            pixel_values=pixel_values, patch_attention_mask=patch_attention_mask, patch_position_ids=position_ids
        ).last_hidden_state

    def merge_vision_text_embeddings(
        self, vision_embeds, inputs_embeds, input_ids=None, attention_mask=None, position_ids=None, **kwargs
    ):
        # Adopted from https://github.com/huggingface/transformers/blob/v4.49.0-SmolVLM-2/src/transformers/models/idefics3/modeling_idefics3.py#L881
        image_features = torch.from_numpy(vision_embeds) if isinstance(vision_embeds, np.ndarray) else vision_embeds
        inputs_embeds = torch.from_numpy(inputs_embeds) if isinstance(inputs_embeds, np.ndarray) else inputs_embeds
        vision_hidden_size = image_features.shape[2]
        special_image_token_mask = input_ids == self.config.image_token_id
        #  Fixes RuntimeError: a leaf Variable that requires grad is being used in an in-place operation.
        new_inputs_embeds = inputs_embeds.clone()
        reshaped_image_hidden_states = image_features.view(-1, vision_hidden_size)
        # cast to the dtype of the input_embeds to support quantized models
        reshaped_image_hidden_states = reshaped_image_hidden_states.to(inputs_embeds.device, inputs_embeds.dtype)
        new_inputs_embeds[special_image_token_mask] = reshaped_image_hidden_states
        inputs_embeds = new_inputs_embeds

        return inputs_embeds, attention_mask, position_ids

    @staticmethod
    def preprocess_inputs(
        text: str,
        image: Optional["Image"] = None,
        processor: Optional[AutoImageProcessor] = None,
        tokenizer: Optional[PreTrainedTokenizer] = None,
        config: Optional[PretrainedConfig] = None,
        video: Optional["VideoInput"] = None,
        audio: Optional[np.ndarray] = None,
    ):
        if processor is None:
            raise ValueError("Processor is required.")

        if video is not None:
            raise ValueError("video input is not supported")

        if audio is not None:
            raise ValueError("Audio input is not supported")
        conversation = [
            {
                "role": "user",
                "content": [
                    {"type": "text", "text": text},
                ],
            }
        ]
        if image is not None:
            conversation[0]["content"].insert(0, {"type": "image"})

        text_prompt = processor.apply_chat_template(conversation, add_generation_prompt=True)

        inputs = processor(images=image, text=text_prompt, return_tensors="pt")
        return inputs


class _OVSmolVLForCasualLM(_OVIdefics3ForCausalLM):
    def merge_vision_text_embeddings(
        self, vision_embeds, inputs_embeds, input_ids=None, attention_mask=None, position_ids=None, **kwargs
    ):
        # Adopted from https://github.com/huggingface/transformers/blob/v4.49.0-SmolVLM-2/src/transformers/models/smolvlm/modeling_smolvlm.py#L803
        image_features = torch.from_numpy(vision_embeds) if isinstance(vision_embeds, np.ndarray) else vision_embeds
        inputs_embeds = torch.from_numpy(inputs_embeds) if isinstance(inputs_embeds, np.ndarray) else inputs_embeds
        _, patch_size, _ = image_features.shape

        image_mask = input_ids == self.config.image_token_id
        num_image_tokens = image_mask.sum(dim=1)
        if not torch.all(num_image_tokens % patch_size == 0):
            raise ValueError("At least one sample has <image> tokens not divisible by patch_size.")

        blocks_per_sample = num_image_tokens // patch_size

        offsets = torch.nn.functional.pad(blocks_per_sample.cumsum(dim=0), (1, 0), value=0)
        block_offset = offsets[:-1]
        row_cum = image_mask.cumsum(dim=-1)
        chunk_idx = (row_cum - 1) // patch_size
        local_idx = (row_cum - 1) % patch_size
        block_idx = block_offset.unsqueeze(1) + chunk_idx

        image_embeds = torch.zeros_like(inputs_embeds)
        image_embeds[image_mask] = image_features[block_idx[image_mask], local_idx[image_mask], :]

        inputs_embeds = torch.where(image_mask.unsqueeze(-1), image_embeds, inputs_embeds)

        return inputs_embeds, attention_mask, position_ids


class _OVPhi4MMForCausalLM(OVModelForVisualCausalLM):
    additional_parts = [
        "vision_projection",
        "audio_embeddings",
        "audio_forward_embeddings",
        "audio_encoder",
        "audio_vision_projection",
        "audio_speech_projection",
    ]

    def __init__(
        self,
        language_model: ov.Model,
        text_embeddings: ov.Model,
        vision_embeddings: ov.Model,
        config: PretrainedConfig = None,
        device: str = "CPU",
        dynamic_shapes: bool = None,
        ov_config: Optional[Dict[str, str]] = None,
        model_save_dir: Optional[Union[str, Path, TemporaryDirectory]] = None,
        quantization_config: Union[OVWeightQuantizationConfig, Dict] = None,
        **kwargs,
    ):
        super().__init__(
            language_model=language_model,
            text_embeddings=text_embeddings,
            vision_embeddings=vision_embeddings,
            config=config,
            device=device,
            dynamic_shapes=dynamic_shapes,
            ov_config=ov_config,
            model_save_dir=model_save_dir,
            quantization_config=quantization_config,
            **kwargs,
        )

        self.sub_GN = torch.tensor(self.config.sub_GN)
        self.glb_GN = torch.tensor(self.config.glb_GN)
        self.audio_config = (
            config.audio_processor["config"] if hasattr(config, "audio_processor") else config.audio_config.to_dict()
        )
        self.chunk_size = self.audio_config.get("chunk_size", -1)
        self.left_chunk = self.audio_config.get("left_chunk", 18)
        self.time_reduction = self.audio_config.get("time_reduction", 8)
        self.image_config = (
            config.img_processor if hasattr(config, "img_processor") else config.vision_config.to_dict()
        )
        self.image_size = self.image_config.get("crop_size", 448)
        self.patch_size = self.image_config.get("patch_size", 14)
        self.num_patches_per_side = self.image_size // self.patch_size
        self._IMAGE_SPECIAL_TOKEN_ID = (
            200010 if "image_token_id" not in self.image_config else self.image_config["image_token_id"]
        )
        self._AUDIO_SPECIAL_TOKEN_ID = (
            200011 if "audio_token_id" not in self.audio_config else self.audio_config["audio_token_id"]
        )
        self._COMPATIBLE_IMAGE_SPECIAL_TOKEN_ID_RANGE = [-9999, -1]  # For backward compatibility
        self._COMPATIBLE_AUDIO_SPECIAL_TOKEN_ID_RANGE = [float("-inf"), -10000]  # For backward compatibility
        self.image_dim_out = self.image_config.get("image_dim_out", self.image_config["hidden_size"])

    # Adopted from https://github.com/huggingface/transformers/blob/v4.51.3/src/transformers/models/phi4_multimodal/modeling_phi4_multimodal.py#L669
    def image_embed(
        self,
        input_ids: torch.LongTensor,
        image_pixel_values: torch.FloatTensor,
        image_attention_mask,
        inputs_embeds,
        image_sizes=None,
        **kwargs,
    ):
        input_shape = input_ids.size()
        input_ids = input_ids.view(-1, input_shape[-1])
        positions_tuple = torch.nonzero(input_ids == self._IMAGE_SPECIAL_TOKEN_ID, as_tuple=True)

        if len(positions_tuple[-1]) == 0:
            return None
        batch_size = image_pixel_values.shape[0]

        img_features = self.get_img_features(
            image_pixel_values.flatten(0, 1),
            image_attention_mask=image_attention_mask.flatten(0, 1).to(dtype=bool),
        )

        base_feat_size = int(np.sqrt(img_features.shape[1]))
        img_features = img_features.view(batch_size, -1, base_feat_size**2, self.image_dim_out)
        image_sizes = image_sizes.view(-1, 2)

        output_imgs = []
        for idx in range(batch_size):
            height, width = image_sizes[idx]
            height_ratio = height // self.image_size
            width_ratio = width // self.image_size
            area_ratio = height_ratio * width_ratio

            global_img = img_features[idx, :1]
            global_img = global_img.reshape(1, base_feat_size, base_feat_size, self.image_dim_out).contiguous()
            temporary_extensor = self.sub_GN.repeat(1, base_feat_size, 1, 1)
            global_img = torch.cat([global_img, temporary_extensor], dim=2).reshape(1, -1, self.image_dim_out)

            sub_img = img_features[idx, 1:]
            sub_img = sub_img[:area_ratio]
            sub_img = (
                sub_img.reshape(height_ratio, width_ratio, base_feat_size, base_feat_size, self.image_dim_out)
                .transpose(1, 2)
                .reshape(1, height_ratio * base_feat_size, width_ratio * base_feat_size, self.image_dim_out)
                .contiguous()
            )

            if image_attention_mask is not None:
                reshaped_image_attention_mask = (
                    image_attention_mask[idx, 1 : area_ratio + 1, 0::2, 0::2]
                    .reshape(height_ratio, width_ratio, base_feat_size, base_feat_size)
                    .transpose(1, 2)
                    .reshape(1, height_ratio * base_feat_size, width_ratio * base_feat_size)
                )
                useful_height = int(reshaped_image_attention_mask[0, :, 0].sum().item())
                useful_width = int(reshaped_image_attention_mask[0, 0, :].sum().item())
                sub_img = sub_img[:, :useful_height, :useful_width]
                temporary_extensor = self.sub_GN.repeat(1, useful_height, 1, 1)
            else:
                temporary_extensor = self.sub_GN.repeat(1, height_ratio * base_feat_size, 1, 1)

            sub_img = torch.cat([sub_img, temporary_extensor], dim=2).reshape(1, -1, self.image_dim_out)

            # Merge global and sub
            output_imgs.append(torch.cat([sub_img, self.glb_GN, global_img], dim=1))

        img_set_tensor = []
        for output_img in output_imgs:
            img_feature_proj = torch.from_numpy(self.vision_projection(output_img))
            img_set_tensor.append(img_feature_proj)

        merged_img_set_tensor = torch.cat(img_set_tensor, dim=1).squeeze(0)

        image_embeds = inputs_embeds.index_put(indices=positions_tuple, values=merged_img_set_tensor, accumulate=False)

        return image_embeds

    # Adopted from https://github.com/huggingface/transformers/blob/v4.51.3/src/transformers/models/phi4_multimodal/modeling_phi4_multimodal.py#L1241
    def audio_embed(
        self,
        input_ids: torch.LongTensor,
        audio_input_embeds: torch.FloatTensor,
        inputs_embeds,
        audio_embed_sizes=None,
        audio_projection_mode="speech",
        **kwargs,
    ):
        input_shape = input_ids.size()
        input_ids = input_ids.view(-1, input_shape[-1])

        positions_tuple = torch.nonzero(input_ids == self._AUDIO_SPECIAL_TOKEN_ID, as_tuple=True)

        if len(positions_tuple[-1]) == 0:
            return None

        audio_embeds = self.get_audio_features(audio_input_embeds, audio_projection_mode)

        merged_audio_embeds = torch.cat(
            [audio_embeds[i, : audio_embed_sizes[i], :] for i in range(len(audio_embed_sizes))], dim=0
        )
        inputs_embeds = inputs_embeds.index_put(indices=positions_tuple, values=merged_audio_embeds, accumulate=False)

        return inputs_embeds

    # Adopted from https://github.com/huggingface/transformers/blob/v4.51.3/src/transformers/models/phi4_multimodal/modeling_phi4_multimodal.py#L1165
    def get_audio_features(
        self,
        input_embeds: torch.FloatTensor,
        audio_projection_mode: str = "speech",
    ):
        xs_pad = self.audio_embeddings(input_embeds)
        input_tensor, pos_k, pos_v, hs_mask, masks = self.forward_embeddings(xs_pad)

        unfolded = False
        ori_bz, seq_len, D = input_tensor.shape
        max_seq_len = 500  # maximum position for absolute positional encoding
        masks_unfold = None
        if seq_len > max_seq_len:
            # audio sequence is longer than max_seq_len, unfold it into chunks of max_seq_len
            unfolded = True
            # the unfold op will drop residual frames, pad it to the multiple of max_seq_len
            if seq_len % max_seq_len > 0:
                chunk_pad_size = max_seq_len - (seq_len % max_seq_len)
            else:
                chunk_pad_size = 0
            if chunk_pad_size > 0:
                input_tensor_pad = torch.nn.functional.pad(
                    torch.from_numpy(input_tensor), (0, 0, 0, chunk_pad_size), "constant", 0
                )
                input_tensor = input_tensor_pad

            input_tensor = self.unfold_tensor(input_tensor, max_seq_len)
            if masks is not None:
                # revise hs_mask here because the previous calculated hs_mask did not consider extra pad
                subsampled_pad_mask = masks.squeeze(1)  # [bz, subsampled_unmask_seq_len]
                extra_padded_subsamlped_pad_mask = torch.nn.functional.pad(
                    subsampled_pad_mask, (0, chunk_pad_size), "constant", False
                )  # extra padding to the pad mask
                extra_padded_subsamlped_pad_mask = extra_padded_subsamlped_pad_mask.unsqueeze(-1).float()
                masks_unfold = self.unfold_tensor(
                    extra_padded_subsamlped_pad_mask, max_seq_len
                )  # unfold the pad mask like we did to the input tensor
                masks_unfold = masks_unfold.squeeze(-1).bool()  # unfold op does not support bool tensor
            else:
                masks_unfold = None
        hs_mask = self.calculate_hs_mask(input_tensor, masks_unfold)
        audio_features = self.audio_encoder(input_tensor, hs_mask)
        if unfolded:
            embed_dim = audio_features.shape[-1]
            audio_features = np.reshape(audio_features, (ori_bz, -1, embed_dim))
            # if we ever padded before unfolding, we need to remove the padding
            if chunk_pad_size > 0:
                audio_features = audio_features[:, :-chunk_pad_size, :]
        audio_encoder = (
            self.audio_vision_projection if audio_projection_mode == "vision" else self.audio_speech_projection
        )
        audio_set_tensor = audio_encoder(audio_features)

        return torch.from_numpy(audio_set_tensor)

    def _chunk_size_selection(self, chunk_size=None, left_chunk=None):
        """If chunk size is a list, we will randomly select a chunk size."""
        if isinstance(chunk_size, list):
            # Variable chunk size during training
            chunk_size_index = int(torch.randint(low=0, high=len(chunk_size), size=(1,)))
            chunk_size_train_eff = chunk_size[chunk_size_index]
            if not isinstance(left_chunk, list):
                raise ValueError("Since chunk_size is a list, left_chunk must be a list")
            if len(left_chunk) != len(chunk_size):
                raise ValueError("The length of left_chunk must be the same as length of chunk_size.")
            left_chunk_train_eff = left_chunk[chunk_size_index]
        else:
            chunk_size_train_eff = chunk_size
            left_chunk_train_eff = left_chunk

        return chunk_size_train_eff, left_chunk_train_eff

    # Adopted from https://github.com/huggingface/transformers/blob/v4.51.3/src/transformers/models/phi4_multimodal/modeling_phi4_multimodal.py#L1121
    def forward_embeddings(self, xs_pad, masks=None, chunk_size_nc=None, left_chunk_nc=None):
        """Forwarding the inputs through the top embedding layers

        Args:
            xs_pad: torch.Tensor
                input tensor
            masks: torch.Tensor
                input mask
            chunk_size_nc: (optional, default is None) chunk size for non-causal layers
            left_chunk_nc: (optional, default is None) # of left chunks for non-causal layers
        """
        seq_len = int(self.compute_lens_change(xs_pad.shape[1]))
        if seq_len <= 0:
            raise ValueError(
                f"""The sequence length after time reduction is invalid: {seq_len}.
                Your input feature is too short. Consider filtering out the very
                short sentence from data loader""",
            )

        batch_size = xs_pad.shape[0]

        enc_streaming_mask = self._streaming_mask(seq_len, batch_size, self.chunk_size, self.left_chunk)

        input_tensor = xs_pad

        input_tensor = self.audio_forward_embeddings(input_tensor)

        streaming_mask = enc_streaming_mask
        if streaming_mask is not None and masks is not None:
            hs_mask = masks & streaming_mask
        else:
            hs_mask = streaming_mask

        if chunk_size_nc is not None:
            enc_streaming_mask_nc = self._streaming_mask(seq_len, batch_size, chunk_size_nc, left_chunk_nc)
            if masks is not None:
                hs_mask_nc = masks & enc_streaming_mask_nc
            else:
                hs_mask_nc = enc_streaming_mask_nc
        else:
            hs_mask_nc = None

        if chunk_size_nc is None:
            return input_tensor, None, None, hs_mask, None
        return input_tensor, None, None, hs_mask, None, hs_mask_nc

    # Adopted from https://github.com/huggingface/transformers/blob/v4.51.3/src/transformers/models/phi4_multimodal/modeling_phi4_multimodal.py#L1101
    def _streaming_mask(self, seq_len, batch_size, chunk_size, left_chunk):
        chunk_size_train_eff, left_chunk_train_eff = self._chunk_size_selection(chunk_size, left_chunk)

        # Create mask matrix for streaming
        # S stores start index. if chunksize is 18, s is [0,18,36,....]
        chunk_start_idx = np.arange(0, seq_len, chunk_size_train_eff)
        # avoid randomness when run evaluation or decoding

        enc_streaming_mask = (
            self.adaptive_enc_mask(seq_len, chunk_start_idx, left_window=left_chunk_train_eff)
            .unsqueeze(0)
            .expand([batch_size, -1, -1])
        )
        return enc_streaming_mask

    def compute_lens_change(self, feature_lens):
        """feature_lens: int
        return updated feature lens.

        This used to return a different lambda function for each case that computed
        the right thing.  That does not work within Torchscript.  If you really
        need this to be faster, create nn.Module()-s for all the cases and return
        one of them.  Torchscript does support that.
        """
        nemo_conv_settings = self.audio_config.get("nemo_conv_settings")
        if nemo_conv_settings is None:
            nemo_conv_settings = {"conv_channels": self.audio_config["nemo_conv_channels"]}
        # Handle the special causal case
        subsampling_causal_cond = nemo_conv_settings.get("subsampling", "dw_striding") in [
            "dw_striding",
            "striding",
            "striding_conv1d",
        ]
        is_causal = nemo_conv_settings.get("is_causal", False)
        if is_causal and subsampling_causal_cond:
            lens_change = (
                torch.ceil(feature_lens / self.time_reduction).long()
                if isinstance(feature_lens, torch.Tensor)
                else math.ceil(feature_lens / self.time_reduction)
            )
            feature_lens_remainder = feature_lens % self.time_reduction
            if isinstance(feature_lens, torch.Tensor):
                lens_change[feature_lens_remainder != 1] += 1
            elif feature_lens_remainder != 1:
                lens_change += 1
                return lens_change
        ceil_func = math.ceil if isinstance(feature_lens, int) else torch.ceil
        return ceil_func(feature_lens / self.time_reduction)

    # Adopted from https://github.com/huggingface/transformers/blob/v4.51.3/src/transformers/models/phi4_multimodal/modeling_phi4_multimodal.py#L1146
    def calculate_hs_mask(self, xs_pad, mask):
        max_audio_length = xs_pad.shape[1]
        batch_size = xs_pad.shape[0]
        enc_streaming_mask = self._streaming_mask(max_audio_length, batch_size, self.chunk_size, self.left_chunk)
        if mask is None:
            return enc_streaming_mask

        feature_lens = mask.sum(1)
        padding_length = feature_lens
        pad_mask = torch.arange(0, max_audio_length).expand(padding_length.size(0), -1) < padding_length.unsqueeze(1)
        pad_mask = pad_mask.unsqueeze(1)
        pad_mask = pad_mask & enc_streaming_mask
        return pad_mask

    # Adopted from https://github.com/huggingface/transformers/blob/v4.51.3/src/transformers/models/phi4_multimodal/modeling_phi4_multimodal.py#L1034
    @staticmethod
    def unfold_tensor(xs_pad, max_seq_len):
        """
        For a given tensor with shape of (N, T, D), if sequence length T is longer than max_seq_len,
        this function unfold it to a (NT', max_seq_len, D) where T' is T // max_seq_len.
        Args:
            xs_pad: N, T, D
        """
        _, _, D = xs_pad.shape
        xs_pad = xs_pad.transpose(-1, -2)  # convert to N, D, T
        # N x D x 1 x T => N x (D x max_seq_len) x T'
        xs_pad = torch.nn.functional.unfold(
            xs_pad[..., None, :],
            kernel_size=(1, max_seq_len),
            stride=(1, max_seq_len),
        )

        new_bsz, _, slen = xs_pad.shape
        # N x D x max_seq_len x T'
        xs_pad = xs_pad.view(new_bsz, -1, max_seq_len, slen)
        # N x T' x max_seq_len x D
        xs_pad = xs_pad.permute(0, 3, 2, 1).contiguous()
        # NT' x max_seq_len x D
        xs_pad = xs_pad.view(-1, max_seq_len, D)
        return xs_pad

    # Adopted from https://github.com/huggingface/transformers/blob/v4.51.3/src/transformers/models/phi4_multimodal/modeling_phi4_multimodal.py#L1053
    @staticmethod
    def adaptive_enc_mask(x_len, chunk_start_idx, left_window=0, right_window=0):
        """
        The function is very important for Transformer Transducer Streaming mode
        Args:
            xs_len (int): sequence length
            chunk_start_idx (list): first idx of each chunk, such as [0,18,36,48]. It also supports adaptive chunk size [0,10,15,45]
            left_window (int): how many left chunks can be seen
            right_window (int): how many right chunks can be seen. It is used for chunk overlap model.
            Returns:
                mask (torch.Tensor): a mask tensor for streaming model
                Torch 1.0.1
                tensor([[1., 1., 0., 0.],
                        [0., 1., 1., 0.],
                        [0., 0., 1., 1.]])
                Torch 1.4.1
                tensor([[True., True., False., False.],
                        [False., True., True., False.],
                        [False., False., True., True.]])
        """
        chunk_start_idx = torch.Tensor(chunk_start_idx).long()  # first idx of each chunk, such as [0,18,36,48].
        start_pad = torch.nn.functional.pad(
            chunk_start_idx, (1, 0)
        )  # append 0 to the beginning, so it becomes [0, 0, 18, 36, 48]
        end_pad = torch.nn.functional.pad(
            chunk_start_idx, (0, 1), value=x_len
        )  # append x_len to the end, so it becomes [0,18,36,48, x_len]
        seq_range = torch.arange(0, x_len).unsqueeze(-1)  # seq_range size: [x_len, 1]
        idx = ((seq_range < end_pad) & (seq_range >= start_pad)).nonzero()[:, 1]  # idx size: [x_len]
        end_pad[idx]  # boundary size: [x_len]
        seq_range_expand = (
            torch.arange(0, x_len).unsqueeze(0).expand(x_len, -1)
        )  # seq_range_expand size [x_len, x_len]
        idx_left = idx - left_window
        idx_left[idx_left < 0] = 0
        boundary_left = start_pad[idx_left]
        mask_left = seq_range_expand >= boundary_left.unsqueeze(-1)
        idx_right = idx + right_window
        idx_right[idx_right > len(chunk_start_idx)] = len(chunk_start_idx)
        boundary_right = end_pad[idx_right]
        mask_right = seq_range_expand < boundary_right.unsqueeze(-1)
        return mask_left & mask_right

    # Adopted from https://github.com/huggingface/transformers/blob/v4.51.3/src/transformers/models/phi4_multimodal/modeling_phi4_multimodal.py#L494-L512
    @staticmethod
    def get_vision_position_ids(pixel_values, patch_attention_mask, patch_size=14, num_patches_per_side=32):
        batch_size = pixel_values.shape[0]
        max_im_h, max_im_w = pixel_values.size(2), pixel_values.size(3)
        max_nb_patches_h, max_nb_patches_w = max_im_h // patch_size, max_im_w // patch_size
        boundaries = torch.arange(1 / num_patches_per_side, 1.0, 1 / num_patches_per_side)
        position_ids = torch.full(
            size=(
                batch_size,
                max_nb_patches_h * max_nb_patches_w,
            ),
            fill_value=0,
        )

        for batch_idx, p_attn_mask in enumerate(patch_attention_mask):
            nb_patches_h = p_attn_mask[:, 0].sum()
            nb_patches_w = p_attn_mask[0].sum()

            fractional_coords_h = torch.arange(0, 1 - 1e-6, 1 / nb_patches_h)
            fractional_coords_w = torch.arange(0, 1 - 1e-6, 1 / nb_patches_w)

            bucket_coords_h = torch.bucketize(fractional_coords_h, boundaries, right=True)
            bucket_coords_w = torch.bucketize(fractional_coords_w, boundaries, right=True)

            pos_ids = (bucket_coords_h[:, None] * num_patches_per_side + bucket_coords_w).flatten()
            position_ids[batch_idx][p_attn_mask.view(-1)] = pos_ids
        return position_ids

    # Adopted from https://github.com/huggingface/transformers/blob/v4.51.3/src/transformers/models/phi4_multimodal/modeling_phi4_multimodal.py#L1561
    def embed_tokens_extend(
        self,
        input_ids: torch.LongTensor,
        input_image_embeds: torch.FloatTensor = None,
        input_audio_embeds: torch.FloatTensor = None,
        image_sizes=None,
        image_attention_mask=None,
        audio_embed_sizes=None,
        audio_projection_mode="speech",
        past_key_values=None,
    ):
        if past_key_values is not None:
            return self.language_model.embed_tokens(input_ids)

        new_input_ids = input_ids.clone()
        new_input_ids[
            (input_ids >= self._COMPATIBLE_IMAGE_SPECIAL_TOKEN_ID_RANGE[0])
            & (input_ids <= self._COMPATIBLE_IMAGE_SPECIAL_TOKEN_ID_RANGE[1])
        ] = self._IMAGE_SPECIAL_TOKEN_ID
        new_input_ids[
            (input_ids >= self._COMPATIBLE_AUDIO_SPECIAL_TOKEN_ID_RANGE[0])
            & (input_ids <= self._COMPATIBLE_AUDIO_SPECIAL_TOKEN_ID_RANGE[1])
        ] = self._AUDIO_SPECIAL_TOKEN_ID
        input_ids = new_input_ids
        image_position_mask = (input_ids == self._IMAGE_SPECIAL_TOKEN_ID).unsqueeze(-1)
        non_image_position_mask = ~image_position_mask
        hidden_states = torch.from_numpy(self.language_model.embed_tokens(input_ids))
        vision_hidden_states = self.image_embed(
            input_ids=input_ids,
            inputs_embeds=hidden_states,
            image_pixel_values=input_image_embeds,
            image_sizes=image_sizes,
            image_attention_mask=image_attention_mask,
        )
        audio_hidden_states = self.audio_embed(
            input_ids=input_ids,
            inputs_embeds=hidden_states,
            audio_input_embeds=input_audio_embeds,
            audio_embed_sizes=audio_embed_sizes,
            audio_projection_mode=audio_projection_mode,
        )
        if vision_hidden_states is not None and audio_hidden_states is not None:
            hidden_states = vision_hidden_states * image_position_mask + audio_hidden_states * non_image_position_mask
        elif vision_hidden_states is not None:
            hidden_states = vision_hidden_states
        elif audio_hidden_states is not None:
            hidden_states = audio_hidden_states

        return hidden_states

    def get_multimodal_embeddings(
        self,
        input_ids,
        pixel_values=None,
        attention_mask=None,
        position_ids=None,
        input_image_embeds: Optional[torch.FloatTensor] = None,
        image_sizes: Optional[torch.LongTensor] = None,
        image_attention_mask=None,
        input_audio_embeds: Optional[torch.FloatTensor] = None,
        audio_embed_sizes=None,
        input_mode=None,
        **kwargs,
    ):
        if pixel_values is not None and input_image_embeds is None:
            input_image_embeds = pixel_values
        audio_projection_mode = None
        if input_audio_embeds is not None:
            if isinstance(input_mode, torch.Tensor):
                assert len(input_mode) == 1
                input_mode = input_mode[0].item()
            if input_mode is None:
                input_mode = 1 if input_image_embeds is not None else 2
            input_mode = InputMode(input_mode)

            if input_mode in [InputMode.VISION_SPEECH, InputMode.VISION]:
                audio_projection_mode = "vision"
            elif input_mode == InputMode.SPEECH:
                audio_projection_mode = "speech"
            elif input_mode == InputMode.LANGUAGE:
                audio_projection_mode = "speech"
            else:
                raise ValueError(f"Invalid input_mode: {input_mode}")
        inputs_embeds = self.embed_tokens_extend(
            input_ids=input_ids,
            input_image_embeds=input_image_embeds,
            input_audio_embeds=input_audio_embeds,
            image_sizes=image_sizes,
            image_attention_mask=image_attention_mask,
            audio_embed_sizes=audio_embed_sizes,
            audio_projection_mode=audio_projection_mode,
            past_key_values=kwargs.get("past_key_values"),
        )
        return inputs_embeds, attention_mask, position_ids

    @staticmethod
    def preprocess_inputs(
        text: str,
        image: Optional["Image"] = None,
        processor: Optional[AutoImageProcessor] = None,
        tokenizer: Optional[PreTrainedTokenizer] = None,
        config: Optional[PretrainedConfig] = None,
        video: Optional["VideoInput"] = None,
        audio: Optional[np.ndarray] = None,
    ):
        if processor is None:
            raise ValueError("Processor is required.")
        if video is not None:
            raise ValueError("Video input is not supported")
        user_prompt = "<|user|>"
        assistant_prompt = "<|assistant|>"
        prompt_suffix = "<|end|>"
        image_token = getattr(processor.tokenizer, "image_token", "<|image_1|>")
        audio_token = getattr(processor.tokenizer, "audio_token", "<|audio_1|>")
        if audio is not None and audio_token not in text:
            text = audio_token + text
        if image is not None and image_token not in text:
            text = image_token + text
        if processor.tokenizer.chat_template is None:
            if not text.startswith(user_prompt):
                text = user_prompt + text + prompt_suffix + assistant_prompt
        else:
            text = processor.tokenizer.apply_chat_template(
                [{"role": "user", "content": text}], tokenize=False, add_generation_prompt=True
            )
        audio_input = {}
        if "audio" in inspect.signature(processor.__call__).parameters:
            sample_rate = None
            if isinstance(audio, tuple):
                audio, sample_rate = audio
            if isinstance(audio, list) and len(audio) == 1 and isinstance(audio[0], tuple):
                audio, sample_rate = audio[0]
            audio_input["audio"] = audio
            if sample_rate is not None:
                audio_input["sampling_rate"] = sample_rate
        else:
            audio_input["audios"] = audio
        inputs = processor(text=text, images=image, **audio_input, return_tensors="pt")
        return inputs

    def get_img_features(self, pixel_values, image_attention_mask):
        patch_position_ids = self.get_vision_position_ids(
            pixel_values, image_attention_mask, self.patch_size, self.num_patches_per_side
        )
        return torch.from_numpy(
            self.vision_embeddings(
                pixel_values=pixel_values,
                patch_attention_mask=image_attention_mask,
                patch_position_ids=patch_position_ids,
            )[0]
        )


class _OVLlama4ForCausalLM(OVModelForVisualCausalLM):
    def get_vision_embeddings(self, pixel_values, input_ids=None, **kwargs):
        if input_ids is not None and input_ids.shape[1] == 1:
            return None
        # Llama4 preprocessor creates bf16 tensor for pixel values, it can not be represented as numpy array
        if pixel_values.dtype != torch.float32:
            pixel_values.to(torch.float32)
        return self.vision_embeddings(pixel_values.to(torch.float32)).last_hidden_state

    # Adopted from https://github.com/huggingface/transformers/blob/v4.51.0/src/transformers/models/llama4/modeling_llama4.py#L1743-L1759
    def merge_vision_text_embeddings(
        self, vision_embeds, inputs_embeds, input_ids=None, attention_mask=None, position_ids=None, **kwargs
    ):
        image_features = torch.from_numpy(vision_embeds) if isinstance(vision_embeds, np.ndarray) else vision_embeds
        inputs_embeds = torch.from_numpy(inputs_embeds) if isinstance(inputs_embeds, np.ndarray) else inputs_embeds
        original_inputs_embeds_shape = inputs_embeds.shape
        special_image_mask = (input_ids == self.config.image_token_index).unsqueeze(-1)
        final_mask = special_image_mask.to(inputs_embeds.device)
        inputs_embeds = inputs_embeds.view(-1, inputs_embeds.size(-1))

        final_mask_1d = final_mask[..., 0].reshape(-1)
        num_tokens_to_fill = final_mask_1d.sum()

        if num_tokens_to_fill != image_features.size(0):
            raise ValueError(
                f"Mismatch: final_mask wants {num_tokens_to_fill} embeddings, "
                f"but multi_modal_projector returned {image_features.size(0)}"
            )

        expanded_mask = final_mask_1d.unsqueeze(-1).expand(-1, inputs_embeds.size(-1))
        inputs_embeds.masked_scatter_(expanded_mask, image_features)

        inputs_embeds = inputs_embeds.view(original_inputs_embeds_shape)

        return inputs_embeds, attention_mask, position_ids

    @staticmethod
    def preprocess_inputs(
        text: str,
        image: Optional["Image"] = None,
        processor: Optional[AutoImageProcessor] = None,
        tokenizer: Optional[PreTrainedTokenizer] = None,
        config: Optional[PretrainedConfig] = None,
        video: Optional["VideoInput"] = None,
    ):
        if processor is None:
            raise ValueError("Processor is required.")

        if video is not None:
            raise ValueError("video input is not supported")
        conversation = [
            {
                "role": "user",
                "content": [
                    {"type": "text", "text": text},
                ],
            }
        ]
        if image is not None:
            conversation[0]["content"].insert(0, {"type": "image"})

        text_prompt = processor.apply_chat_template(conversation, add_generation_prompt=True)

        inputs = processor(images=image, text=text_prompt, return_tensors="pt")
        return inputs


MODEL_TYPE_TO_CLS_MAPPING = {
    "llava": _OVLlavaForCausalLM,
    "llava_next": _OVLlavaNextForCausalLM,
    "llava_next_video": _OVLlavaNextVideoForCausalLM,
    "minicpmv": _OVMiniCPMVForCausalLM,
    "llava-qwen2": _OVNanoLlavaForCausalLM,
    "maira2": _OVMaira2ForCausalLM,
    "phi3_v": _OVPhi3VisionForCausalLM,
    "internvl_chat": _OVInternVLForCausalLM,
    "qwen2_vl": _OVQwen2VLForCausalLM,
    "qwen2_vl_text": _OVQwen2VLForCausalLM,
    "qwen2_5_vl": _OVQwen2_5_VLForCausalLM,
    "qwen2_5_vl_text": _OVQwen2_5_VLForCausalLM,
    "got_ocr2": _OVGotOCR2ForCausalLM,
    "gemma3": _OVGemma3ForCausalLM,
    "idefics3": _OVIdefics3ForCausalLM,
    "smolvlm": _OVSmolVLForCasualLM,
    "phi4mm": _OVPhi4MMForCausalLM,
    "phi4_multimodal": _OVPhi4MMForCausalLM,
    "llama4": _OVLlama4ForCausalLM,
    "qwen3_vl": _OVQwen3VLForCausalLM,
    "minicpmo": _OVMiniCPMOForCausalLM,
}