Adding all project files

venv/Lib/site-packages/transformers/models/llava_onevision/__init__.py

@@ -0,0 +1,31 @@
+# Copyright 2024 The HuggingFace Team. All rights reserved.
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+from typing import TYPE_CHECKING
+
+from ...utils import _LazyModule
+from ...utils.import_utils import define_import_structure
+
+
+if TYPE_CHECKING:
+    from .configuration_llava_onevision import *
+    from .image_processing_llava_onevision import *
+    from .image_processing_llava_onevision_fast import *
+    from .modeling_llava_onevision import *
+    from .processing_llava_onevision import *
+    from .video_processing_llava_onevision import *
+else:
+    import sys
+
+    _file = globals()["__file__"]
+    sys.modules[__name__] = _LazyModule(__name__, _file, define_import_structure(_file), module_spec=__spec__)

venv/Lib/site-packages/transformers/models/llava_onevision/configuration_llava_onevision.py

@@ -0,0 +1,196 @@
+# coding=utf-8
+# Copyright 2024 HuggingFace Inc. team. All rights reserved.
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+
+
+from ...configuration_utils import PretrainedConfig
+from ...utils import (
+    logging,
+)
+from ..auto import CONFIG_MAPPING, AutoConfig
+
+
+logger = logging.get_logger(__name__)
+
+
+class LlavaOnevisionConfig(PretrainedConfig):
+    r"""
+    This is the configuration class to store the configuration of a [`LlavaOnevisionForConditionalGeneration`]. It is used to instantiate an
+    Llava-NeXT model according to the specified arguments, defining the model architecture. Instantiating a configuration
+    with the defaults will yield a similar configuration to that of the [llava-hf/llava-onevision-qwen2-7b-ov-hf](https://huggingface.co/llava-hf/llava-onevision-qwen2-7b-ov-hf)
+    model.
+
+    Configuration objects inherit from [`PretrainedConfig`] and can be used to control the model outputs. Read the
+    documentation from [`PretrainedConfig`] for more information.
+
+    Args:
+        vision_config (`Union[AutoConfig, dict]`,  *optional*, defaults to `SiglipVisionConfig`):
+            The config object or dictionary of the vision backbone.
+        text_config (`Union[AutoConfig, dict]`, *optional*, defaults to `Qwen2Config`):
+            The config object or dictionary of the text backbone.
+        image_token_index (`int`, *optional*, defaults to 151646):
+            The image token index to encode the image prompt.
+        video_token_index (`int`, *optional*, defaults to 151647):
+            The video token index to encode the video prompt.
+        projector_hidden_act (`str`, *optional*, defaults to `"gelu"`):
+            The activation function used by the multimodal projector.
+        vision_feature_select_strategy (`str`, *optional*, defaults to `"full"`):
+            The feature selection strategy used to select the vision feature from the vision backbone.
+            Can be one of `"default"` or `"full"`. If `"default"`, the CLS token is removed from the vision features.
+            If `"full"`, the full vision features are used.
+        vision_feature_layer (`Union[int, list[int]]`, *optional*, defaults to -1):
+            The index of the layer to select the vision feature. If multiple indices are provided,
+            the vision feature of the corresponding indices will be concatenated to form the
+            vision features.
+        vision_aspect_ratio (`str`, *optional*, defaults to `"anyres_max_9"`):
+            Aspect ratio used when processong image features. The default value is "anyres_max_9".
+        image_grid_pinpoints (`List`, *optional*):
+            A list of possible resolutions to use for processing high resolution images. Each item in the list should be a tuple or list
+            of the form `(height, width)`.
+        tie_word_embeddings (`bool`, *optional*, defaults to `False`):
+            Whether the model's input and output word embeddings should be tied.
+        multimodal_projector_bias (`bool`, *optional*, defaults to `True`):
+            Whether to use bias in the multimodal projector.
+
+    Example:
+
+    ```python
+    >>> from transformers import LlavaOnevisionForConditionalGeneration, LlavaOnevisionConfig, SiglipVisionConfig, Qwen2Config
+
+    >>> # Initializing a CLIP-vision config
+    >>> vision_config = SiglipVisionConfig()
+
+    >>> # Initializing a Llama config
+    >>> text_config = Qwen2Config()
+
+    >>> # Initializing a Llava-Next llava-hf/llava-onevision-qwen2-7b-ov-hf style configuration
+    >>> configuration = LlavaOnevisionConfig(vision_config, text_config)
+
+    >>> # Initializing a model from the llava-hf/llava-onevision-qwen2-7b-ov-hf style configuration
+    >>> model = LlavaOnevisionForConditionalGeneration(configuration)
+
+    >>> # Accessing the model configuration
+    >>> configuration = model.config
+    ```"""
+
+    model_type = "llava_onevision"
+    attribute_map = {
+        "image_token_id": "image_token_index",
+        "video_token_id": "video_token_index",
+    }
+    sub_configs = {"text_config": AutoConfig, "vision_config": AutoConfig}
+
+    def __init__(
+        self,
+        vision_config=None,
+        text_config=None,
+        image_token_index=151646,
+        video_token_index=151647,
+        projector_hidden_act="gelu",
+        vision_feature_select_strategy="full",
+        vision_feature_layer=-1,
+        vision_aspect_ratio="anyres_max_9",
+        image_grid_pinpoints=None,
+        tie_word_embeddings=False,
+        multimodal_projector_bias=True,
+        **kwargs,
+    ):
+        self.image_token_index = image_token_index
+        self.video_token_index = video_token_index
+        self.projector_hidden_act = projector_hidden_act
+        self.multimodal_projector_bias = multimodal_projector_bias
+
+        if vision_feature_select_strategy not in ["default", "full"]:
+            raise ValueError(
+                "vision_feature_select_strategy should be one of 'default', 'full'."
+                f"Got: {vision_feature_select_strategy}"
+            )
+
+        self.vision_feature_select_strategy = vision_feature_select_strategy
+        self.vision_feature_layer = vision_feature_layer
+        self.vision_aspect_ratio = vision_aspect_ratio
+        image_grid_pinpoints = (
+            image_grid_pinpoints
+            if image_grid_pinpoints is not None
+            else [
+                [384, 384],
+                [384, 768],
+                [384, 1152],
+                [384, 1536],
+                [384, 1920],
+                [384, 2304],
+                [768, 384],
+                [768, 768],
+                [768, 1152],
+                [768, 1536],
+                [768, 1920],
+                [768, 2304],
+                [1152, 384],
+                [1152, 768],
+                [1152, 1152],
+                [1152, 1536],
+                [1152, 1920],
+                [1152, 2304],
+                [1536, 384],
+                [1536, 768],
+                [1536, 1152],
+                [1536, 1536],
+                [1536, 1920],
+                [1536, 2304],
+                [1920, 384],
+                [1920, 768],
+                [1920, 1152],
+                [1920, 1536],
+                [1920, 1920],
+                [1920, 2304],
+                [2304, 384],
+                [2304, 768],
+                [2304, 1152],
+                [2304, 1536],
+                [2304, 1920],
+                [2304, 2304],
+            ]
+        )
+        self.image_grid_pinpoints = image_grid_pinpoints
+
+        if isinstance(vision_config, dict):
+            vision_config["model_type"] = (
+                vision_config["model_type"] if "model_type" in vision_config else "siglip_vision_model"
+            )
+            vision_config = CONFIG_MAPPING[vision_config["model_type"]](**vision_config)
+        elif vision_config is None:
+            vision_config = CONFIG_MAPPING["siglip_vision_model"](
+                hidden_size=1152,
+                intermediate_size=4304,
+                patch_size=14,
+                image_size=384,
+                num_hidden_layers=26,
+                num_attention_heads=16,
+                vision_use_head=False,
+            )
+
+        self.vision_config = vision_config
+
+        if isinstance(text_config, dict):
+            text_config["model_type"] = text_config["model_type"] if "model_type" in text_config else "qwen2"
+            text_config = CONFIG_MAPPING[text_config["model_type"]](**text_config)
+        elif text_config is None:
+            text_config = CONFIG_MAPPING["qwen2"]()
+
+        self.text_config = text_config
+
+        super().__init__(tie_word_embeddings=tie_word_embeddings, **kwargs)
+
+
+__all__ = ["LlavaOnevisionConfig"]

venv/Lib/site-packages/transformers/models/llava_onevision/image_processing_llava_onevision.py

@@ -0,0 +1,784 @@
+# coding=utf-8
+# Copyright 2024 The HuggingFace Inc. team. All rights reserved.
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+"""Image processor class for LLaVa-Onevision."""
+
+from collections.abc import Iterable
+from typing import Optional, Union
+
+import numpy as np
+
+from ...image_processing_utils import (
+    BaseImageProcessor,
+    BatchFeature,
+    get_patch_output_size,
+    get_size_dict,
+    select_best_resolution,
+)
+from ...image_transforms import (
+    PaddingMode,
+    convert_to_rgb,
+    pad,
+    resize,
+    to_channel_dimension_format,
+)
+from ...image_utils import (
+    OPENAI_CLIP_MEAN,
+    OPENAI_CLIP_STD,
+    ChannelDimension,
+    ImageInput,
+    PILImageResampling,
+    get_image_size,
+    infer_channel_dimension_format,
+    is_scaled_image,
+    make_flat_list_of_images,
+    to_numpy_array,
+    valid_images,
+    validate_preprocess_arguments,
+)
+from ...utils import TensorType, is_vision_available, logging
+
+
+logger = logging.get_logger(__name__)
+
+
+if is_vision_available():
+    from PIL import Image
+
+
+# Copied from transformers.models.llava_next.image_processing_llava_next.divide_to_patches
+def divide_to_patches(image: np.array, patch_size: int, input_data_format) -> list[np.array]:
+    """
+    Divides an image into patches of a specified size.
+
+    Args:
+        image (`np.array`):
+            The input image.
+        patch_size (`int`):
+            The size of each patch.
+        input_data_format (`ChannelDimension` or `str`):
+            The channel dimension format of the input image.
+
+    Returns:
+        list: A list of np.array representing the patches.
+    """
+    patches = []
+    height, width = get_image_size(image, channel_dim=input_data_format)
+    for i in range(0, height, patch_size):
+        for j in range(0, width, patch_size):
+            if input_data_format == ChannelDimension.LAST:
+                patch = image[i : i + patch_size, j : j + patch_size]
+            else:
+                patch = image[:, i : i + patch_size, j : j + patch_size]
+            patches.append(patch)
+
+    return patches
+
+
+# Copied from transformers.models.llava_next.image_processing_llava_next.expand_to_square
+def expand_to_square(image: np.array, background_color, input_data_format) -> np.array:
+    """
+    Expands an image to a square by adding a background color.
+    """
+
+    height, width = get_image_size(image, channel_dim=input_data_format)
+    if width == height:
+        return image
+    elif width > height:
+        result = np.ones((width, width, image.shape[2]), dtype=image.dtype) * background_color
+        result[(width - height) // 2 : (width - height) // 2 + height, :] = image
+        return result
+    else:
+        result = np.ones((height, height, image.shape[2]), dtype=image.dtype) * background_color
+        result[:, (height - width) // 2 : (height - width) // 2 + width] = image
+        return result
+
+
+class LlavaOnevisionImageProcessor(BaseImageProcessor):
+    r"""
+    Constructs a LLaVa-Onevision image processor. Based on [`SiglipImageProcessor`] with incorporation of processing each video frame.
+
+    Args:
+        do_resize (`bool`, *optional*, defaults to `True`):
+            Whether to resize the image's (height, width) dimensions to the specified `size`. Can be overridden by
+            `do_resize` in the `preprocess` method.
+        size (`dict[str, int]` *optional*, defaults to `{"shortest_edge": 224}`):
+            Size of the image after resizing. The shortest edge of the image is resized to size["shortest_edge"], with
+            the longest edge resized to keep the input aspect ratio. Can be overridden by `size` in the `preprocess`
+            method.
+        image_grid_pinpoints (`List` *optional*, defaults to `[[672, 336], [336, 672], [672, 672], [336, 1008], [1008, 336]]`):
+            A list of possible resolutions to use for processing high resolution images. The best resolution is selected
+            based on the original size of the image. Can be overridden by `image_grid_pinpoints` in the `preprocess`
+            method. Not used for processing videos.
+        resample (`PILImageResampling`, *optional*, defaults to `Resampling.BICUBIC`):
+            Resampling filter to use if resizing the image. Can be overridden by `resample` in the `preprocess` method.
+        do_rescale (`bool`, *optional*, defaults to `True`):
+            Whether to rescale the image by the specified scale `rescale_factor`. Can be overridden by `do_rescale` in
+            the `preprocess` method.
+        rescale_factor (`int` or `float`, *optional*, defaults to `1/255`):
+            Scale factor to use if rescaling the image. Can be overridden by `rescale_factor` in the `preprocess`
+            method.
+        do_normalize (`bool`, *optional*, defaults to `True`):
+            Whether to normalize the image. Can be overridden by `do_normalize` in the `preprocess` method.
+        image_mean (`float` or `list[float]`, *optional*, defaults to `[0.48145466, 0.4578275, 0.40821073]`):
+            Mean to use if normalizing the image. This is a float or list of floats the length of the number of
+            channels in the image. Can be overridden by the `image_mean` parameter in the `preprocess` method.
+        image_std (`float` or `list[float]`, *optional*, defaults to `[0.26862954, 0.26130258, 0.27577711]`):
+            Standard deviation to use if normalizing the image. This is a float or list of floats the length of the
+            number of channels in the image. Can be overridden by the `image_std` parameter in the `preprocess` method.
+            Can be overridden by the `image_std` parameter in the `preprocess` method.
+        do_pad (`bool`, *optional*, defaults to `True`):
+            Whether to pad the image. If `True`, will pad the patch dimension of the images in the batch to the largest
+            number of patches in the batch. Padding will be applied to the bottom and right with zeros.
+        do_convert_rgb (`bool`, *optional*, defaults to `True`):
+            Whether to convert the image to RGB.
+    """
+
+    model_input_names = ["pixel_values_videos"]
+
+    def __init__(
+        self,
+        do_resize: bool = True,
+        size: Optional[dict[str, int]] = None,
+        image_grid_pinpoints: Optional[list] = None,
+        resample: PILImageResampling = PILImageResampling.BICUBIC,
+        do_rescale: bool = True,
+        rescale_factor: Union[int, float] = 1 / 255,
+        do_normalize: bool = True,
+        image_mean: Optional[Union[float, list[float]]] = None,
+        image_std: Optional[Union[float, list[float]]] = None,
+        do_pad: Optional[bool] = True,
+        do_convert_rgb: bool = True,
+        **kwargs,
+    ) -> None:
+        super().__init__(**kwargs)
+        size = size if size is not None else {"height": 384, "width": 384}
+        size = get_size_dict(size, default_to_square=False)
+        image_grid_pinpoints = (
+            image_grid_pinpoints
+            if image_grid_pinpoints is not None
+            else [
+                [384, 384],
+                [384, 768],
+                [384, 1152],
+                [384, 1536],
+                [384, 1920],
+                [384, 2304],
+                [768, 384],
+                [768, 768],
+                [768, 1152],
+                [768, 1536],
+                [768, 1920],
+                [768, 2304],
+                [1152, 384],
+                [1152, 768],
+                [1152, 1152],
+                [1152, 1536],
+                [1152, 1920],
+                [1152, 2304],
+                [1536, 384],
+                [1536, 768],
+                [1536, 1152],
+                [1536, 1536],
+                [1536, 1920],
+                [1536, 2304],
+                [1920, 384],
+                [1920, 768],
+                [1920, 1152],
+                [1920, 1536],
+                [1920, 1920],
+                [1920, 2304],
+                [2304, 384],
+                [2304, 768],
+                [2304, 1152],
+                [2304, 1536],
+                [2304, 1920],
+                [2304, 2304],
+            ]
+        )
+
+        self.do_resize = do_resize
+        self.size = size
+        self.image_grid_pinpoints = image_grid_pinpoints
+        self.resample = resample
+        self.do_rescale = do_rescale
+        self.rescale_factor = rescale_factor
+        self.do_normalize = do_normalize
+        self.image_mean = image_mean if image_mean is not None else OPENAI_CLIP_MEAN
+        self.image_std = image_std if image_std is not None else OPENAI_CLIP_STD
+        self.do_pad = do_pad
+        self.do_convert_rgb = do_convert_rgb
+
+    # Copied from transformers.models.llava_next.image_processing_llava_next.LlavaNextImageProcessor.pad
+    def pad(
+        self,
+        image: np.ndarray,
+        padding: Union[int, tuple[int, int], Iterable[tuple[int, int]]],
+        mode: PaddingMode = PaddingMode.CONSTANT,
+        constant_values: Union[float, Iterable[float]] = 0.0,
+        data_format: Optional[Union[str, ChannelDimension]] = None,
+        input_data_format: Optional[Union[str, ChannelDimension]] = None,
+    ) -> np.ndarray:
+        """
+        Pads the `image` with the specified `padding` and `mode`. Padding can be in the (`height`, `width`)
+        dimension of in the (`num_patches`) dimension. In the second case an iterable if tuples is expected
+        as input.
+
+        Args:
+            image (`np.ndarray`):
+                The image to pad.
+            padding (`int` or `tuple[int, int]` or `Iterable[tuple[int, int]]`):
+                Padding to apply to the edges of the height, width axes. Can be one of three formats:
+                - `((before_height, after_height), (before_width, after_width))` unique pad widths for each axis.
+                - `((before, after),)` yields same before and after pad for height and width.
+                - `(pad,)` or int is a shortcut for before = after = pad width for all axes.
+            mode (`PaddingMode`):
+                The padding mode to use. Can be one of:
+                    - `"constant"`: pads with a constant value.
+                    - `"reflect"`: pads with the reflection of the vector mirrored on the first and last values of the
+                    vector along each axis.
+                    - `"replicate"`: pads with the replication of the last value on the edge of the array along each axis.
+                    - `"symmetric"`: pads with the reflection of the vector mirrored along the edge of the array.
+            constant_values (`float` or `Iterable[float]`, *optional*):
+                The value to use for the padding if `mode` is `"constant"`.
+            data_format (`str` or `ChannelDimension`, *optional*):
+                The channel dimension format for the output image. Can be one of:
+                    - `"channels_first"` or `ChannelDimension.FIRST`: image in (num_channels, height, width) format.
+                    - `"channels_last"` or `ChannelDimension.LAST`: image in (height, width, num_channels) format.
+                If unset, will use same as the input image.
+            input_data_format (`str` or `ChannelDimension`, *optional*):
+                The channel dimension format for the input image. Can be one of:
+                    - `"channels_first"` or `ChannelDimension.FIRST`: image in (num_channels, height, width) format.
+                    - `"channels_last"` or `ChannelDimension.LAST`: image in (height, width, num_channels) format.
+                If unset, will use the inferred format of the input image.
+
+        Returns:
+            `np.ndarray`: The padded image.
+
+        """
+
+        # call the general `pad` if padding on `height/width`, otherwise it's the `num_patched` dim
+        if isinstance(padding, int) or len(padding) != 4:
+            return pad(image, padding, mode, constant_values, data_format, input_data_format)
+
+        if input_data_format is None:
+            input_data_format = infer_channel_dimension_format(image)
+        if mode == PaddingMode.CONSTANT:
+            image = np.pad(image, padding, mode="constant", constant_values=constant_values)
+        elif mode == PaddingMode.REFLECT:
+            image = np.pad(image, padding, mode="reflect")
+        elif mode == PaddingMode.REPLICATE:
+            image = np.pad(image, padding, mode="edge")
+        elif mode == PaddingMode.SYMMETRIC:
+            image = np.pad(image, padding, mode="symmetric")
+        else:
+            raise ValueError(f"Invalid padding mode: {mode}")
+        image = (
+            to_channel_dimension_format(image, data_format, input_data_format) if data_format is not None else image
+        )
+        return image
+
+    # Copied from transformers.models.llava_next.image_processing_llava_next.LlavaNextImageProcessor._resize_for_patching
+    def _resize_for_patching(
+        self, image: np.array, target_resolution: tuple, resample, input_data_format: ChannelDimension
+    ) -> np.array:
+        """
+        Resizes an image to a target resolution while maintaining aspect ratio.
+
+        Args:
+            image (np.array):
+                The input image.
+            target_resolution (tuple):
+                The target resolution (height, width) of the image.
+            resample (`PILImageResampling`):
+                Resampling filter to use if resizing the image.
+            input_data_format (`ChannelDimension` or `str`):
+                The channel dimension format of the input image.
+
+        Returns:
+            np.array: The resized and padded image.
+        """
+        new_height, new_width = get_patch_output_size(image, target_resolution, input_data_format)
+
+        # Resize the image
+        resized_image = resize(image, (new_height, new_width), resample=resample, input_data_format=input_data_format)
+
+        return resized_image
+
+    # Copied from transformers.models.llava_next.image_processing_llava_next.LlavaNextImageProcessor._get_padding_size
+    def _get_padding_size(self, original_resolution: tuple, target_resolution: tuple):
+        original_height, original_width = original_resolution
+        target_height, target_width = target_resolution
+        paste_x, r_x = divmod(target_width - original_width, 2)
+        paste_y, r_y = divmod(target_height - original_height, 2)
+        return (paste_y, paste_y + r_y), (paste_x, paste_x + r_x)
+
+    # Copied from transformers.models.llava_next.image_processing_llava_next.LlavaNextImageProcessor._pad_for_patching
+    def _pad_for_patching(
+        self, image: np.array, target_resolution: tuple, input_data_format: ChannelDimension
+    ) -> np.array:
+        """
+        Pad an image to a target resolution while maintaining aspect ratio.
+        """
+        new_resolution = get_patch_output_size(image, target_resolution, input_data_format)
+        padding = self._get_padding_size(new_resolution, target_resolution)
+
+        padded_image = self.pad(image, padding=padding)
+
+        return padded_image
+
+    # Copied from transformers.models.llava_next.image_processing_llava_next.LlavaNextImageProcessor.get_image_patches
+    def get_image_patches(
+        self,
+        image: np.array,
+        grid_pinpoints,
+        size: tuple,
+        patch_size: int,
+        resample: PILImageResampling,
+        data_format: ChannelDimension,
+        input_data_format: ChannelDimension,
+    ) -> list[np.array]:
+        """
+        Process an image with variable resolutions by dividing it into patches.
+
+        Args:
+            image (np.array):
+                The input image to be processed.
+            grid_pinpoints (List):
+                A string representation of a list of possible resolutions.
+            size (`tuple`):
+                Size to resize the original image to.
+            patch_size (`int`):
+                Size of the patches to divide the image into.
+            resample (`PILImageResampling`):
+                Resampling filter to use if resizing the image.
+            data_format (`ChannelDimension` or `str`):
+                The channel dimension format for the output image.
+            input_data_format (`ChannelDimension` or `str`):
+                The channel dimension format of the input image.
+
+        Returns:
+            list[np.array]: A list of NumPy arrays containing the processed image patches.
+        """
+        if not isinstance(grid_pinpoints, list):
+            raise TypeError("grid_pinpoints must be a list of possible resolutions.")
+
+        possible_resolutions = grid_pinpoints
+
+        image_size = get_image_size(image, channel_dim=input_data_format)
+        best_resolution = select_best_resolution(image_size, possible_resolutions)
+        resized_image = self._resize_for_patching(
+            image, best_resolution, resample=resample, input_data_format=input_data_format
+        )
+        padded_image = self._pad_for_patching(resized_image, best_resolution, input_data_format=input_data_format)
+
+        patches = divide_to_patches(padded_image, patch_size=patch_size, input_data_format=input_data_format)
+
+        # make sure that all patches are in the input data format
+        patches = [
+            to_channel_dimension_format(patch, channel_dim=data_format, input_channel_dim=input_data_format)
+            for patch in patches
+        ]
+
+        resized_original_image = resize(
+            image,
+            size=size,
+            resample=resample,
+            data_format=data_format,
+            input_data_format=input_data_format,
+        )
+
+        image_patches = [resized_original_image] + patches
+
+        return image_patches
+
+    # Copied from transformers.models.llava_next.image_processing_llava_next.LlavaNextImageProcessor._pad_for_batching
+    def _pad_for_batching(
+        self,
+        pixel_values: list[np.ndarray],
+        data_format: Optional[Union[str, ChannelDimension]] = None,
+        input_data_format: Optional[Union[str, ChannelDimension]] = None,
+    ):
+        """
+        Pads images on the `num_of_patches` dimension with zeros to form a batch of same number of patches.
+
+        Args:
+            pixel_values (`list[np.ndarray]`):
+                An array of pixel values of each images of shape (`batch_size`, `num_patches`, `image_in_3D`)
+            data_format (`str` or `ChannelDimension`, *optional*):
+                The channel dimension format for the output image. Can be one of:
+                    - `"channels_first"` or `ChannelDimension.FIRST`: image in (num_channels, height, width) format.
+                    - `"channels_last"` or `ChannelDimension.LAST`: image in (height, width, num_channels) format.
+                If unset, will use same as the input image.
+            input_data_format (`str` or `ChannelDimension`, *optional*):
+                The channel dimension format for the input image. Can be one of:
+                    - `"channels_first"` or `ChannelDimension.FIRST`: image in (num_channels, height, width) format.
+                    - `"channels_last"` or `ChannelDimension.LAST`: image in (height, width, num_channels) format.
+                If unset, will use the inferred format of the input image.
+
+        Returns:
+            list[`np.ndarray`]: The padded images.
+        """
+        max_patch = max(len(x) for x in pixel_values)
+        pixel_values = [
+            self.pad(
+                image,
+                padding=((0, max_patch - image.shape[0]), (0, 0), (0, 0), (0, 0)),
+                data_format=data_format,
+                input_data_format=input_data_format,
+            )
+            for image in pixel_values
+        ]
+
+        return pixel_values
+
+    # Copied from transformers.models.llava.image_processing_llava.LlavaImageProcessor.pad_to_square
+    def pad_to_square(
+        self,
+        image: np.ndarray,
+        background_color: Union[int, tuple[int, int, int]] = 0,
+        data_format: Optional[Union[str, ChannelDimension]] = None,
+        input_data_format: Optional[Union[str, ChannelDimension]] = None,
+    ) -> np.array:
+        """
+        Pads an image to a square based on the longest edge.
+
+        Args:
+            image (`np.ndarray`):
+                The image to pad.
+            background_color (`int` or `tuple[int, int, int]`, *optional*, defaults to 0):
+                The color to use for the padding. Can be an integer for single channel or a
+                tuple of integers representing for multi-channel images. If passed as integer
+                in mutli-channel mode, it will default to `0` in subsequent channels.
+            data_format (`str` or `ChannelDimension`, *optional*):
+                The channel dimension format for the output image. Can be one of:
+                    - `"channels_first"` or `ChannelDimension.FIRST`: image in (num_channels, height, width) format.
+                    - `"channels_last"` or `ChannelDimension.LAST`: image in (height, width, num_channels) format.
+                If unset, will use same as the input image.
+            input_data_format (`str` or `ChannelDimension`, *optional*):
+                The channel dimension format for the input image. Can be one of:
+                    - `"channels_first"` or `ChannelDimension.FIRST`: image in (num_channels, height, width) format.
+                    - `"channels_last"` or `ChannelDimension.LAST`: image in (height, width, num_channels) format.
+                If unset, will use the inferred format of the input image.
+
+        Returns:
+            `np.ndarray`: The padded image.
+        """
+        height, width = get_image_size(image, input_data_format)
+        num_channels = image.shape[0] if input_data_format == ChannelDimension.FIRST else image.shape[-1]
+
+        if height == width:
+            image = (
+                to_channel_dimension_format(image, data_format, input_data_format)
+                if data_format is not None
+                else image
+            )
+            return image
+
+        max_dim = max(height, width)
+
+        # Ensure background_color is the correct shape
+        if isinstance(background_color, int):
+            background_color = [background_color]
+        elif len(background_color) != num_channels:
+            raise ValueError(
+                f"background_color must have no more than {num_channels} elements to match the number of channels"
+            )
+
+        if input_data_format == ChannelDimension.FIRST:
+            result = np.zeros((num_channels, max_dim, max_dim), dtype=image.dtype)
+            for i, color in enumerate(background_color):
+                result[i, :, :] = color
+            if width > height:
+                start = (max_dim - height) // 2
+                result[:, start : start + height, :] = image
+            else:
+                start = (max_dim - width) // 2
+                result[:, :, start : start + width] = image
+        else:
+            result = np.zeros((max_dim, max_dim, num_channels), dtype=image.dtype)
+            for i, color in enumerate(background_color):
+                result[:, :, i] = color
+            if width > height:
+                start = (max_dim - height) // 2
+                result[start : start + height, :, :] = image
+            else:
+                start = (max_dim - width) // 2
+                result[:, start : start + width, :] = image
+
+        image = (
+            to_channel_dimension_format(result, data_format, input_data_format) if data_format is not None else result
+        )
+        return image
+
+    def _preprocess(
+        self,
+        images: ImageInput,
+        do_resize: Optional[bool] = None,
+        size: Optional[dict[str, int]] = None,
+        resample: PILImageResampling = None,
+        do_rescale: Optional[bool] = None,
+        rescale_factor: Optional[float] = None,
+        do_normalize: Optional[bool] = None,
+        image_mean: Optional[Union[float, list[float]]] = None,
+        image_std: Optional[Union[float, list[float]]] = None,
+        do_convert_rgb: Optional[bool] = None,
+        data_format: Optional[ChannelDimension] = ChannelDimension.FIRST,
+        input_data_format: Optional[Union[str, ChannelDimension]] = None,
+    ) -> Image.Image:
+        """
+        Args:
+            images (`ImageInput`):
+                Batch of frames (one video) to preprocess. Expects a batch of frames with pixel values ranging from 0 to 255. If
+                passing in images with pixel values between 0 and 1, set `do_rescale=False`.
+            do_resize (`bool`, *optional*, defaults to `self.do_resize`):
+                Whether to resize the image.
+            size (`dict[str, int]`, *optional*, defaults to `self.size`):
+                Size of the image after resizing. Shortest edge of the image is resized to size["shortest_edge"], with
+                the longest edge resized to keep the input aspect ratio.
+            resample (`int`, *optional*, defaults to `self.resample`):
+                Resampling filter to use if resizing the image. This can be one of the enum `PILImageResampling`. Only
+                has an effect if `do_resize` is set to `True`.
+            do_rescale (`bool`, *optional*, defaults to `self.do_rescale`):
+                Whether to rescale the image.
+            rescale_factor (`float`, *optional*, defaults to `self.rescale_factor`):
+                Rescale factor to rescale the image by if `do_rescale` is set to `True`.
+            do_normalize (`bool`, *optional*, defaults to `self.do_normalize`):
+                Whether to normalize the image.
+            image_mean (`float` or `list[float]`, *optional*, defaults to `self.image_mean`):
+                Image mean to use for normalization. Only has an effect if `do_normalize` is set to `True`.
+            image_std (`float` or `list[float]`, *optional*, defaults to `self.image_std`):
+                Image standard deviation to use for normalization. Only has an effect if `do_normalize` is set to
+                `True`.
+            data_format (`ChannelDimension` or `str`, *optional*, defaults to `ChannelDimension.FIRST`):
+                The channel dimension format for the output image. Can be one of:
+                - `"channels_first"` or `ChannelDimension.FIRST`: image in (num_channels, height, width) format.
+                - `"channels_last"` or `ChannelDimension.LAST`: image in (height, width, num_channels) format.
+                - Unset: Use the channel dimension format of the input image.
+            input_data_format (`ChannelDimension` or `str`, *optional*):
+                The channel dimension format for the input image. If unset, the channel dimension format is inferred
+                from the input image. Can be one of:
+                - `"channels_first"` or `ChannelDimension.FIRST`: image in (num_channels, height, width) format.
+                - `"channels_last"` or `ChannelDimension.LAST`: image in (height, width, num_channels) format.
+                - `"none"` or `ChannelDimension.NONE`: image in (height, width) format.
+        """
+        if do_resize:
+            images = [
+                resize(image=image, size=size, resample=resample, input_data_format=input_data_format)
+                for image in images
+            ]
+
+        if do_rescale:
+            images = [
+                self.rescale(image=image, scale=rescale_factor, input_data_format=input_data_format)
+                for image in images
+            ]
+
+        if do_normalize:
+            images = [
+                self.normalize(image=image, mean=image_mean, std=image_std, input_data_format=input_data_format)
+                for image in images
+            ]
+
+        images = [
+            to_channel_dimension_format(image, data_format, input_channel_dim=input_data_format) for image in images
+        ]
+
+        return images
+
+    def preprocess(
+        self,
+        images: ImageInput,
+        do_resize: Optional[bool] = None,
+        size: Optional[dict[str, int]] = None,
+        image_grid_pinpoints: Optional[list] = None,
+        resample: PILImageResampling = None,
+        do_rescale: Optional[bool] = None,
+        rescale_factor: Optional[float] = None,
+        do_normalize: Optional[bool] = None,
+        image_mean: Optional[Union[float, list[float]]] = None,
+        image_std: Optional[Union[float, list[float]]] = None,
+        do_pad: Optional[bool] = None,
+        do_convert_rgb: Optional[bool] = None,
+        return_tensors: Optional[Union[str, TensorType]] = None,
+        data_format: Optional[ChannelDimension] = ChannelDimension.FIRST,
+        input_data_format: Optional[Union[str, ChannelDimension]] = None,
+    ):
+        """
+        Args:
+            images (`PIL.Image.Image`, `np.ndarray`, `torch.Tensor`, `list[PIL.Image.Image]`, `list[np.ndarray]`, `list[torch.Tensor]`):
+                The image or batch of images to be prepared. Each image can be a PIL image, NumPy array or PyTorch
+                tensor. Both channels-first and channels-last formats are supported.
+            do_resize (`bool`, *optional*, defaults to `self.do_resize`):
+                Whether to resize the image.
+            size (`dict[str, int]`, *optional*, defaults to `self.size`):
+                Size of the image after resizing. Shortest edge of the image is resized to size["shortest_edge"], with
+                the longest edge resized to keep the input aspect ratio.
+            image_grid_pinpoints (`List` *optional*, defaults to `self.image_grid_pinpoints`):
+                A list of possible resolutions to use for processing high resolution images. The best resolution is
+                selected based on the original size of the image.
+            resample (`int`, *optional*, defaults to `self.resample`):
+                Resampling filter to use if resizing the image. This can be one of the enum `PILImageResampling`. Only
+                has an effect if `do_resize` is set to `True`.
+            do_rescale (`bool`, *optional*, defaults to `self.do_rescale`):
+                Whether to rescale the image.
+            rescale_factor (`float`, *optional*, defaults to `self.rescale_factor`):
+                Rescale factor to rescale the image by if `do_rescale` is set to `True`.
+            do_normalize (`bool`, *optional*, defaults to `self.do_normalize`):
+                Whether to normalize the image.
+            image_mean (`float` or `list[float]`, *optional*, defaults to `self.image_mean`):
+                Image mean to use for normalization. Only has an effect if `do_normalize` is set to `True`.
+            image_std (`float` or `list[float]`, *optional*, defaults to `self.image_std`):
+                Image standard deviation to use for normalization. Only has an effect if `do_normalize` is set to
+                `True`.
+            do_pad (`bool`, *optional*, defaults to `self.do_pad`):
+                Whether to pad the image. If `True`, will pad the patch dimension of the images in the batch to the largest
+                number of patches in the batch. Padding will be applied to the bottom and right with zeros.
+            do_convert_rgb (`bool`, *optional*, defaults to `self.do_convert_rgb`):
+                Whether to convert the image to RGB.
+            return_tensors (`str` or `TensorType`, *optional*):
+                The type of tensors to return. Can be one of:
+                - Unset: Return a list of `np.ndarray`.
+                - `TensorType.TENSORFLOW` or `'tf'`: Return a batch of type `tf.Tensor`.
+                - `TensorType.PYTORCH` or `'pt'`: Return a batch of type `torch.Tensor`.
+                - `TensorType.NUMPY` or `'np'`: Return a batch of type `np.ndarray`.
+                - `TensorType.JAX` or `'jax'`: Return a batch of type `jax.numpy.ndarray`.
+            data_format (`ChannelDimension` or `str`, *optional*, defaults to `ChannelDimension.FIRST`):
+                The channel dimension format for the output image. Can be one of:
+                - `"channels_first"` or `ChannelDimension.FIRST`: image in (num_channels, height, width) format.
+                - `"channels_last"` or `ChannelDimension.LAST`: image in (height, width, num_channels) format.
+                - Unset: Use the channel dimension format of the input image.
+            input_data_format (`ChannelDimension` or `str`, *optional*):
+                The channel dimension format for the input image. If unset, the channel dimension format is inferred
+                from the input image. Can be one of:
+                - `"channels_first"` or `ChannelDimension.FIRST`: image in (num_channels, height, width) format.
+                - `"channels_last"` or `ChannelDimension.LAST`: image in (height, width, num_channels) format.
+                - `"none"` or `ChannelDimension.NONE`: image in (height, width) format.
+
+        """
+        do_resize = do_resize if do_resize is not None else self.do_resize
+        size = size if size is not None else self.size
+        size = get_size_dict(size, default_to_square=False)
+        image_grid_pinpoints = image_grid_pinpoints if image_grid_pinpoints is not None else self.image_grid_pinpoints
+        resample = resample if resample is not None else self.resample
+        do_rescale = do_rescale if do_rescale is not None else self.do_rescale
+        rescale_factor = rescale_factor if rescale_factor is not None else self.rescale_factor
+        do_normalize = do_normalize if do_normalize is not None else self.do_normalize
+        image_mean = image_mean if image_mean is not None else self.image_mean
+        image_std = image_std if image_std is not None else self.image_std
+        do_pad = do_pad if do_pad is not None else self.do_pad
+        do_convert_rgb = do_convert_rgb if do_convert_rgb is not None else self.do_convert_rgb
+
+        if isinstance(images, (tuple, list)) and isinstance(images[0], (tuple, list)):
+            # if the first element is a list, we assume that all elements are lists
+            batch_num_images = [len(x) for x in images]
+        elif isinstance(images, (tuple, list)):
+            # treat this as a single-image case for backward compatibility
+            batch_num_images = [1] * len(images)
+        else:
+            batch_num_images = [1]
+        # only single image patching is supported
+        need_patching = [n == 1 for n in batch_num_images for _ in range(n)]
+
+        images = make_flat_list_of_images(images)
+
+        if not valid_images(images):
+            raise ValueError(
+                "Invalid image type. Must be of type PIL.Image.Image, numpy.ndarray, "
+                "torch.Tensor, tf.Tensor or jax.ndarray."
+            )
+
+        validate_preprocess_arguments(
+            do_rescale=do_rescale,
+            rescale_factor=rescale_factor,
+            do_normalize=do_normalize,
+            image_mean=image_mean,
+            image_std=image_std,
+            do_resize=do_resize,
+            size=size,
+            resample=resample,
+        )
+
+        if do_convert_rgb:
+            images = [convert_to_rgb(image) for image in images]
+
+        # All transformations expect numpy arrays.
+        images = [to_numpy_array(image) for image in images]
+
+        if do_rescale and is_scaled_image(images[0]):
+            logger.warning_once(
+                "It looks like you are trying to rescale already rescaled images. If the input"
+                " images have pixel values between 0 and 1, set `do_rescale=False` to avoid rescaling them again."
+            )
+
+        if input_data_format is None:
+            # We assume that all images have the same channel dimension format.
+            input_data_format = infer_channel_dimension_format(images[0])
+
+        size_tuple = (
+            (size["height"], size["width"])
+            if "height" in size and "width" in size
+            else (size["shortest_edge"], size["shortest_edge"])
+        )
+
+        new_images = []
+        image_sizes = [get_image_size(image, channel_dim=input_data_format) for image in images]
+        for i, image in enumerate(images):
+            if need_patching[i]:
+                # convert image into a list of patches
+                # we intentionally use the same data format as the input data format
+                image_patches = self.get_image_patches(
+                    image,
+                    image_grid_pinpoints,
+                    size=size_tuple,
+                    patch_size=size_tuple[0],
+                    resample=resample,
+                    data_format=input_data_format,
+                    input_data_format=input_data_format,
+                )
+            else:
+                padded_image = self.pad_to_square(
+                    image=image,
+                    background_color=tuple(int(x * 255) for x in self.image_mean),
+                    input_data_format=input_data_format,
+                )
+                image_patches = [padded_image]
+
+            # preprocess patches
+            pixel_values = self._preprocess(
+                image_patches,
+                do_resize=do_resize,
+                size=size_tuple,
+                resample=resample,
+                do_rescale=do_rescale,
+                rescale_factor=rescale_factor,
+                do_normalize=do_normalize,
+                image_mean=image_mean,
+                image_std=image_std,
+                data_format=data_format,
+                input_data_format=input_data_format,
+            )
+            pixel_values = np.array(pixel_values)
+            new_images.append(pixel_values)
+
+        if do_pad:
+            processed_images = self._pad_for_batching(new_images)
+
+        return BatchFeature(
+            data={"pixel_values": processed_images, "image_sizes": image_sizes, "batch_num_images": batch_num_images},
+            tensor_type=return_tensors,
+        )
+
+
+__all__ = ["LlavaOnevisionImageProcessor"]

venv/Lib/site-packages/transformers/models/llava_onevision/image_processing_llava_onevision_fast.py

@@ -0,0 +1,355 @@
+#                🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨
+#           This file was automatically generated from src/transformers/models/llava_onevision/modular_llava_onevision.py.
+#               Do NOT edit this file manually as any edits will be overwritten by the generation of
+#             the file from the modular. If any change should be done, please apply the change to the
+#                          modular_llava_onevision.py file directly. One of our CI enforces this.
+#                🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨
+# coding=utf-8
+# Copyright 2024 the HuggingFace Inc. team. All rights reserved.
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+
+from typing import Optional, Union
+
+import torch
+
+from ...image_processing_utils import BatchFeature, get_patch_output_size, select_best_resolution
+from ...image_processing_utils_fast import (
+    BaseImageProcessorFast,
+    DefaultFastImageProcessorKwargs,
+    divide_to_patches,
+    group_images_by_shape,
+    reorder_images,
+)
+from ...image_utils import (
+    OPENAI_CLIP_MEAN,
+    OPENAI_CLIP_STD,
+    ChannelDimension,
+    ImageInput,
+    PILImageResampling,
+    SizeDict,
+    get_image_size,
+)
+from ...processing_utils import Unpack
+from ...utils import TensorType, auto_docstring, is_torchvision_v2_available
+
+
+if is_torchvision_v2_available():
+    from torchvision.transforms.v2 import functional as F
+else:
+    from torchvision.transforms import functional as F
+
+
+class LlavaOnevisionFastImageProcessorKwargs(DefaultFastImageProcessorKwargs):
+    """
+    image_grid_pinpoints (`list[list[int]]`, *optional*):
+        A list of possible resolutions to use for processing high resolution images. The best resolution is selected
+        based on the original size of the image. Can be overridden by `image_grid_pinpoints` in the `preprocess`
+        method.
+    do_pad (`bool`, *optional*):
+        Whether to pad the image. If `True`, will pad the patch dimension of the images in the batch to the largest
+        number of patches in the batch. Padding will be applied to the bottom and right with zeros.
+    """
+
+    image_grid_pinpoints: Optional[list[list[int]]]
+    do_pad: Optional[bool]
+
+
+@auto_docstring
+class LlavaOnevisionImageProcessorFast(BaseImageProcessorFast):
+    resample = PILImageResampling.BICUBIC
+    image_mean = OPENAI_CLIP_MEAN
+    image_std = OPENAI_CLIP_STD
+    size = {"height": 384, "width": 384}
+    default_to_square = False
+    crop_size = None
+    do_resize = True
+    do_center_crop = None
+    do_rescale = True
+    do_normalize = True
+    do_convert_rgb = True
+    do_pad = True
+    image_grid_pinpoints = [[384, 384], [384, 768], [384, 1152], [384, 1536], [384, 1920], [384, 2304], [768, 384], [768, 768], [768, 1152], [768, 1536], [768, 1920], [768, 2304], [1152, 384], [1152, 768], [1152, 1152], [1152, 1536], [1152, 1920], [1152, 2304], [1536, 384], [1536, 768], [1536, 1152], [1536, 1536], [1536, 1920], [1536, 2304], [1920, 384], [1920, 768], [1920, 1152], [1920, 1536], [1920, 1920], [1920, 2304], [2304, 384], [2304, 768], [2304, 1152], [2304, 1536], [2304, 1920], [2304, 2304]]  # fmt: skip
+    valid_kwargs = LlavaOnevisionFastImageProcessorKwargs
+    model_input_names = ["pixel_values_videos"]
+
+    def __init__(self, **kwargs: Unpack[LlavaOnevisionFastImageProcessorKwargs]):
+        super().__init__(**kwargs)
+
+    @auto_docstring
+    def preprocess(self, images: ImageInput, **kwargs: Unpack[LlavaOnevisionFastImageProcessorKwargs]) -> BatchFeature:
+        if isinstance(images, (tuple, list)) and isinstance(images[0], (tuple, list)):
+            # if the first element is a list, we assume that all elements are lists
+            batch_num_images = [len(x) for x in images]
+        elif isinstance(images, (tuple, list)):
+            # treat this as a single-image case for backward compatibility
+            batch_num_images = [1] * len(images)
+        else:
+            batch_num_images = [1]
+        kwargs["batch_num_images"] = batch_num_images
+        return super().preprocess(images, **kwargs)
+
+    def _resize_for_patching(
+        self,
+        image: "torch.Tensor",
+        target_resolution: tuple,
+        interpolation: "F.InterpolationMode",
+        input_data_format: ChannelDimension,
+    ) -> "torch.Tensor":
+        """
+        Resizes an image to a target resolution while maintaining aspect ratio.
+
+        Args:
+            image ("torch.Tensor"):
+                The input image.
+            target_resolution (tuple):
+                The target resolution (height, width) of the image.
+            interpolation (`InterpolationMode`):
+                Resampling filter to use if resizing the image.
+            input_data_format (`ChannelDimension` or `str`):
+                The channel dimension format of the input image.
+
+        Returns:
+            "torch.Tensor": The resized and padded image.
+        """
+        new_height, new_width = get_patch_output_size(image, target_resolution, input_data_format)
+
+        # Resize the image
+        resized_image = self.resize(
+            image=image,
+            size=SizeDict(height=new_height, width=new_width),
+            interpolation=interpolation,
+        )
+
+        return resized_image
+
+    def _get_padding_size(self, original_resolution: tuple, target_resolution: tuple):
+        original_height, original_width = original_resolution
+        target_height, target_width = target_resolution
+        paste_x, r_x = divmod(target_width - original_width, 2)
+        paste_y, r_y = divmod(target_height - original_height, 2)
+        return [paste_x, paste_y, paste_x + r_x, paste_y + r_y]
+
+    def _pad_for_patching(
+        self, image: "torch.Tensor", target_resolution: tuple, input_data_format: ChannelDimension
+    ) -> "torch.Tensor":
+        """
+        Pad an image to a target resolution while maintaining aspect ratio.
+        """
+        new_resolution = get_patch_output_size(image, target_resolution, input_data_format)
+        padding = self._get_padding_size(new_resolution, target_resolution)
+
+        padded_image = F.pad(image, padding=padding)
+
+        return padded_image
+
+    def _get_image_patches(
+        self,
+        image: "torch.Tensor",
+        grid_pinpoints,
+        size: tuple,
+        patch_size: int,
+        interpolation: "F.InterpolationMode",
+    ) -> list["torch.Tensor"]:
+        """
+        Process an image with variable resolutions by dividing it into patches.
+
+        Args:
+            image ("torch.Tensor"):
+                The input image to be processed.
+            grid_pinpoints (List):
+                A string representation of a list of possible resolutions.
+            size (`tuple`):
+                Size to resize the original image to.
+            patch_size (`int`):
+                Size of the patches to divide the image into.
+            interpolation (`"InterpolationMode"`):
+                Resampling filter to use if resizing the image.
+
+        Returns:
+            list["torch.Tensor"]: A list of NumPy arrays containing the processed image patches.
+        """
+        if not isinstance(grid_pinpoints, list):
+            raise TypeError("grid_pinpoints must be a list of possible resolutions.")
+
+        possible_resolutions = grid_pinpoints
+
+        image_size = get_image_size(image, channel_dim=ChannelDimension.FIRST)
+        best_resolution = select_best_resolution(image_size, possible_resolutions)
+        resized_image = self._resize_for_patching(
+            image, best_resolution, interpolation=interpolation, input_data_format=ChannelDimension.FIRST
+        )
+        padded_image = self._pad_for_patching(resized_image, best_resolution, input_data_format=ChannelDimension.FIRST)
+        patches = divide_to_patches(padded_image, patch_size=patch_size)
+        resized_original_image = F.resize(image, size=size, interpolation=interpolation)
+
+        image_patches = [resized_original_image] + patches
+
+        return image_patches
+
+    def _pad_for_batching(
+        self,
+        pixel_values: list["torch.Tensor"],
+    ) -> list["torch.Tensor"]:
+        """
+        Pads images on the `num_of_patches` dimension with zeros to form a batch of same number of patches.
+
+        Args:
+            pixel_values (`list[torch.Tensor]`):
+                An array of pixel values of each images of shape (`batch_size`, `num_patches`, `image_in_3D`)
+
+        Returns:
+            list[`torch.Tensor`]: The padded images.
+        """
+        max_patch = max(len(x) for x in pixel_values)
+        pixel_values = [
+            torch.nn.functional.pad(image, pad=[0, 0, 0, 0, 0, 0, 0, max_patch - image.shape[0]])
+            for image in pixel_values
+        ]
+
+        return pixel_values
+
+    def _preprocess(
+        self,
+        images: list["torch.Tensor"],
+        do_resize: bool,
+        size: SizeDict,
+        image_grid_pinpoints: list[list[int]],
+        interpolation: Optional["F.InterpolationMode"],
+        do_center_crop: bool,
+        crop_size: SizeDict,
+        do_rescale: bool,
+        rescale_factor: float,
+        do_normalize: bool,
+        image_mean: Optional[Union[float, list[float]]],
+        image_std: Optional[Union[float, list[float]]],
+        do_pad: bool,
+        batch_num_images: list[int],
+        disable_grouping: Optional[bool],
+        return_tensors: Optional[Union[str, TensorType]],
+        **kwargs,
+    ) -> BatchFeature:
+        processed_images = []
+        image_sizes = []
+
+        # only single image patching is supported
+        need_patching = [n == 1 for n in batch_num_images for _ in range(n)]
+
+        # Determine the size tuple
+        if size and size.height and size.width:
+            size_tuple = (size.height, size.width)
+        else:
+            size_tuple = (size.shortest_edge, size.shortest_edge)
+
+        # Determine the patch size
+        if crop_size and crop_size.height:
+            patch_size = crop_size.height
+        elif size and size.height:
+            patch_size = size.height
+        else:
+            patch_size = size.shortest_edge
+
+        for i, image in enumerate(images):
+            if need_patching[i]:
+                image_patches = self._get_image_patches(
+                    image,
+                    image_grid_pinpoints,
+                    size=size_tuple,
+                    patch_size=patch_size,
+                    interpolation=interpolation,
+                )
+            else:
+                padded_image = self.pad_to_square(
+                    images=image, background_color=tuple(int(x * 255) for x in self.image_mean)
+                )
+                image_patches = [padded_image]
+
+            # Group images by size for batched processing
+            processed_image_patches_grouped = {}
+            grouped_image_patches, grouped_image_patches_index = group_images_by_shape(
+                image_patches, disable_grouping=disable_grouping
+            )
+            for shape, stacked_image_patches in grouped_image_patches.items():
+                if do_resize:
+                    stacked_image_patches = self.resize(
+                        image=stacked_image_patches,
+                        size=size,
+                        interpolation=interpolation,
+                    )
+                if do_center_crop:
+                    stacked_image_patches = self.center_crop(stacked_image_patches, crop_size)
+                # Fused rescale and normalize
+                stacked_image_patches = self.rescale_and_normalize(
+                    stacked_image_patches, do_rescale, rescale_factor, do_normalize, image_mean, image_std
+                )
+                processed_image_patches_grouped[shape] = stacked_image_patches
+            processed_image_patches = reorder_images(processed_image_patches_grouped, grouped_image_patches_index)
+            processed_image_patches = (
+                torch.stack(processed_image_patches, dim=0) if return_tensors else processed_image_patches
+            )
+            processed_images.append(processed_image_patches)
+            image_sizes.append(get_image_size(image, ChannelDimension.FIRST))
+
+        if do_pad:
+            processed_images = self._pad_for_batching(processed_images)
+        processed_images = torch.stack(processed_images, dim=0) if return_tensors else processed_images
+        return BatchFeature(
+            data={"pixel_values": processed_images, "image_sizes": image_sizes, "batch_num_images": batch_num_images},
+            tensor_type=return_tensors,
+        )
+
+    # Copied from transformers.models.llava.image_processing_llava_fast.LlavaImageProcessorFast.pad_to_square
+    def pad_to_square(
+        self,
+        images: "torch.Tensor",
+        background_color: Union[int, tuple[int, int, int]] = 0,
+    ) -> "torch.Tensor":
+        """
+        Pads an image to a square based on the longest edge.
+
+        Args:
+            images (`np.ndarray`):
+                The images to pad.
+            background_color (`int` or `tuple[int, int, int]`, *optional*, defaults to 0):
+                The color to use for the padding. Can be an integer for single channel or a
+                tuple of integers representing for multi-channel images. If passed as integer
+                in mutli-channel mode, it will default to `0` in subsequent channels.
+        Returns:
+            `torch.Tensor`: The padded images.
+        """
+        height, width = get_image_size(images, ChannelDimension.FIRST)
+
+        if height == width:
+            return images
+
+        num_channels = images.shape[1] if len(images.shape) == 4 else images.shape[0]
+        if isinstance(background_color, int):
+            background_color = [background_color] + [0] * (num_channels - 1)
+        elif len(background_color) != num_channels:
+            raise ValueError(
+                f"background_color must have no more than {num_channels} elements to match the number of channels"
+            )
+
+        max_dim = max(height, width)
+        paste_x_left = (max_dim - width) // 2
+        paste_y_left = (max_dim - height) // 2
+        paste_x_right = max_dim - width - paste_x_left
+        paste_y_right = max_dim - height - paste_y_left
+        padded_images = F.pad(
+            images, padding=[paste_x_left, paste_y_left, paste_x_right, paste_y_right], fill=background_color
+        )
+
+        return padded_images
+
+
+__all__ = ["LlavaOnevisionImageProcessorFast"]

venv/Lib/site-packages/transformers/models/llava_onevision/modeling_llava_onevision.py

@@ -0,0 +1,999 @@
+#                🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨
+#           This file was automatically generated from src/transformers/models/llava_onevision/modular_llava_onevision.py.
+#               Do NOT edit this file manually as any edits will be overwritten by the generation of
+#             the file from the modular. If any change should be done, please apply the change to the
+#                          modular_llava_onevision.py file directly. One of our CI enforces this.
+#                🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨
+# coding=utf-8
+# Copyright 2024 the HuggingFace Inc. team. All rights reserved.
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+
+import math
+from dataclasses import dataclass
+from typing import Optional, Union
+
+import numpy as np
+import torch
+from torch import nn
+
+from ...activations import ACT2FN
+from ...cache_utils import Cache
+from ...generation import GenerationMixin
+from ...image_processing_utils import select_best_resolution
+from ...modeling_flash_attention_utils import FlashAttentionKwargs
+from ...modeling_outputs import BaseModelOutputWithPast, ModelOutput
+from ...modeling_utils import PreTrainedModel
+from ...processing_utils import Unpack
+from ...utils import (
+    TransformersKwargs,
+    auto_docstring,
+    can_return_tuple,
+    is_torchdynamo_compiling,
+    logging,
+)
+from ..auto import AutoModel
+from .configuration_llava_onevision import LlavaOnevisionConfig
+
+
+logger = logging.get_logger(__name__)
+
+
+@dataclass
+@auto_docstring(
+    custom_intro="""
+    Base class for Llava outputs, with hidden states and attentions.
+    """
+)
+class LlavaOnevisionModelOutputWithPast(BaseModelOutputWithPast):
+    r"""
+    past_key_values (`Cache`, *optional*, returned when `use_cache=True` is passed or when `config.use_cache=True`):
+        Tuple of `tuple(torch.FloatTensor)` of length `config.n_layers`, with each tuple having 2 tensors of shape
+        `(batch_size, num_heads, sequence_length, embed_size_per_head)`)
+
+        Contains pre-computed hidden-states (key and values in the self-attention blocks) that can be used (see
+        `past_key_values` input) to speed up sequential decoding.
+    image_hidden_states (`torch.FloatTensor`, *optional*):
+        A `torch.FloatTensor` of size `(batch_size, num_images, sequence_length, hidden_size)`.
+        image_hidden_states of the model produced by the vision encoder and after projecting the last hidden state.
+    video_hidden_states (`torch.FloatTensor`, *optional*):
+        A `torch.FloatTensor`  of size `(batch_size * num_frames, num_videos, sequence_length, hidden_size)`.
+        video_hidden_states of the model produced by the vision encoder and after projecting the last hidden state.
+    """
+
+    image_hidden_states: Optional[torch.FloatTensor] = None
+
+    video_hidden_states: Optional[torch.FloatTensor] = None
+
+
+@dataclass
+@auto_docstring(
+    custom_intro="""
+    Base class for LlavaOnevision causal language model (or autoregressive) outputs.
+    """
+)
+class LlavaOnevisionCausalLMOutputWithPast(ModelOutput):
+    r"""
+    loss (`torch.FloatTensor` of shape `(1,)`, *optional*, returned when `labels` is provided):
+        Language modeling loss (for next-token prediction).
+    logits (`torch.FloatTensor` of shape `(batch_size, sequence_length, config.vocab_size)`):
+        Prediction scores of the language modeling head (scores for each vocabulary token before SoftMax).
+    past_key_values (`Cache`, *optional*, returned when `use_cache=True` is passed or when `config.use_cache=True`):
+        Tuple of `tuple(torch.FloatTensor)` of length `config.n_layers`, with each tuple having 2 tensors of shape
+        `(batch_size, num_heads, sequence_length, embed_size_per_head)`)
+
+        Contains pre-computed hidden-states (key and values in the self-attention blocks) that can be used (see
+        `past_key_values` input) to speed up sequential decoding.
+    image_hidden_states (`torch.FloatTensor`, *optional*):
+        A `torch.FloatTensor` of size (batch_size * num_patches, num_images, sequence_length, hidden_size)`.
+        image_hidden_states of the model produced by the vision encoder and after projecting the last hidden state.
+    video_hidden_states (`torch.FloatTensor`, *optional*):
+        A `torch.FloatTensor`  of size `(batch_size * num_frames, num_videos, sequence_length, hidden_size)`.
+        video_hidden_states of the model produced by the vision encoder and after projecting the last hidden state.
+    """
+
+    loss: Optional[torch.FloatTensor] = None
+    logits: Optional[torch.FloatTensor] = None
+    past_key_values: Optional[list[torch.FloatTensor]] = None
+    hidden_states: Optional[tuple[torch.FloatTensor]] = None
+    attentions: Optional[tuple[torch.FloatTensor]] = None
+    image_hidden_states: Optional[torch.FloatTensor] = None
+
+    video_hidden_states: Optional[torch.FloatTensor] = None
+
+
+class LlavaOnevisionPooler(nn.Module):
+    def __init__(self, config):
+        super().__init__()
+
+        mode = config.spatial_pool_mode
+        stride = config.spatial_pool_stride
+        out_channels = getattr(config, "spatial_pool_out_channels", config.vision_config.hidden_size)
+        self.image_size = (config.vision_config.image_size // config.vision_config.patch_size) ** 2
+
+        if mode == "average":
+            self.pool = nn.AvgPool2d(kernel_size=stride, stride=stride)
+        elif mode == "max":
+            self.pool = nn.MaxPool2d(kernel_size=stride, stride=stride)
+        elif mode == "conv":
+            self.pool = nn.Conv2d(
+                in_channels=config.vision_config.hidden_size,
+                out_channels=out_channels,
+                kernel_size=stride,
+                stride=stride,
+            )
+        else:
+            raise ValueError(f"Unknown pooling mode: {mode}. Has to be one of [`average`, `max`, `conv`]")
+
+    def forward(self, image_features):
+        ori_width = int(math.sqrt(image_features.shape[1] * self.image_size // self.image_size))
+        ori_height = int(ori_width * self.image_size // self.image_size)
+
+        batch_size, _, dim = image_features.shape
+        image_features_spatial = image_features.view(batch_size, ori_height, ori_height, dim).permute(0, 3, 1, 2)
+        image_features_spatial_pool = self.pool(image_features_spatial)
+
+        return image_features_spatial_pool.flatten(2).transpose(1, 2).contiguous()
+
+
+class LlavaOnevisionMultiModalProjector(nn.Module):
+    def __init__(self, config: LlavaOnevisionConfig):
+        super().__init__()
+        # We have hidden_size * the number of vision feature layers
+        num_feature_layers = 1 if isinstance(config.vision_feature_layer, int) else len(config.vision_feature_layer)
+        self.linear_1 = nn.Linear(
+            config.vision_config.hidden_size * num_feature_layers,
+            config.text_config.hidden_size,
+            bias=config.multimodal_projector_bias,
+        )
+        self.act = ACT2FN[config.projector_hidden_act]
+        self.linear_2 = nn.Linear(
+            config.text_config.hidden_size, config.text_config.hidden_size, bias=config.multimodal_projector_bias
+        )
+
+    def forward(self, image_features):
+        hidden_states = self.linear_1(image_features)
+        hidden_states = self.act(hidden_states)
+        hidden_states = self.linear_2(hidden_states)
+        return hidden_states
+
+
+def get_anyres_image_grid_shape(image_size, grid_pinpoints, patch_size):
+    """
+    Calculate the shape of the image patch grid after the preprocessing for images of any resolution.
+
+    Args:
+        image_size (`tuple`):
+            The size of the input image in the format (width, height).
+        grid_pinpoints (`List`):
+            A list containing possible resolutions. Each item in the list should be a tuple or list
+            of the form `(height, width)`.
+        patch_size (`int`):
+            The size of each image patch.
+
+    Returns:
+        tuple: The shape of the image patch grid in the format (width, height).
+    """
+    if not isinstance(grid_pinpoints, list):
+        raise TypeError("grid_pinpoints should be a list of tuples or lists")
+
+    # ! VERY IMPORTANT if image_size is tensor, must convert to into tuple, otherwise it will cause wrong calculate
+    if not isinstance(image_size, (list, tuple)):
+        if not isinstance(image_size, (torch.Tensor, np.ndarray)):
+            raise TypeError(
+                f"image_size invalid type: {type(image_size)} not valid, should be either list, tuple, np.ndarray or tensor"
+            )
+        image_size = image_size.tolist()
+
+    height, width = select_best_resolution(image_size, grid_pinpoints)
+    return height // patch_size, width // patch_size
+
+
+def image_size_to_num_patches(image_size, grid_pinpoints, patch_size: int):
+    """
+    Calculate the number of patches after the preprocessing for images of any resolution.
+
+    Args:
+        image_size (`torch.LongTensor` or `np.ndarray` or `tuple[int, int]`):
+            The size of the input image in the format (height, width). ?
+        grid_pinpoints (`List`):
+            A list containing possible resolutions. Each item in the list should be a tuple or list
+            of the form `(height, width)`.
+        patch_size (`int`):
+            The size of each image patch.
+
+    Returns:
+        int: the number of patches
+    """
+    if not isinstance(grid_pinpoints, list):
+        raise TypeError("grid_pinpoints should be a list of tuples or lists")
+
+    # ! VERY IMPORTANT if image_size is tensor, must convert to into tuple, otherwise it will cause wrong calculate
+    if not isinstance(image_size, (list, tuple)):
+        if not isinstance(image_size, (torch.Tensor, np.ndarray)):
+            raise TypeError(f"image_size invalid type {type(image_size)} with value {image_size}")
+        image_size = image_size.tolist()
+
+    best_resolution = select_best_resolution(image_size, grid_pinpoints)
+    height, width = best_resolution
+    num_patches = 0
+    # consider change to ceil(height/patch_size)*ceil(width/patch_size) + 1
+    for i in range(0, height, patch_size):
+        for j in range(0, width, patch_size):
+            num_patches += 1
+    # add the base patch
+    num_patches += 1
+    return num_patches
+
+
+def unpad_image(tensor, original_size):
+    """
+    Unpads a PyTorch tensor of a padded and resized image.
+
+    Args:
+        tensor (`torch.Tensor`):
+            The image tensor, assumed to be of shape (num_channels, height, width).
+        original_size (`tuple`):
+            The original size of the image (height, width).
+
+    Returns:
+        `torch.Tensor`: The unpadded image tensor.
+    """
+    if not isinstance(original_size, (list, tuple)):
+        if not isinstance(original_size, (torch.Tensor, np.ndarray)):
+            raise TypeError(
+                f"image_size invalid type: {type(original_size)} not valid, should be either list, tuple, np.ndarray or tensor"
+            )
+        original_size = original_size.tolist()
+    original_height, original_width = original_size
+    current_height, current_width = tensor.shape[1:]
+
+    original_aspect_ratio = original_width / original_height
+    current_aspect_ratio = current_width / current_height
+
+    if original_aspect_ratio > current_aspect_ratio:
+        scale_factor = current_width / original_width
+        new_height = int(round(original_height * scale_factor, 7))
+        padding = (current_height - new_height) // 2
+        unpadded_tensor = tensor[:, padding : current_height - padding, :]
+    else:
+        scale_factor = current_height / original_height
+        new_width = int(round(original_width * scale_factor, 7))
+        padding = (current_width - new_width) // 2
+        unpadded_tensor = tensor[:, :, padding : current_width - padding]
+
+    return unpadded_tensor
+
+
+@auto_docstring
+class LlavaOnevisionPreTrainedModel(PreTrainedModel):
+    config: LlavaOnevisionConfig
+    base_model_prefix = ""
+    supports_gradient_checkpointing = True
+    _no_split_modules = ["LlamaDecoderLayer"]
+    _skip_keys_device_placement = "past_key_values"
+
+    _supports_flash_attn = True
+    _supports_sdpa = True
+
+    _can_compile_fullgraph = True
+    _supports_flex_attn = True
+    _supports_attention_backend = True
+
+    def _init_weights(self, module):
+        std = getattr(self.config, "initializer_range", self.config.get_text_config().initializer_range)
+
+        if isinstance(module, nn.Linear):
+            module.weight.data.normal_(mean=0.0, std=std)
+            if module.bias is not None:
+                module.bias.data.zero_()
+        elif isinstance(module, LlavaOnevisionModel):
+            embed_std = 1 / math.sqrt(self.config.text_config.hidden_size)
+            module.image_newline.data.normal_(mean=0.0, std=embed_std)
+
+
+@auto_docstring(
+    custom_intro="""
+    The Llava-Next model which consists of a vision backbone and a language model without language modeling head.
+    """
+)
+class LlavaOnevisionModel(LlavaOnevisionPreTrainedModel):
+    _checkpoint_conversion_mapping = {"language_model.model": "language_model"}
+
+    def __init__(self, config):
+        super().__init__(config)
+        self.vision_tower = AutoModel.from_config(config.vision_config)
+
+        self.multi_modal_projector = LlavaOnevisionMultiModalProjector(config)
+        embed_std = 1 / math.sqrt(config.text_config.hidden_size)
+        self.image_newline = nn.Parameter(torch.randn(config.text_config.hidden_size, dtype=self.dtype) * embed_std)
+
+        self.vocab_size = config.text_config.vocab_size
+        self.language_model = AutoModel.from_config(config.text_config)
+        self.pad_token_id = self.config.pad_token_id if self.config.pad_token_id is not None else -1
+        self.post_init()
+
+    def get_input_embeddings(self):
+        return self.language_model.get_input_embeddings()
+
+    def set_input_embeddings(self, value):
+        self.language_model.set_input_embeddings(value)
+
+    def set_decoder(self, decoder):
+        self.language_model = decoder
+
+    def get_decoder(self):
+        return self.language_model
+
+    def pack_image_features(self, image_features, image_sizes, image_newline=None, vision_aspect_ratio="anyres_max_9"):
+        """
+        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.
+            vision_aspect_ratio (`str`, *optional*, "anyres_max_9"):
+                Aspect ratio used when processong image features. The default value is "anyres_max_9".
+        Returns:
+            image_features (`torch.Tensor` of shape `(all_feat_len, embed_dim)`)
+            feature_lens (`list[int]`)
+                token length of each image in image_features
+        """
+        new_image_features = []
+        feature_lens = []
+        for image_idx, image_feature in enumerate(image_features):
+            if image_feature.shape[0] > 1:
+                base_image_feature = image_feature[0]
+                image_feature = image_feature[1:]
+                height = width = self.config.vision_config.image_size // self.config.vision_config.patch_size
+                if height * width != base_image_feature.shape[0]:
+                    raise ValueError("The number of patches is not consistent with the image size.")
+                num_patch_height, num_patch_width = get_anyres_image_grid_shape(
+                    image_sizes[image_idx],
+                    self.config.image_grid_pinpoints,
+                    self.config.vision_config.image_size,
+                )
+                image_feature = image_feature.view(num_patch_height, num_patch_width, height, width, -1)
+                image_feature = image_feature.permute(4, 0, 2, 1, 3).contiguous()
+                image_feature = image_feature.flatten(1, 2).flatten(2, 3)
+                image_feature = unpad_image(image_feature, image_sizes[image_idx])
+                max_num_patches = int(vision_aspect_ratio.strip("anyres_max_"))
+                channels, curr_height, curr_width = image_feature.shape
+                ratio = math.sqrt(curr_height * curr_width / (max_num_patches * height**2))
+                if ratio > 1.1:
+                    image_feature = image_feature[None]
+                    image_feature = nn.functional.interpolate(
+                        image_feature, [int(curr_height // ratio), int(curr_width // ratio)], mode="bilinear"
+                    )[0]
+                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)
+                image_feature = image_feature.flatten(0, 1)
+            new_image_features.append(image_feature)
+            feature_lens.append(image_feature.size(0))
+        feature_lens = torch.tensor(feature_lens, dtype=torch.long, device=image_features[0].device)
+        return new_image_features, feature_lens
+
+    def get_image_features(
+        self,
+        pixel_values: torch.FloatTensor,
+        image_sizes: torch.Tensor,
+        vision_feature_layer: Optional[Union[int, list[int]]] = None,
+        vision_feature_select_strategy: Optional[str] = None,
+        vision_aspect_ratio: Optional[str] = None,
+        batch_num_images: Optional[torch.LongTensor] = None,
+    ):
+        """
+        Obtains image last hidden states from the vision tower and apply multimodal projection.
+
+        Args:
+            pixel_values (`torch.FloatTensor]` of shape `(batch_size, num_patches, channels, height, width)`)
+               The tensors corresponding to the input images.
+            image_sizes (`torch.Tensor` of shape `(num_images, 2)`)
+                Actual image size of each images (H, W).
+            vision_feature_layer (`Union[int, list[int]]`):
+                The index of the layer to select the vision feature. If multiple indices are provided,
+                the vision feature of the corresponding indices will be concatenated to form the
+                vision features.
+            vision_feature_select_strategy (`str`):
+                The feature selection strategy used to select the vision feature from the vision backbone.
+                Can be one of `"default"` or `"full"`
+            batch_num_images (`torch.LongTensor`, *optional*):
+                Number of images in each sample.
+        Returns:
+            image_features (list[`torch.Tensor`]): List of image feature tensor, each contains all the visual feature of all patches
+            and are of shape `(num_patches, image_length, embed_dim)`).
+        """
+        vision_feature_layer = (
+            vision_feature_layer if vision_feature_layer is not None else self.config.vision_feature_layer
+        )
+        vision_feature_select_strategy = (
+            vision_feature_select_strategy
+            if vision_feature_select_strategy is not None
+            else self.config.vision_feature_select_strategy
+        )
+        vision_aspect_ratio = (
+            vision_aspect_ratio if vision_aspect_ratio is not None else self.config.vision_aspect_ratio
+        )
+
+        # ! infer image_num_patches from image_sizes
+        if batch_num_images is None:
+            # treat this as a single-image case for backward compatibility
+            need_patching = [True] * len(image_sizes)
+        else:
+            need_patching = [n == 1 for n in batch_num_images for _ in range(n)]
+        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,
+            )
+            if should_patch
+            else 1
+            for imsize, should_patch in zip(image_sizes, need_patching)
+        ]
+        if pixel_values.dim() == 5:
+            # stacked if input is (batch_size, num_patches, num_channels, height, width)
+            _pixel_values_list = [pix_val[:num_patch] for pix_val, num_patch in zip(pixel_values, image_num_patches)]
+            pixel_values = torch.cat(_pixel_values_list, dim=0)
+        elif pixel_values.dim() != 4:
+            # otherwise has to be stacked from list of (num_patches, num_channels, height, width)
+            raise ValueError(f"pixel_values of shape {pixel_values.shape}, expect to be of 4 or 5 dimensions")
+
+        image_features = self.vision_tower(pixel_values, output_hidden_states=True)
+        # If we have one vision feature layer, return the corresponding hidden states,
+        # otherwise, select the hidden states of each feature layer and concatenate them
+        if isinstance(vision_feature_layer, int):
+            selected_image_feature = image_features.hidden_states[vision_feature_layer]
+        else:
+            hs_pool = [image_features.hidden_states[layer_idx] for layer_idx in vision_feature_layer]
+            selected_image_feature = torch.cat(hs_pool, dim=-1)
+
+        if vision_feature_select_strategy == "default":
+            selected_image_feature = selected_image_feature[:, 1:]
+        elif vision_feature_select_strategy == "full":
+            selected_image_feature = selected_image_feature
+        image_features = self.multi_modal_projector(selected_image_feature)
+        image_features = torch.split(image_features, image_num_patches, dim=0)
+
+        image_features, feature_lens = self.pack_image_features(
+            image_features,
+            image_sizes,
+            image_newline=self.image_newline,
+            vision_aspect_ratio=vision_aspect_ratio,
+        )
+        return image_features
+
+    @can_return_tuple
+    @auto_docstring
+    def forward(
+        self,
+        input_ids: torch.LongTensor = None,
+        pixel_values: torch.FloatTensor = None,
+        image_sizes: Optional[torch.LongTensor] = None,
+        pixel_values_videos: torch.FloatTensor = None,
+        image_sizes_videos: Optional[torch.LongTensor] = None,
+        attention_mask: Optional[torch.Tensor] = None,
+        position_ids: Optional[torch.LongTensor] = None,
+        past_key_values: Optional[Cache] = None,
+        inputs_embeds: Optional[torch.FloatTensor] = None,
+        vision_feature_layer: Optional[Union[int, list[int]]] = None,
+        vision_feature_select_strategy: Optional[str] = None,
+        vision_aspect_ratio: Optional[str] = None,
+        batch_num_images: Optional[torch.LongTensor] = None,
+        use_cache: Optional[bool] = None,
+        output_attentions: Optional[bool] = None,
+        output_hidden_states: Optional[bool] = None,
+        return_dict: Optional[bool] = None,
+        cache_position: Optional[torch.LongTensor] = None,
+        **kwargs: Unpack[FlashAttentionKwargs],
+    ) -> Union[tuple, LlavaOnevisionModelOutputWithPast]:
+        r"""
+        image_sizes_videos (`torch.LongTensor` of shape `(batch_size, frames, 2)`, *optional*):
+            The sizes of the videos in the batch, being (height, width) for each frame in the video.
+        vision_aspect_ratio (`str`, *optional*, defaults to `"anyres_max_9"`):
+            Aspect ratio used when processong image features. The default value is "anyres_max_9".
+        batch_num_images (`torch.LongTensor`, *optional*):
+            Number of images in each sample.
+        """
+
+        output_attentions = output_attentions if output_attentions is not None else self.config.output_attentions
+        output_hidden_states = (
+            output_hidden_states if output_hidden_states is not None else self.config.output_hidden_states
+        )
+        return_dict = return_dict if return_dict is not None else self.config.use_return_dict
+        vision_feature_layer = (
+            vision_feature_layer if vision_feature_layer is not None else self.config.vision_feature_layer
+        )
+        vision_feature_select_strategy = (
+            vision_feature_select_strategy
+            if vision_feature_select_strategy is not None
+            else self.config.vision_feature_select_strategy
+        )
+        vision_aspect_ratio = (
+            vision_aspect_ratio if vision_aspect_ratio is not None else self.config.vision_aspect_ratio
+        )
+
+        if (input_ids is None) ^ (inputs_embeds is not None):
+            raise ValueError("You must specify exactly one of input_ids or inputs_embeds")
+
+        if inputs_embeds is None:
+            inputs_embeds = self.get_input_embeddings()(input_ids)
+
+        # Images are processed with Anyres
+        if pixel_values is not None:
+            image_features = self.get_image_features(
+                pixel_values,
+                image_sizes,
+                vision_feature_layer=vision_feature_layer,
+                vision_feature_select_strategy=vision_feature_select_strategy,
+                batch_num_images=batch_num_images,
+            )
+            image_features = torch.cat(image_features, dim=0)
+
+            if input_ids is None:
+                special_image_mask = inputs_embeds == self.get_input_embeddings()(
+                    torch.tensor(self.config.image_token_id, dtype=torch.long, device=inputs_embeds.device)
+                )
+                special_image_mask = special_image_mask.all(-1)
+            else:
+                special_image_mask = input_ids == self.config.image_token_id
+
+            n_image_tokens = (special_image_mask).sum()
+            special_image_mask = special_image_mask.unsqueeze(-1).expand_as(inputs_embeds).to(inputs_embeds.device)
+
+            if not is_torchdynamo_compiling() and inputs_embeds[special_image_mask].numel() != image_features.numel():
+                n_image_features = image_features.shape[0]
+                raise ValueError(
+                    f"Image features and image tokens do not match: tokens: {n_image_tokens}, features {n_image_features}"
+                )
+            image_features = image_features.to(inputs_embeds.device, inputs_embeds.dtype)
+            inputs_embeds = inputs_embeds.masked_scatter(special_image_mask, image_features)
+
+        # Video are simply embedded and further pooled to decrease seq len
+        if pixel_values_videos is not None:
+            video_features = self.get_video_features(
+                pixel_values_videos,
+                vision_feature_layer=vision_feature_layer,
+                vision_feature_select_strategy=vision_feature_select_strategy,
+            )
+            image_newline = (
+                self.image_newline[None, None, :].repeat(video_features.shape[0], 1, 1).to(video_features.device)
+            )
+            video_features = torch.cat((video_features, image_newline), dim=1)
+            video_features = video_features.flatten(0, 1)
+
+            if input_ids is None:
+                special_video_mask = inputs_embeds == self.get_input_embeddings()(
+                    torch.tensor(self.config.video_token_id, dtype=torch.long, device=inputs_embeds.device)
+                )
+                special_video_mask = special_video_mask.all(-1)
+            else:
+                special_video_mask = input_ids == self.config.video_token_id
+
+            n_video_tokens = (special_video_mask).sum()
+            special_video_mask = special_video_mask.unsqueeze(-1).expand_as(inputs_embeds).to(inputs_embeds.device)
+
+            if not is_torchdynamo_compiling() and inputs_embeds[special_video_mask].numel() != video_features.numel():
+                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}"
+                )
+            video_features = video_features.to(inputs_embeds.device, inputs_embeds.dtype)
+            inputs_embeds = inputs_embeds.masked_scatter(special_video_mask, video_features)
+
+        outputs = self.language_model(
+            attention_mask=attention_mask,
+            position_ids=position_ids,
+            past_key_values=past_key_values,
+            inputs_embeds=inputs_embeds,
+            use_cache=use_cache,
+            output_attentions=output_attentions,
+            output_hidden_states=output_hidden_states,
+            return_dict=True,
+            cache_position=cache_position,
+            **kwargs,
+        )
+
+        return LlavaOnevisionModelOutputWithPast(
+            last_hidden_state=outputs.last_hidden_state,
+            past_key_values=outputs.past_key_values,
+            hidden_states=outputs.hidden_states,
+            attentions=outputs.attentions,
+            image_hidden_states=image_features if pixel_values is not None else None,
+            video_hidden_states=video_features if pixel_values_videos is not None else None,
+        )
+
+    def get_video_features(
+        self,
+        pixel_values: torch.FloatTensor,
+        vision_feature_layer: Union[int, list[int]],
+        vision_feature_select_strategy: str,
+    ):
+        """
+        Obtains video last hidden states from the vision tower, apply multimodal projection and pooling.
+
+        Args:
+            pixel_values (`torch.FloatTensor]` of shape `(batch_size, num_frames, channels, height, width)`)
+               The tensors corresponding to the input video.
+            vision_feature_layer (`Union[int, list[int]], *optional*, defaults to -2`):
+                The index of the layer to select the vision feature. If multiple indices are provided,
+                the vision feature of the corresponding indices will be concatenated to form the
+                vision features.
+            vision_feature_select_strategy (`str`):
+                The feature selection strategy used to select the vision feature from the vision backbone.
+                Can be one of `"default"` or `"full"`
+        Returns:
+            video_features (list[`torch.Tensor`]): List of video feature tensor, each contains all the visual feature of all patches
+            and are of shape `(num_videos, video_length, embed_dim)`).
+        """
+        batch_size, frames, channels, height, width = pixel_values.shape
+        pixel_values = pixel_values.view(batch_size * frames, channels, height, width)
+        video_features = self.vision_tower(pixel_values, output_hidden_states=True)
+
+        # If we have one vision feature layer, return the corresponding hidden states,
+        # otherwise, select the hidden states of each feature layer and concatenate them
+        if isinstance(vision_feature_layer, int):
+            selected_video_feature = video_features.hidden_states[vision_feature_layer]
+        else:
+            hs_pool = [video_features.hidden_states[layer_idx] for layer_idx in vision_feature_layer]
+            selected_video_feature = torch.cat(hs_pool, dim=-1)
+
+        if vision_feature_select_strategy == "default":
+            selected_video_feature = selected_video_feature[:, 1:]
+        elif vision_feature_select_strategy == "full":
+            selected_video_feature = selected_video_feature
+        video_features = self.multi_modal_projector(selected_video_feature)
+
+        video_features = self.apply_pooling(video_features)
+        video_features = video_features.reshape(batch_size, frames * video_features.shape[1], -1)
+
+        return video_features
+
+    def apply_pooling(self, image_features):
+        height = width = self.config.vision_config.image_size // self.config.vision_config.patch_size
+        batch_frames, seq_len, dim = image_features.shape
+        image_features = image_features.view(batch_frames, height, width, -1)
+        image_features = image_features.permute(0, 3, 1, 2).contiguous()
+
+        height, width = image_features.shape[2:]
+        scaled_shape = [math.ceil(height / 2), math.ceil(width / 2)]
+        image_features = nn.functional.interpolate(image_features, size=scaled_shape, mode="bilinear")
+
+        image_features = image_features.permute(0, 2, 3, 1)
+        image_features = image_features.view(batch_frames, -1, dim)
+        return image_features
+
+
+@auto_docstring(
+    custom_intro="""
+    The LLAVA-NeXT model which consists of a vision backbone and a language model.
+    """
+)
+class LlavaOnevisionForConditionalGeneration(LlavaOnevisionPreTrainedModel, GenerationMixin):
+    _checkpoint_conversion_mapping = {
+        "^language_model.model": "model.language_model",
+        "^vision_tower": "model.vision_tower",
+        "^multi_modal_projector": "model.multi_modal_projector",
+        "^image_newline": "model.image_newline",
+        "^language_model.lm_head": "lm_head",
+    }
+    _tied_weights_keys = ["lm_head.weight"]
+
+    def __init__(self, config: LlavaOnevisionConfig):
+        super().__init__(config)
+        self.model = LlavaOnevisionModel(config)
+        self.lm_head = nn.Linear(config.text_config.hidden_size, config.text_config.vocab_size, bias=False)
+        self.post_init()
+
+    def get_input_embeddings(self):
+        return self.model.get_input_embeddings()
+
+    def set_input_embeddings(self, value):
+        self.model.set_input_embeddings(value)
+
+    def get_output_embeddings(self) -> nn.Module:
+        return self.lm_head
+
+    def set_decoder(self, decoder):
+        self.model.set_decoder(decoder)
+
+    def get_decoder(self):
+        return self.model.get_decoder()
+
+    def pack_image_features(self, image_features, image_sizes, vision_feature_select_strategy, image_newline=None):
+        return self.model.pack_image_features(
+            image_features=image_features,
+            image_sizes=image_sizes,
+            vision_feature_select_strategy=vision_feature_select_strategy,
+            image_newline=image_newline,
+        )
+
+    def get_image_features(
+        self,
+        pixel_values: torch.FloatTensor,
+        image_sizes: torch.Tensor,
+        vision_feature_layer: Optional[Union[int, list[int]]] = None,
+        vision_feature_select_strategy: Optional[str] = None,
+    ):
+        return self.model.get_image_features(
+            pixel_values=pixel_values,
+            image_sizes=image_sizes,
+            vision_feature_layer=vision_feature_layer,
+            vision_feature_select_strategy=vision_feature_select_strategy,
+        )
+
+    # Make modules available throught conditional class for BC
+    @property
+    def language_model(self):
+        return self.model.language_model
+
+    @property
+    def vision_tower(self):
+        return self.model.vision_tower
+
+    @property
+    def multi_modal_projector(self):
+        return self.model.multi_modal_projector
+
+    @can_return_tuple
+    @auto_docstring
+    def forward(
+        self,
+        input_ids: torch.LongTensor = None,
+        pixel_values: torch.FloatTensor = None,
+        image_sizes: Optional[torch.LongTensor] = None,
+        pixel_values_videos: torch.FloatTensor = None,
+        image_sizes_videos: Optional[torch.LongTensor] = None,
+        attention_mask: Optional[torch.Tensor] = None,
+        position_ids: Optional[torch.LongTensor] = None,
+        past_key_values: Optional[Cache] = None,
+        inputs_embeds: Optional[torch.FloatTensor] = None,
+        vision_feature_layer: Optional[Union[int, list[int]]] = None,
+        vision_feature_select_strategy: Optional[str] = None,
+        vision_aspect_ratio: Optional[str] = None,
+        batch_num_images: Optional[torch.LongTensor] = None,
+        labels: Optional[torch.LongTensor] = None,
+        use_cache: Optional[bool] = None,
+        output_attentions: Optional[bool] = None,
+        output_hidden_states: Optional[bool] = None,
+        return_dict: Optional[bool] = None,
+        cache_position: Optional[torch.LongTensor] = None,
+        logits_to_keep: Union[int, torch.Tensor] = 0,
+        **kwargs: Unpack[TransformersKwargs],
+    ) -> Union[tuple, LlavaOnevisionCausalLMOutputWithPast]:
+        r"""
+        image_sizes_videos (`torch.LongTensor` of shape `(batch_size, frames, 2)`, *optional*):
+            The sizes of the videos in the batch, being (height, width) for each frame in the video.
+        vision_aspect_ratio (`str`, *optional*, defaults to `"anyres_max_9"`):
+            Aspect ratio used when processong image features. The default value is "anyres_max_9".
+        batch_num_images (`torch.LongTensor`, *optional*):
+            Number of images in each sample.
+        labels (`torch.LongTensor` of shape `(batch_size, sequence_length)`, *optional*):
+            Labels for computing the masked language modeling loss. Indices should either be in `[0, ...,
+            config.vocab_size]` or -100 (see `input_ids` docstring). Tokens with indices set to `-100` are ignored
+            (masked), the loss is only computed for the tokens with labels in `[0, ..., config.vocab_size]`.
+
+        Example:
+
+        ```python
+        >>> from PIL import Image
+        >>> import requests
+        >>> import torch
+        >>> from transformers import LlavaOnevisionProcessor, LlavaOnevisionForConditionalGeneration
+
+        >>> model = LlavaOnevisionForConditionalGeneration.from_pretrained("llava-hf/llava-onevision-qwen2-7b-ov-hf", torch_dtype="float16", device_map="cuda:0")
+        >>> processor = LlavaOnevisionProcessor.from_pretrained("llava-hf/llava-onevision-qwen2-7b-ov-hf")
+
+        >>> conversation = [
+        ...     {
+        ...       "role": "user",
+        ...       "content": [
+        ...           {"type": "text", "text": "What is shown in this image?"},
+        ...           {"type": "image"},
+        ...         ],
+        ...     },
+        ... ]
+        >>> prompt = processor.apply_chat_template(conversation, add_generation_prompt=True)
+
+        >>> image_file = "http://images.cocodataset.org/val2017/000000039769.jpg"
+        >>> raw_image = Image.open(requests.get(image_file, stream=True).raw)
+        >>> inputs = processor(text=prompt, images=raw_image, return_tensors='pt').to(0, torch.float16)
+
+        >>> output = model.generate(**inputs, max_new_tokens=20, do_sample=False)
+        >>> processor.batch_decode(output, skip_special_tokens=True)[0]
+        "user\n\nWhat is shown in this image?\nassistant\ncat"
+        ```"""
+        output_attentions = output_attentions if output_attentions is not None else self.config.output_attentions
+        output_hidden_states = (
+            output_hidden_states if output_hidden_states is not None else self.config.output_hidden_states
+        )
+        return_dict = return_dict if return_dict is not None else self.config.use_return_dict
+        vision_feature_layer = (
+            vision_feature_layer if vision_feature_layer is not None else self.config.vision_feature_layer
+        )
+        vision_feature_select_strategy = (
+            vision_feature_select_strategy
+            if vision_feature_select_strategy is not None
+            else self.config.vision_feature_select_strategy
+        )
+        vision_aspect_ratio = (
+            vision_aspect_ratio if vision_aspect_ratio is not None else self.config.vision_aspect_ratio
+        )
+
+        outputs = self.model(
+            input_ids=input_ids,
+            pixel_values=pixel_values,
+            pixel_values_videos=pixel_values_videos,
+            image_sizes=image_sizes,
+            image_sizes_videos=image_sizes_videos,
+            vision_aspect_ratio=vision_aspect_ratio,
+            vision_feature_layer=vision_feature_layer,
+            vision_feature_select_strategy=vision_feature_select_strategy,
+            batch_num_images=batch_num_images,
+            attention_mask=attention_mask,
+            position_ids=position_ids,
+            past_key_values=past_key_values,
+            inputs_embeds=inputs_embeds,
+            use_cache=use_cache,
+            output_attentions=output_attentions,
+            output_hidden_states=output_hidden_states,
+            return_dict=True,
+            cache_position=cache_position,
+            logits_to_keep=logits_to_keep,
+            **kwargs,
+        )
+
+        hidden_states = outputs[0]
+        # Only compute necessary logits, and do not upcast them to float if we are not computing the loss
+        slice_indices = slice(-logits_to_keep, None) if isinstance(logits_to_keep, int) else logits_to_keep
+        logits = self.lm_head(hidden_states[:, slice_indices, :])
+
+        loss = None
+        if labels is not None:
+            loss = self.loss_function(
+                logits=logits, labels=labels, vocab_size=self.config.text_config.vocab_size, **kwargs
+            )
+
+        return LlavaOnevisionCausalLMOutputWithPast(
+            loss=loss,
+            logits=logits,
+            past_key_values=outputs.past_key_values,
+            hidden_states=outputs.hidden_states,
+            attentions=outputs.attentions,
+            image_hidden_states=outputs.image_hidden_states,
+            video_hidden_states=outputs.video_hidden_states,
+        )
+
+    def prepare_inputs_for_generation(
+        self,
+        input_ids,
+        past_key_values=None,
+        inputs_embeds=None,
+        pixel_values=None,
+        image_sizes=None,
+        pixel_values_videos=None,
+        image_sizes_videos=None,
+        attention_mask=None,
+        cache_position=None,
+        logits_to_keep=None,
+        **kwargs,
+    ):
+        # Overwritten -- in specific circumstances we don't want to forward image inputs to the model
+
+        model_inputs = super().prepare_inputs_for_generation(
+            input_ids,
+            past_key_values=past_key_values,
+            inputs_embeds=inputs_embeds,
+            attention_mask=attention_mask,
+            cache_position=cache_position,
+            logits_to_keep=logits_to_keep,
+            **kwargs,
+        )
+
+        if cache_position[0] == 0:
+            # If we're in cached decoding stage, pixel values should be None because input ids do not contain special image token anymore
+            # Otherwise we need pixel values to be passed to model
+            model_inputs["pixel_values"] = pixel_values
+            model_inputs["image_sizes"] = image_sizes
+            model_inputs["pixel_values_videos"] = pixel_values_videos
+            model_inputs["image_sizes_videos"] = image_sizes_videos
+
+        return model_inputs
+
+    @staticmethod
+    def _prepare_4d_causal_attention_mask_with_cache_position(
+        attention_mask: torch.Tensor,
+        sequence_length: int,
+        target_length: int,
+        dtype: torch.dtype,
+        cache_position: torch.Tensor,
+        batch_size: int,
+        **kwargs,
+    ):
+        """
+        Creates a causal 4D mask of shape `(batch_size, 1, query_length, key_value_length)` from a 2D mask of shape
+        `(batch_size, key_value_length)`, or if the input `attention_mask` is already 4D, do nothing.
+
+        Args:
+            attention_mask (`torch.Tensor`):
+                A 2D attention mask of shape `(batch_size, key_value_length)` or a 4D attention mask of shape
+                `(batch_size, 1, query_length, key_value_length)`.
+            sequence_length (`int`):
+                The sequence length being processed.
+            target_length (`int`):
+                The target length: when generating with static cache, the mask should be as long as the static cache,
+                to account for the 0 padding, the part of the cache that is not filled yet.
+            dtype (`torch.dtype`):
+                The dtype to use for the 4D attention mask.
+            cache_position (`torch.Tensor`):
+                Indices depicting the position of the input sequence tokens in the sequence.
+            batch_size (`torch.Tensor`):
+                Batch size.
+        """
+        if attention_mask is not None and attention_mask.dim() == 4:
+            # In this case we assume that the mask comes already in inverted form and requires no inversion or slicing.
+            causal_mask = attention_mask
+        else:
+            min_dtype = torch.finfo(dtype).min
+            causal_mask = torch.full(
+                (sequence_length, target_length), fill_value=min_dtype, dtype=dtype, device=cache_position.device
+            )
+            if sequence_length != 1:
+                causal_mask = torch.triu(causal_mask, diagonal=1)
+            causal_mask *= torch.arange(target_length, device=cache_position.device) > cache_position.reshape(-1, 1)
+            causal_mask = causal_mask[None, None, :, :].expand(batch_size, 1, -1, -1)
+            if attention_mask is not None:
+                causal_mask = causal_mask.clone()  # copy to contiguous memory for in-place edit
+                mask_length = attention_mask.shape[-1]
+                padding_mask = causal_mask[:, :, :, :mask_length] + attention_mask[:, None, None, :].to(
+                    causal_mask.device
+                )
+                padding_mask = padding_mask == 0
+                causal_mask[:, :, :, :mask_length] = causal_mask[:, :, :, :mask_length].masked_fill(
+                    padding_mask, min_dtype
+                )
+
+        return causal_mask
+
+    def get_video_features(
+        self,
+        pixel_values: torch.FloatTensor,
+        vision_feature_layer: Optional[Union[int, list[int]]] = None,
+        vision_feature_select_strategy: Optional[str] = None,
+    ):
+        return self.model.get_video_features(
+            pixel_values=pixel_values,
+            vision_feature_layer=vision_feature_layer,
+            vision_feature_select_strategy=vision_feature_select_strategy,
+        )
+
+
+__all__ = ["LlavaOnevisionModel", "LlavaOnevisionForConditionalGeneration", "LlavaOnevisionPreTrainedModel"]

venv/Lib/site-packages/transformers/models/llava_onevision/modular_llava_onevision.py

@@ -0,0 +1,791 @@
+# coding=utf-8
+# Copyright 2024 the HuggingFace Inc. team. All rights reserved.
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+
+import math
+from typing import Optional, Union
+
+import torch
+from torch import nn
+
+from transformers.models.llava_next.image_processing_llava_next_fast import LlavaNextImageProcessorFast
+from transformers.models.llava_next_video.modeling_llava_next_video import (
+    LlavaNextVideoCausalLMOutputWithPast,
+    LlavaNextVideoForConditionalGeneration,
+    LlavaNextVideoModel,
+    LlavaNextVideoModelOutputWithPast,
+    LlavaNextVideoPreTrainedModel,
+    TransformersKwargs,
+    get_anyres_image_grid_shape,
+    image_size_to_num_patches,
+    unpad_image,
+)
+
+from ...cache_utils import Cache
+from ...image_processing_utils import BatchFeature
+from ...image_processing_utils_fast import DefaultFastImageProcessorKwargs, group_images_by_shape, reorder_images
+from ...image_utils import (
+    OPENAI_CLIP_MEAN,
+    OPENAI_CLIP_STD,
+    ChannelDimension,
+    ImageInput,
+    PILImageResampling,
+    SizeDict,
+    get_image_size,
+)
+from ...modeling_flash_attention_utils import FlashAttentionKwargs
+from ...processing_utils import Unpack
+from ...utils import (
+    TensorType,
+    auto_docstring,
+    can_return_tuple,
+    is_torchdynamo_compiling,
+    is_torchvision_available,
+    is_torchvision_v2_available,
+    logging,
+)
+
+
+if is_torchvision_available():
+    if is_torchvision_v2_available():
+        from torchvision.transforms.v2 import functional as F
+    else:
+        from torchvision.transforms import functional as F
+
+
+logger = logging.get_logger(__name__)
+
+
+class LlavaOnevisionFastImageProcessorKwargs(DefaultFastImageProcessorKwargs):
+    """
+    image_grid_pinpoints (`list[list[int]]`, *optional*):
+        A list of possible resolutions to use for processing high resolution images. The best resolution is selected
+        based on the original size of the image. Can be overridden by `image_grid_pinpoints` in the `preprocess`
+        method.
+    do_pad (`bool`, *optional*):
+        Whether to pad the image. If `True`, will pad the patch dimension of the images in the batch to the largest
+        number of patches in the batch. Padding will be applied to the bottom and right with zeros.
+    """
+
+    image_grid_pinpoints: Optional[list[list[int]]]
+    do_pad: Optional[bool]
+
+
+class LlavaOnevisionImageProcessorFast(LlavaNextImageProcessorFast):
+    resample = PILImageResampling.BICUBIC
+    image_mean = OPENAI_CLIP_MEAN
+    image_std = OPENAI_CLIP_STD
+    size = {"height": 384, "width": 384}
+    crop_size = None
+    default_to_square = False
+    do_resize = True
+    do_center_crop = None
+    do_rescale = True
+    do_normalize = True
+    do_convert_rgb = True
+    do_pad = True
+    image_grid_pinpoints = [[384, 384], [384, 768], [384, 1152], [384, 1536], [384, 1920], [384, 2304], [768, 384], [768, 768], [768, 1152], [768, 1536], [768, 1920], [768, 2304], [1152, 384], [1152, 768], [1152, 1152], [1152, 1536], [1152, 1920], [1152, 2304], [1536, 384], [1536, 768], [1536, 1152], [1536, 1536], [1536, 1920], [1536, 2304], [1920, 384], [1920, 768], [1920, 1152], [1920, 1536], [1920, 1920], [1920, 2304], [2304, 384], [2304, 768], [2304, 1152], [2304, 1536], [2304, 1920], [2304, 2304]]  # fmt: skip
+    model_input_names = ["pixel_values_videos"]
+
+    # Copied from transformers.models.llava.image_processing_llava_fast.LlavaImageProcessorFast.pad_to_square
+    def pad_to_square(
+        self,
+        images: "torch.Tensor",
+        background_color: Union[int, tuple[int, int, int]] = 0,
+    ) -> "torch.Tensor":
+        """
+        Pads an image to a square based on the longest edge.
+
+        Args:
+            images (`np.ndarray`):
+                The images to pad.
+            background_color (`int` or `tuple[int, int, int]`, *optional*, defaults to 0):
+                The color to use for the padding. Can be an integer for single channel or a
+                tuple of integers representing for multi-channel images. If passed as integer
+                in mutli-channel mode, it will default to `0` in subsequent channels.
+        Returns:
+            `torch.Tensor`: The padded images.
+        """
+        height, width = get_image_size(images, ChannelDimension.FIRST)
+
+        if height == width:
+            return images
+
+        num_channels = images.shape[1] if len(images.shape) == 4 else images.shape[0]
+        if isinstance(background_color, int):
+            background_color = [background_color] + [0] * (num_channels - 1)
+        elif len(background_color) != num_channels:
+            raise ValueError(
+                f"background_color must have no more than {num_channels} elements to match the number of channels"
+            )
+
+        max_dim = max(height, width)
+        paste_x_left = (max_dim - width) // 2
+        paste_y_left = (max_dim - height) // 2
+        paste_x_right = max_dim - width - paste_x_left
+        paste_y_right = max_dim - height - paste_y_left
+        padded_images = F.pad(
+            images, padding=[paste_x_left, paste_y_left, paste_x_right, paste_y_right], fill=background_color
+        )
+
+        return padded_images
+
+    @auto_docstring
+    def preprocess(self, images: ImageInput, **kwargs: Unpack[LlavaOnevisionFastImageProcessorKwargs]) -> BatchFeature:
+        if isinstance(images, (tuple, list)) and isinstance(images[0], (tuple, list)):
+            # if the first element is a list, we assume that all elements are lists
+            batch_num_images = [len(x) for x in images]
+        elif isinstance(images, (tuple, list)):
+            # treat this as a single-image case for backward compatibility
+            batch_num_images = [1] * len(images)
+        else:
+            batch_num_images = [1]
+        kwargs["batch_num_images"] = batch_num_images
+        return super().preprocess(images, **kwargs)
+
+    def _preprocess(
+        self,
+        images: list["torch.Tensor"],
+        do_resize: bool,
+        size: SizeDict,
+        image_grid_pinpoints: list[list[int]],
+        interpolation: Optional["F.InterpolationMode"],
+        do_center_crop: bool,
+        crop_size: SizeDict,
+        do_rescale: bool,
+        rescale_factor: float,
+        do_normalize: bool,
+        image_mean: Optional[Union[float, list[float]]],
+        image_std: Optional[Union[float, list[float]]],
+        do_pad: bool,
+        batch_num_images: list[int],
+        disable_grouping: Optional[bool],
+        return_tensors: Optional[Union[str, TensorType]],
+        **kwargs,
+    ) -> BatchFeature:
+        processed_images = []
+        image_sizes = []
+
+        # only single image patching is supported
+        need_patching = [n == 1 for n in batch_num_images for _ in range(n)]
+
+        # Determine the size tuple
+        if size and size.height and size.width:
+            size_tuple = (size.height, size.width)
+        else:
+            size_tuple = (size.shortest_edge, size.shortest_edge)
+
+        # Determine the patch size
+        if crop_size and crop_size.height:
+            patch_size = crop_size.height
+        elif size and size.height:
+            patch_size = size.height
+        else:
+            patch_size = size.shortest_edge
+
+        for i, image in enumerate(images):
+            if need_patching[i]:
+                image_patches = self._get_image_patches(
+                    image,
+                    image_grid_pinpoints,
+                    size=size_tuple,
+                    patch_size=patch_size,
+                    interpolation=interpolation,
+                )
+            else:
+                padded_image = self.pad_to_square(
+                    images=image, background_color=tuple(int(x * 255) for x in self.image_mean)
+                )
+                image_patches = [padded_image]
+
+            # Group images by size for batched processing
+            processed_image_patches_grouped = {}
+            grouped_image_patches, grouped_image_patches_index = group_images_by_shape(
+                image_patches, disable_grouping=disable_grouping
+            )
+            for shape, stacked_image_patches in grouped_image_patches.items():
+                if do_resize:
+                    stacked_image_patches = self.resize(
+                        image=stacked_image_patches,
+                        size=size,
+                        interpolation=interpolation,
+                    )
+                if do_center_crop:
+                    stacked_image_patches = self.center_crop(stacked_image_patches, crop_size)
+                # Fused rescale and normalize
+                stacked_image_patches = self.rescale_and_normalize(
+                    stacked_image_patches, do_rescale, rescale_factor, do_normalize, image_mean, image_std
+                )
+                processed_image_patches_grouped[shape] = stacked_image_patches
+            processed_image_patches = reorder_images(processed_image_patches_grouped, grouped_image_patches_index)
+            processed_image_patches = (
+                torch.stack(processed_image_patches, dim=0) if return_tensors else processed_image_patches
+            )
+            processed_images.append(processed_image_patches)
+            image_sizes.append(get_image_size(image, ChannelDimension.FIRST))
+
+        if do_pad:
+            processed_images = self._pad_for_batching(processed_images)
+        processed_images = torch.stack(processed_images, dim=0) if return_tensors else processed_images
+        return BatchFeature(
+            data={"pixel_values": processed_images, "image_sizes": image_sizes, "batch_num_images": batch_num_images},
+            tensor_type=return_tensors,
+        )
+
+
+class LlavaOnevisionModelOutputWithPast(LlavaNextVideoModelOutputWithPast):
+    pass
+
+
+class LlavaOnevisionCausalLMOutputWithPast(LlavaNextVideoCausalLMOutputWithPast):
+    pass
+
+
+class LlavaOnevisionPreTrainedModel(LlavaNextVideoPreTrainedModel):
+    pass
+
+
+class LlavaOnevisionModel(LlavaNextVideoModel):
+    def __init__(self, config):
+        super().__init__(config)
+        del self.vision_resampler
+
+    def pack_image_features(self, image_features, image_sizes, image_newline=None, vision_aspect_ratio="anyres_max_9"):
+        """
+        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.
+            vision_aspect_ratio (`str`, *optional*, "anyres_max_9"):
+                Aspect ratio used when processong image features. The default value is "anyres_max_9".
+        Returns:
+            image_features (`torch.Tensor` of shape `(all_feat_len, embed_dim)`)
+            feature_lens (`list[int]`)
+                token length of each image in image_features
+        """
+        new_image_features = []
+        feature_lens = []
+        for image_idx, image_feature in enumerate(image_features):
+            if image_feature.shape[0] > 1:
+                base_image_feature = image_feature[0]
+                image_feature = image_feature[1:]
+                height = width = self.config.vision_config.image_size // self.config.vision_config.patch_size
+                if height * width != base_image_feature.shape[0]:
+                    raise ValueError("The number of patches is not consistent with the image size.")
+                num_patch_height, num_patch_width = get_anyres_image_grid_shape(
+                    image_sizes[image_idx],
+                    self.config.image_grid_pinpoints,
+                    self.config.vision_config.image_size,
+                )
+                image_feature = image_feature.view(num_patch_height, num_patch_width, height, width, -1)
+                image_feature = image_feature.permute(4, 0, 2, 1, 3).contiguous()
+                image_feature = image_feature.flatten(1, 2).flatten(2, 3)
+                image_feature = unpad_image(image_feature, image_sizes[image_idx])
+                max_num_patches = int(vision_aspect_ratio.strip("anyres_max_"))
+                channels, curr_height, curr_width = image_feature.shape
+                ratio = math.sqrt(curr_height * curr_width / (max_num_patches * height**2))
+                if ratio > 1.1:
+                    image_feature = image_feature[None]
+                    image_feature = nn.functional.interpolate(
+                        image_feature, [int(curr_height // ratio), int(curr_width // ratio)], mode="bilinear"
+                    )[0]
+                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)
+                image_feature = image_feature.flatten(0, 1)
+            new_image_features.append(image_feature)
+            feature_lens.append(image_feature.size(0))
+        feature_lens = torch.tensor(feature_lens, dtype=torch.long, device=image_features[0].device)
+        return new_image_features, feature_lens
+
+    def apply_pooling(self, image_features):
+        height = width = self.config.vision_config.image_size // self.config.vision_config.patch_size
+        batch_frames, seq_len, dim = image_features.shape
+        image_features = image_features.view(batch_frames, height, width, -1)
+        image_features = image_features.permute(0, 3, 1, 2).contiguous()
+
+        height, width = image_features.shape[2:]
+        scaled_shape = [math.ceil(height / 2), math.ceil(width / 2)]
+        image_features = nn.functional.interpolate(image_features, size=scaled_shape, mode="bilinear")
+
+        image_features = image_features.permute(0, 2, 3, 1)
+        image_features = image_features.view(batch_frames, -1, dim)
+        return image_features
+
+    def get_image_features(
+        self,
+        pixel_values: torch.FloatTensor,
+        image_sizes: torch.Tensor,
+        vision_feature_layer: Optional[Union[int, list[int]]] = None,
+        vision_feature_select_strategy: Optional[str] = None,
+        vision_aspect_ratio: Optional[str] = None,
+        batch_num_images: Optional[torch.LongTensor] = None,
+    ):
+        """
+        Obtains image last hidden states from the vision tower and apply multimodal projection.
+
+        Args:
+            pixel_values (`torch.FloatTensor]` of shape `(batch_size, num_patches, channels, height, width)`)
+               The tensors corresponding to the input images.
+            image_sizes (`torch.Tensor` of shape `(num_images, 2)`)
+                Actual image size of each images (H, W).
+            vision_feature_layer (`Union[int, list[int]]`):
+                The index of the layer to select the vision feature. If multiple indices are provided,
+                the vision feature of the corresponding indices will be concatenated to form the
+                vision features.
+            vision_feature_select_strategy (`str`):
+                The feature selection strategy used to select the vision feature from the vision backbone.
+                Can be one of `"default"` or `"full"`
+            batch_num_images (`torch.LongTensor`, *optional*):
+                Number of images in each sample.
+        Returns:
+            image_features (list[`torch.Tensor`]): List of image feature tensor, each contains all the visual feature of all patches
+            and are of shape `(num_patches, image_length, embed_dim)`).
+        """
+        vision_feature_layer = (
+            vision_feature_layer if vision_feature_layer is not None else self.config.vision_feature_layer
+        )
+        vision_feature_select_strategy = (
+            vision_feature_select_strategy
+            if vision_feature_select_strategy is not None
+            else self.config.vision_feature_select_strategy
+        )
+        vision_aspect_ratio = (
+            vision_aspect_ratio if vision_aspect_ratio is not None else self.config.vision_aspect_ratio
+        )
+
+        # ! infer image_num_patches from image_sizes
+        if batch_num_images is None:
+            # treat this as a single-image case for backward compatibility
+            need_patching = [True] * len(image_sizes)
+        else:
+            need_patching = [n == 1 for n in batch_num_images for _ in range(n)]
+        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,
+            )
+            if should_patch
+            else 1
+            for imsize, should_patch in zip(image_sizes, need_patching)
+        ]
+        if pixel_values.dim() == 5:
+            # stacked if input is (batch_size, num_patches, num_channels, height, width)
+            _pixel_values_list = [pix_val[:num_patch] for pix_val, num_patch in zip(pixel_values, image_num_patches)]
+            pixel_values = torch.cat(_pixel_values_list, dim=0)
+        elif pixel_values.dim() != 4:
+            # otherwise has to be stacked from list of (num_patches, num_channels, height, width)
+            raise ValueError(f"pixel_values of shape {pixel_values.shape}, expect to be of 4 or 5 dimensions")
+
+        image_features = self.vision_tower(pixel_values, output_hidden_states=True)
+        # If we have one vision feature layer, return the corresponding hidden states,
+        # otherwise, select the hidden states of each feature layer and concatenate them
+        if isinstance(vision_feature_layer, int):
+            selected_image_feature = image_features.hidden_states[vision_feature_layer]
+        else:
+            hs_pool = [image_features.hidden_states[layer_idx] for layer_idx in vision_feature_layer]
+            selected_image_feature = torch.cat(hs_pool, dim=-1)
+
+        if vision_feature_select_strategy == "default":
+            selected_image_feature = selected_image_feature[:, 1:]
+        elif vision_feature_select_strategy == "full":
+            selected_image_feature = selected_image_feature
+        image_features = self.multi_modal_projector(selected_image_feature)
+        image_features = torch.split(image_features, image_num_patches, dim=0)
+
+        image_features, feature_lens = self.pack_image_features(
+            image_features,
+            image_sizes,
+            image_newline=self.image_newline,
+            vision_aspect_ratio=vision_aspect_ratio,
+        )
+        return image_features
+
+    def get_video_features(
+        self,
+        pixel_values: torch.FloatTensor,
+        vision_feature_layer: Union[int, list[int]],
+        vision_feature_select_strategy: str,
+    ):
+        """
+        Obtains video last hidden states from the vision tower, apply multimodal projection and pooling.
+
+        Args:
+            pixel_values (`torch.FloatTensor]` of shape `(batch_size, num_frames, channels, height, width)`)
+               The tensors corresponding to the input video.
+            vision_feature_layer (`Union[int, list[int]], *optional*, defaults to -2`):
+                The index of the layer to select the vision feature. If multiple indices are provided,
+                the vision feature of the corresponding indices will be concatenated to form the
+                vision features.
+            vision_feature_select_strategy (`str`):
+                The feature selection strategy used to select the vision feature from the vision backbone.
+                Can be one of `"default"` or `"full"`
+        Returns:
+            video_features (list[`torch.Tensor`]): List of video feature tensor, each contains all the visual feature of all patches
+            and are of shape `(num_videos, video_length, embed_dim)`).
+        """
+        batch_size, frames, channels, height, width = pixel_values.shape
+        pixel_values = pixel_values.view(batch_size * frames, channels, height, width)
+        video_features = self.vision_tower(pixel_values, output_hidden_states=True)
+
+        # If we have one vision feature layer, return the corresponding hidden states,
+        # otherwise, select the hidden states of each feature layer and concatenate them
+        if isinstance(vision_feature_layer, int):
+            selected_video_feature = video_features.hidden_states[vision_feature_layer]
+        else:
+            hs_pool = [video_features.hidden_states[layer_idx] for layer_idx in vision_feature_layer]
+            selected_video_feature = torch.cat(hs_pool, dim=-1)
+
+        if vision_feature_select_strategy == "default":
+            selected_video_feature = selected_video_feature[:, 1:]
+        elif vision_feature_select_strategy == "full":
+            selected_video_feature = selected_video_feature
+        video_features = self.multi_modal_projector(selected_video_feature)
+
+        video_features = self.apply_pooling(video_features)
+        video_features = video_features.reshape(batch_size, frames * video_features.shape[1], -1)
+
+        return video_features
+
+    def forward(
+        self,
+        input_ids: torch.LongTensor = None,
+        pixel_values: torch.FloatTensor = None,
+        image_sizes: Optional[torch.LongTensor] = None,
+        pixel_values_videos: torch.FloatTensor = None,
+        image_sizes_videos: Optional[torch.LongTensor] = None,
+        attention_mask: Optional[torch.Tensor] = None,
+        position_ids: Optional[torch.LongTensor] = None,
+        past_key_values: Optional[Cache] = None,
+        inputs_embeds: Optional[torch.FloatTensor] = None,
+        vision_feature_layer: Optional[Union[int, list[int]]] = None,
+        vision_feature_select_strategy: Optional[str] = None,
+        vision_aspect_ratio: Optional[str] = None,
+        batch_num_images: Optional[torch.LongTensor] = None,
+        use_cache: Optional[bool] = None,
+        output_attentions: Optional[bool] = None,
+        output_hidden_states: Optional[bool] = None,
+        return_dict: Optional[bool] = None,
+        cache_position: Optional[torch.LongTensor] = None,
+        **kwargs: Unpack[FlashAttentionKwargs],
+    ) -> Union[tuple, LlavaOnevisionModelOutputWithPast]:
+        r"""
+        image_sizes_videos (`torch.LongTensor` of shape `(batch_size, frames, 2)`, *optional*):
+            The sizes of the videos in the batch, being (height, width) for each frame in the video.
+        vision_aspect_ratio (`str`, *optional*, defaults to `"anyres_max_9"`):
+            Aspect ratio used when processong image features. The default value is "anyres_max_9".
+        batch_num_images (`torch.LongTensor`, *optional*):
+            Number of images in each sample.
+        """
+
+        output_attentions = output_attentions if output_attentions is not None else self.config.output_attentions
+        output_hidden_states = (
+            output_hidden_states if output_hidden_states is not None else self.config.output_hidden_states
+        )
+        return_dict = return_dict if return_dict is not None else self.config.use_return_dict
+        vision_feature_layer = (
+            vision_feature_layer if vision_feature_layer is not None else self.config.vision_feature_layer
+        )
+        vision_feature_select_strategy = (
+            vision_feature_select_strategy
+            if vision_feature_select_strategy is not None
+            else self.config.vision_feature_select_strategy
+        )
+        vision_aspect_ratio = (
+            vision_aspect_ratio if vision_aspect_ratio is not None else self.config.vision_aspect_ratio
+        )
+
+        if (input_ids is None) ^ (inputs_embeds is not None):
+            raise ValueError("You must specify exactly one of input_ids or inputs_embeds")
+
+        if inputs_embeds is None:
+            inputs_embeds = self.get_input_embeddings()(input_ids)
+
+        # Images are processed with Anyres
+        if pixel_values is not None:
+            image_features = self.get_image_features(
+                pixel_values,
+                image_sizes,
+                vision_feature_layer=vision_feature_layer,
+                vision_feature_select_strategy=vision_feature_select_strategy,
+                batch_num_images=batch_num_images,
+            )
+            image_features = torch.cat(image_features, dim=0)
+
+            if input_ids is None:
+                special_image_mask = inputs_embeds == self.get_input_embeddings()(
+                    torch.tensor(self.config.image_token_id, dtype=torch.long, device=inputs_embeds.device)
+                )
+                special_image_mask = special_image_mask.all(-1)
+            else:
+                special_image_mask = input_ids == self.config.image_token_id
+
+            n_image_tokens = (special_image_mask).sum()
+            special_image_mask = special_image_mask.unsqueeze(-1).expand_as(inputs_embeds).to(inputs_embeds.device)
+
+            if not is_torchdynamo_compiling() and inputs_embeds[special_image_mask].numel() != image_features.numel():
+                n_image_features = image_features.shape[0]
+                raise ValueError(
+                    f"Image features and image tokens do not match: tokens: {n_image_tokens}, features {n_image_features}"
+                )
+            image_features = image_features.to(inputs_embeds.device, inputs_embeds.dtype)
+            inputs_embeds = inputs_embeds.masked_scatter(special_image_mask, image_features)
+
+        # Video are simply embedded and further pooled to decrease seq len
+        if pixel_values_videos is not None:
+            video_features = self.get_video_features(
+                pixel_values_videos,
+                vision_feature_layer=vision_feature_layer,
+                vision_feature_select_strategy=vision_feature_select_strategy,
+            )
+            image_newline = (
+                self.image_newline[None, None, :].repeat(video_features.shape[0], 1, 1).to(video_features.device)
+            )
+            video_features = torch.cat((video_features, image_newline), dim=1)
+            video_features = video_features.flatten(0, 1)
+
+            if input_ids is None:
+                special_video_mask = inputs_embeds == self.get_input_embeddings()(
+                    torch.tensor(self.config.video_token_id, dtype=torch.long, device=inputs_embeds.device)
+                )
+                special_video_mask = special_video_mask.all(-1)
+            else:
+                special_video_mask = input_ids == self.config.video_token_id
+
+            n_video_tokens = (special_video_mask).sum()
+            special_video_mask = special_video_mask.unsqueeze(-1).expand_as(inputs_embeds).to(inputs_embeds.device)
+
+            if not is_torchdynamo_compiling() and inputs_embeds[special_video_mask].numel() != video_features.numel():
+                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}"
+                )
+            video_features = video_features.to(inputs_embeds.device, inputs_embeds.dtype)
+            inputs_embeds = inputs_embeds.masked_scatter(special_video_mask, video_features)
+
+        outputs = self.language_model(
+            attention_mask=attention_mask,
+            position_ids=position_ids,
+            past_key_values=past_key_values,
+            inputs_embeds=inputs_embeds,
+            use_cache=use_cache,
+            output_attentions=output_attentions,
+            output_hidden_states=output_hidden_states,
+            return_dict=True,
+            cache_position=cache_position,
+            **kwargs,
+        )
+
+        return LlavaOnevisionModelOutputWithPast(
+            last_hidden_state=outputs.last_hidden_state,
+            past_key_values=outputs.past_key_values,
+            hidden_states=outputs.hidden_states,
+            attentions=outputs.attentions,
+            image_hidden_states=image_features if pixel_values is not None else None,
+            video_hidden_states=video_features if pixel_values_videos is not None else None,
+        )
+
+
+class LlavaOnevisionForConditionalGeneration(LlavaNextVideoForConditionalGeneration):
+    @can_return_tuple
+    @auto_docstring
+    def forward(
+        self,
+        input_ids: torch.LongTensor = None,
+        pixel_values: torch.FloatTensor = None,
+        image_sizes: Optional[torch.LongTensor] = None,
+        pixel_values_videos: torch.FloatTensor = None,
+        image_sizes_videos: Optional[torch.LongTensor] = None,
+        attention_mask: Optional[torch.Tensor] = None,
+        position_ids: Optional[torch.LongTensor] = None,
+        past_key_values: Optional[Cache] = None,
+        inputs_embeds: Optional[torch.FloatTensor] = None,
+        vision_feature_layer: Optional[Union[int, list[int]]] = None,
+        vision_feature_select_strategy: Optional[str] = None,
+        vision_aspect_ratio: Optional[str] = None,
+        batch_num_images: Optional[torch.LongTensor] = None,
+        labels: Optional[torch.LongTensor] = None,
+        use_cache: Optional[bool] = None,
+        output_attentions: Optional[bool] = None,
+        output_hidden_states: Optional[bool] = None,
+        return_dict: Optional[bool] = None,
+        cache_position: Optional[torch.LongTensor] = None,
+        logits_to_keep: Union[int, torch.Tensor] = 0,
+        **kwargs: Unpack[TransformersKwargs],
+    ) -> Union[tuple, LlavaOnevisionCausalLMOutputWithPast]:
+        r"""
+        image_sizes_videos (`torch.LongTensor` of shape `(batch_size, frames, 2)`, *optional*):
+            The sizes of the videos in the batch, being (height, width) for each frame in the video.
+        vision_aspect_ratio (`str`, *optional*, defaults to `"anyres_max_9"`):
+            Aspect ratio used when processong image features. The default value is "anyres_max_9".
+        batch_num_images (`torch.LongTensor`, *optional*):
+            Number of images in each sample.
+        labels (`torch.LongTensor` of shape `(batch_size, sequence_length)`, *optional*):
+            Labels for computing the masked language modeling loss. Indices should either be in `[0, ...,
+            config.vocab_size]` or -100 (see `input_ids` docstring). Tokens with indices set to `-100` are ignored
+            (masked), the loss is only computed for the tokens with labels in `[0, ..., config.vocab_size]`.
+
+        Example:
+
+        ```python
+        >>> from PIL import Image
+        >>> import requests
+        >>> import torch
+        >>> from transformers import LlavaOnevisionProcessor, LlavaOnevisionForConditionalGeneration
+
+        >>> model = LlavaOnevisionForConditionalGeneration.from_pretrained("llava-hf/llava-onevision-qwen2-7b-ov-hf", torch_dtype="float16", device_map="cuda:0")
+        >>> processor = LlavaOnevisionProcessor.from_pretrained("llava-hf/llava-onevision-qwen2-7b-ov-hf")
+
+        >>> conversation = [
+        ...     {
+        ...       "role": "user",
+        ...       "content": [
+        ...           {"type": "text", "text": "What is shown in this image?"},
+        ...           {"type": "image"},
+        ...         ],
+        ...     },
+        ... ]
+        >>> prompt = processor.apply_chat_template(conversation, add_generation_prompt=True)
+
+        >>> image_file = "http://images.cocodataset.org/val2017/000000039769.jpg"
+        >>> raw_image = Image.open(requests.get(image_file, stream=True).raw)
+        >>> inputs = processor(text=prompt, images=raw_image, return_tensors='pt').to(0, torch.float16)
+
+        >>> output = model.generate(**inputs, max_new_tokens=20, do_sample=False)
+        >>> processor.batch_decode(output, skip_special_tokens=True)[0]
+        "user\n\nWhat is shown in this image?\nassistant\ncat"
+        ```"""
+        output_attentions = output_attentions if output_attentions is not None else self.config.output_attentions
+        output_hidden_states = (
+            output_hidden_states if output_hidden_states is not None else self.config.output_hidden_states
+        )
+        return_dict = return_dict if return_dict is not None else self.config.use_return_dict
+        vision_feature_layer = (
+            vision_feature_layer if vision_feature_layer is not None else self.config.vision_feature_layer
+        )
+        vision_feature_select_strategy = (
+            vision_feature_select_strategy
+            if vision_feature_select_strategy is not None
+            else self.config.vision_feature_select_strategy
+        )
+        vision_aspect_ratio = (
+            vision_aspect_ratio if vision_aspect_ratio is not None else self.config.vision_aspect_ratio
+        )
+
+        outputs = self.model(
+            input_ids=input_ids,
+            pixel_values=pixel_values,
+            pixel_values_videos=pixel_values_videos,
+            image_sizes=image_sizes,
+            image_sizes_videos=image_sizes_videos,
+            vision_aspect_ratio=vision_aspect_ratio,
+            vision_feature_layer=vision_feature_layer,
+            vision_feature_select_strategy=vision_feature_select_strategy,
+            batch_num_images=batch_num_images,
+            attention_mask=attention_mask,
+            position_ids=position_ids,
+            past_key_values=past_key_values,
+            inputs_embeds=inputs_embeds,
+            use_cache=use_cache,
+            output_attentions=output_attentions,
+            output_hidden_states=output_hidden_states,
+            return_dict=True,
+            cache_position=cache_position,
+            logits_to_keep=logits_to_keep,
+            **kwargs,
+        )
+
+        hidden_states = outputs[0]
+        # Only compute necessary logits, and do not upcast them to float if we are not computing the loss
+        slice_indices = slice(-logits_to_keep, None) if isinstance(logits_to_keep, int) else logits_to_keep
+        logits = self.lm_head(hidden_states[:, slice_indices, :])
+
+        loss = None
+        if labels is not None:
+            loss = self.loss_function(
+                logits=logits, labels=labels, vocab_size=self.config.text_config.vocab_size, **kwargs
+            )
+
+        return LlavaOnevisionCausalLMOutputWithPast(
+            loss=loss,
+            logits=logits,
+            past_key_values=outputs.past_key_values,
+            hidden_states=outputs.hidden_states,
+            attentions=outputs.attentions,
+            image_hidden_states=outputs.image_hidden_states,
+            video_hidden_states=outputs.video_hidden_states,
+        )
+
+    def prepare_inputs_for_generation(
+        self,
+        input_ids,
+        past_key_values=None,
+        inputs_embeds=None,
+        pixel_values=None,
+        image_sizes=None,
+        pixel_values_videos=None,
+        image_sizes_videos=None,
+        attention_mask=None,
+        cache_position=None,
+        logits_to_keep=None,
+        **kwargs,
+    ):
+        # Overwritten -- in specific circumstances we don't want to forward image inputs to the model
+
+        model_inputs = super().prepare_inputs_for_generation(
+            input_ids,
+            past_key_values=past_key_values,
+            inputs_embeds=inputs_embeds,
+            attention_mask=attention_mask,
+            cache_position=cache_position,
+            logits_to_keep=logits_to_keep,
+            **kwargs,
+        )
+
+        if cache_position[0] == 0:
+            # If we're in cached decoding stage, pixel values should be None because input ids do not contain special image token anymore
+            # Otherwise we need pixel values to be passed to model
+            model_inputs["pixel_values"] = pixel_values
+            model_inputs["image_sizes"] = image_sizes
+            model_inputs["pixel_values_videos"] = pixel_values_videos
+            model_inputs["image_sizes_videos"] = image_sizes_videos
+
+        return model_inputs
+
+
+__all__ = [
+    "LlavaOnevisionImageProcessorFast",
+    "LlavaOnevisionModel",
+    "LlavaOnevisionForConditionalGeneration",
+    "LlavaOnevisionPreTrainedModel",
+]

venv/Lib/site-packages/transformers/models/llava_onevision/processing_llava_onevision.py

@@ -0,0 +1,355 @@
+# coding=utf-8
+# Copyright 2024 The HuggingFace Inc. team.
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+"""
+Processor class for LLaVa-Onevision.
+"""
+
+import math
+from collections.abc import Iterable
+from typing import Union
+
+import numpy as np
+
+from ...feature_extraction_utils import BatchFeature
+from ...image_processing_utils import select_best_resolution
+from ...image_utils import ImageInput, get_image_size, to_numpy_array
+from ...processing_utils import MultiModalData, ProcessingKwargs, ProcessorMixin, Unpack
+from ...tokenization_utils_base import PreTokenizedInput, TextInput
+from ...utils import logging
+from ...video_utils import VideoInput
+
+
+logger = logging.get_logger(__name__)
+
+
+class LlavaOnevisionProcessorKwargs(ProcessingKwargs, total=False):
+    # see processing_utils.ProcessingKwargs documentation for usage.
+    _defaults = {
+        "text_kwargs": {
+            "padding": False,
+            "return_mm_token_type_ids": False,
+        },
+        "image_kwargs": {},
+        "videos_kwargs": {},
+    }
+
+
+class LlavaOnevisionProcessor(ProcessorMixin):
+    r"""
+    Constructs a LLaVa-Onevision processor which wraps a LLaVa-Onevision video processor, LLaVa-NeXT image processor and a LLaMa tokenizer into a single processor.
+
+    [`LlavaNextProcessor`] offers all the functionalities of [`LlavaOnevisionVideoProcessor`], [`LlavaOnevisionImageProcessor`] and [`LlamaTokenizerFast`]. See the
+    [`~LlavaOnevisionVideoProcessor.__call__`], [`~LlavaNextProcessor.__call__`] and [`~LlavaNextProcessor.decode`] for more information.
+
+    Args:
+        image_processor ([`LlavaOnevisionImageProcessor`], *optional*):
+            The image processor is a required input.
+        tokenizer ([`LlamaTokenizerFast`], *optional*):
+            The tokenizer is a required input.
+        video_processor ([`LlavaOnevisionVideoProcessor`], *optional*):
+            The video processor is a required input.
+        num_image_tokens (`int`, *optional*):
+            Number of image tokens for one imagethat will be returned by vision tower.
+        vision_feature_select_strategy (`str`, *optional*):
+            The feature selection strategy used to select the vision feature from the vision backbone.
+            Should be same as in model's config
+        chat_template (`str`, *optional*): A Jinja template which will be used to convert lists of messages
+            in a chat into a tokenizable string.
+        image_token (`str`, *optional*, defaults to `"<image>"`):
+            Special token used to denote image location.
+        video_token (`str`, *optional*, defaults to `"<video>"`):
+            Special token used to denote video location.
+        vision_aspect_ratio (`str`, *optional*, defaults to `"anyres_max_9"`):
+            Aspect ratio used when processong image features. The default value is "anyres_max_9".
+    """
+
+    attributes = ["image_processor", "tokenizer", "video_processor"]
+    image_processor_class = "AutoImageProcessor"
+    tokenizer_class = "AutoTokenizer"
+    video_processor_class = "AutoVideoProcessor"
+
+    def __init__(
+        self,
+        image_processor=None,
+        tokenizer=None,
+        video_processor=None,
+        num_image_tokens=None,
+        vision_feature_select_strategy=None,
+        chat_template=None,
+        image_token="<image>",
+        video_token="<video>",
+        vision_aspect_ratio="anyres_max_9",
+        **kwargs,
+    ):
+        self.num_image_tokens = num_image_tokens
+        self.vision_feature_select_strategy = vision_feature_select_strategy
+        self.image_token = tokenizer.image_token if hasattr(tokenizer, "image_token") else image_token
+        self.video_token = tokenizer.video_token if hasattr(tokenizer, "video_token") else video_token
+        self.image_token_id = (
+            tokenizer.image_token_id
+            if getattr(tokenizer, "image_token_id", None)
+            else tokenizer.convert_tokens_to_ids(self.image_token)
+        )
+        self.video_token_id = (
+            tokenizer.video_token_id
+            if getattr(tokenizer, "video_token_id", None)
+            else tokenizer.convert_tokens_to_ids(self.video_token)
+        )
+        self.vision_aspect_ratio = vision_aspect_ratio
+        super().__init__(image_processor, tokenizer, video_processor, chat_template=chat_template)
+
+    def __call__(
+        self,
+        images: ImageInput = None,
+        text: Union[TextInput, PreTokenizedInput, list[TextInput], list[PreTokenizedInput]] = None,
+        audio=None,
+        videos: VideoInput = None,
+        **kwargs: Unpack[LlavaOnevisionProcessorKwargs],
+    ) -> BatchFeature:
+        """
+        Main method to prepare for the model one or several sequences(s) and image(s). This method forwards the `text`
+        and `kwargs` arguments to LlamaTokenizerFast's [`~LlamaTokenizerFast.__call__`] if `text` is not `None` to encode
+        the text. To prepare the image(s), this method forwards the `images` and `kwrags` arguments to
+        LlavaNextImageProcessor's [`~LlavaNextImageProcessor.__call__`] if `images` is not `None`. Please refer to the docstring
+        of the above two methods for more information.
+
+        Args:
+            images (`PIL.Image.Image`, `np.ndarray`, `torch.Tensor`, `list[PIL.Image.Image]`, `list[np.ndarray]`, `list[torch.Tensor]`):
+                The image or batch of images to be prepared. Each image can be a PIL image, NumPy array or PyTorch
+                tensor. Both channels-first and channels-last formats are supported.
+            text (`str`, `list[str]`, `list[list[str]]`):
+                The sequence or batch of sequences to be encoded. Each sequence can be a string or a list of strings
+                (pretokenized string). If the sequences are provided as list of strings (pretokenized), you must set
+                `is_split_into_words=True` (to lift the ambiguity with a batch of sequences).
+            videos (`np.ndarray`, `torch.Tensor`, `list[np.ndarray]`, `list[torch.Tensor]`):
+                The image or batch of videos to be prepared. Each video can be a 4D NumPy array or PyTorch
+
+        Returns:
+            [`BatchFeature`]: A [`BatchFeature`] with the following fields:
+
+            - **input_ids** -- List of token ids to be fed to a model. Returned when `text` is not `None`.
+            - **attention_mask** -- List of indices specifying which tokens should be attended to by the model (when
+              `return_attention_mask=True` or if *"attention_mask"* is in `self.model_input_names` and if `text` is not
+              `None`).
+            - **pixel_values** -- Pixel values to be fed to a model. Returned when `images` is not `None`.
+            - **pixel_values_videos** -- Pixel values of a video input to be fed to a model. Returned when `videos` is not `None`.
+            - **image_sizes** -- Size of each image that will be used to unpad an image. Returned when `images` is not `None`.
+        """
+
+        output_kwargs = self._merge_kwargs(
+            LlavaOnevisionProcessorKwargs,
+            tokenizer_init_kwargs=self.tokenizer.init_kwargs,
+            **kwargs,
+        )
+
+        if isinstance(text, str):
+            text = [text]
+        elif not isinstance(text, list) and not isinstance(text[0], str):
+            raise TypeError("Invalid input text. Please provide a string, or a list of strings")
+
+        image_inputs = video_inputs = {}
+
+        if images is not None:
+            image_inputs = self.image_processor(images, **output_kwargs["images_kwargs"])
+
+            batch_num_images = iter(image_inputs["batch_num_images"])
+            image_sizes = iter(image_inputs["image_sizes"])
+            height, width = get_image_size(
+                to_numpy_array(image_inputs["pixel_values"][0][0]),
+                channel_dim=output_kwargs["images_kwargs"].get("data_format"),
+            )
+            text, num_image_tokens = self._expand_image_tokens(
+                text, image_sizes, height, width, self.image_token, batch_num_images
+            )
+
+        if videos is not None:
+            video_inputs = self.video_processor(videos, **output_kwargs["videos_kwargs"])
+
+            one_video = video_inputs.get("pixel_values_videos")[0]
+            if isinstance(video_inputs.get("pixel_values_videos")[0], (list, tuple)):
+                one_video = np.array(one_video)
+            else:
+                one_video = to_numpy_array(one_video)
+            height, width = get_image_size(one_video[0], channel_dim=output_kwargs["images_kwargs"].get("data_format"))
+            num_frames = one_video.shape[0]  # frame dim is always after batch dim
+            patches_height_width = int(math.sqrt(self.num_image_tokens))
+            pooled_height_width = math.ceil(patches_height_width / 2)
+            num_video_tokens = (num_frames * pooled_height_width * pooled_height_width) + 1  # +1 for newline token
+            text = [sample.replace(self.video_token, self.video_token * num_video_tokens) for sample in text]
+
+        return_tensors = output_kwargs["text_kwargs"].pop("return_tensors", None)
+        return_mm_token_type_ids = output_kwargs["text_kwargs"].pop("return_mm_token_type_ids", None)
+        text_inputs = self.tokenizer(text, **output_kwargs["text_kwargs"])
+        self._check_special_mm_tokens(text, text_inputs, modalities=["image"])
+
+        if return_mm_token_type_ids:
+            array_ids = np.array(text_inputs["input_ids"])
+            mm_token_type_ids = np.zeros_like(text_inputs["input_ids"])
+            mm_token_type_ids[array_ids == self.image_token_id] = 1
+            text_inputs["mm_token_type_ids"] = mm_token_type_ids.tolist()
+
+        return BatchFeature(data={**text_inputs, **image_inputs, **video_inputs}, tensor_type=return_tensors)
+
+    def _expand_image_tokens(
+        self,
+        text: list[TextInput],
+        image_sizes: Iterable[Union[list[int], int]],
+        height: int,
+        width: int,
+        special_token: str,
+        batch_num_images: Iterable[int],
+    ):
+        prompt_strings = []
+        max_num_vision_tokens = 0
+        for sample in text:
+            if special_token in sample:
+                is_multi_image = next(batch_num_images) != 1
+            else:
+                is_multi_image = False
+            while special_token in sample:
+                if is_multi_image:
+                    num_image_tokens = self.num_image_tokens + 1  # one for image_newline
+                else:
+                    original_size = next(image_sizes)
+                    if not isinstance(original_size, (list, tuple)):
+                        # cast to list to avoid numerical precision errors when calculating unpadding
+                        original_size = original_size.tolist()
+                    orig_height, orig_width = original_size
+                    num_image_tokens = self._get_number_of_features(orig_height, orig_width, height, width)
+                max_num_vision_tokens = max(max_num_vision_tokens, num_image_tokens)
+                if self.vision_feature_select_strategy == "default":
+                    num_image_tokens -= 1
+                sample = sample.replace(special_token, "<placeholder>" * num_image_tokens, 1)
+            prompt_strings.append(sample)
+        text = [sample.replace("<placeholder>", special_token) for sample in prompt_strings]
+        return text, max_num_vision_tokens
+
+    def _get_number_of_features(self, orig_height: int, orig_width: int, height: int, width: int) -> int:
+        image_grid_pinpoints = self.image_processor.image_grid_pinpoints
+
+        height_best_resolution, width_best_resolution = select_best_resolution(
+            [orig_height, orig_width], image_grid_pinpoints
+        )
+        scale_height, scale_width = height_best_resolution // height, width_best_resolution // width
+
+        patches_height = patches_width = int(math.sqrt(self.num_image_tokens))
+        unpadded_features, newline_features = self._get_unpadded_features(
+            orig_height, orig_width, patches_height, patches_width, scale_height, scale_width
+        )
+
+        # The base patch covers the entire image (no CLS for SigLIP)
+        base_features = self.num_image_tokens
+        num_image_tokens = unpadded_features + newline_features + base_features
+        return num_image_tokens
+
+    # Adapted from transformers.models.llava_next.processing_llava_next.LlavaNextProcessor._get_unpadded_features
+    def _get_unpadded_features(self, height, width, patches_height, patches_width, scale_height, scale_width):
+        """
+        Get number of features for a given image with height/width. LLaVA-NeXT is different from LLaVA
+        because it divided each image into patches depending on its resolution. Therefore we need to calculate how many
+        patches an image is divided into and get the number of features from that.
+        """
+        current_height = patches_height * scale_height
+        current_width = patches_width * scale_width
+
+        original_aspect_ratio = width / height
+        current_aspect_ratio = current_width / current_height
+        if original_aspect_ratio > current_aspect_ratio:
+            new_height = int(round(height * (current_width / width), 7))
+            padding = (current_height - new_height) // 2
+            current_height -= padding * 2
+        else:
+            new_width = int(round(width * (current_height / height), 7))
+            padding = (current_width - new_width) // 2
+            current_width -= padding * 2
+
+        unpadded_features = current_height * current_width
+        newline_features = current_height
+
+        max_num_patches = int(self.vision_aspect_ratio.strip("anyres_max_"))
+        ratio = math.sqrt(current_height * current_width / (max_num_patches * patches_height**2))
+        if ratio > 1.1:
+            unpadded_features = int(current_height // ratio) * int(current_width // ratio)
+            newline_features = int(current_height // ratio)
+
+        return (unpadded_features, newline_features)
+
+    def _get_num_multimodal_tokens(self, image_sizes=None, video_sizes=None, **kwargs):
+        """
+        Computes the number of placeholder tokens needed for multimodal inputs with the given sizes.
+        Args:
+            image_sizes (list[list[str]], *optional*):
+                The input sizes formatted as (height, width) per each image.
+            video_sizes (list[list[str]], *optional*):
+                The input sizes formatted as (num_frames, height, width) per each video.
+            audio_lengths (list[int], *optional*):
+                The input length formatted as per each audio.
+        Returns:
+            dict[str, list[int]]: A dictionary mapping each modality ("image", "video", "audio")
+            to a list containing the number of placeholder tokens required. If the model doesn't accept
+            a certain modality or no input sizes are provided, the dict value is set to an empty list.
+        """
+        vision_data = {}
+        if image_sizes is not None:
+            images_kwargs = LlavaOnevisionProcessorKwargs._defaults.get("images_kwargs", {})
+            images_kwargs.update(kwargs)
+
+            size = images_kwargs.get("size", None) or self.image_processor.size
+            size = (
+                (size["shortest_edge"], size["shortest_edge"])
+                if "shortest_edge" in size
+                else (min(size["height"], size["width"]), min(size["height"], size["width"]))
+            )
+            processed_height, processed_width = size
+
+            batch_num_image_tokens = []
+            num_image_patches = [1] * len(image_sizes)  # llava-ov doesn't batch pixels as Idefics, thus `1` patch`
+            for image_size in image_sizes:
+                orig_height, orig_width = image_size
+                num_image_tokens = self._get_number_of_features(
+                    orig_height, orig_width, processed_height, processed_width
+                )
+                if self.vision_feature_select_strategy == "default":
+                    num_image_tokens -= 1
+                batch_num_image_tokens.append(num_image_tokens)
+            vision_data.update({"num_image_tokens": batch_num_image_tokens, "num_image_patches": num_image_patches})
+
+        return MultiModalData(**vision_data)
+
+    # Copied from transformers.models.clip.processing_clip.CLIPProcessor.batch_decode with CLIP->Llama
+    def batch_decode(self, *args, **kwargs):
+        """
+        This method forwards all its arguments to LlamaTokenizerFast's [`~PreTrainedTokenizer.batch_decode`]. Please
+        refer to the docstring of this method for more information.
+        """
+        return self.tokenizer.batch_decode(*args, **kwargs)
+
+    # Copied from transformers.models.clip.processing_clip.CLIPProcessor.decode with CLIP->Llama
+    def decode(self, *args, **kwargs):
+        """
+        This method forwards all its arguments to LlamaTokenizerFast's [`~PreTrainedTokenizer.decode`]. Please refer to
+        the docstring of this method for more information.
+        """
+        return self.tokenizer.decode(*args, **kwargs)
+
+    @property
+    # Copied from transformers.models.clip.processing_clip.CLIPProcessor.model_input_names
+    def model_input_names(self):
+        tokenizer_input_names = self.tokenizer.model_input_names
+        image_processor_input_names = self.image_processor.model_input_names
+        return list(dict.fromkeys(tokenizer_input_names + image_processor_input_names))
+
+
+__all__ = ["LlavaOnevisionProcessor"]

venv/Lib/site-packages/transformers/models/llava_onevision/video_processing_llava_onevision.py

@@ -0,0 +1,58 @@
+# coding=utf-8
+# Copyright 2025 The HuggingFace Inc. team. All rights reserved.
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+"""Video processor class for LLaVa-Onevision."""
+
+from ...image_utils import (
+    OPENAI_CLIP_MEAN,
+    OPENAI_CLIP_STD,
+)
+from ...processing_utils import Unpack, VideosKwargs
+from ...utils import is_vision_available
+from ...utils.import_utils import requires
+from ...video_processing_utils import (
+    BaseVideoProcessor,
+)
+
+
+if is_vision_available():
+    from ...image_utils import PILImageResampling
+
+
+class LlavaOnevisionFastVideoProcessorInitKwargs(VideosKwargs): ...
+
+
+@requires(backends=("torchvision",))
+class LlavaOnevisionVideoProcessor(BaseVideoProcessor):
+    resample = PILImageResampling.BICUBIC
+    image_mean = OPENAI_CLIP_MEAN
+    image_std = OPENAI_CLIP_STD
+    size = {"height": 384, "width": 384}
+    rescale_factor = 1 / 255
+    default_to_square = False
+    crop_size = None
+    do_resize = True
+    do_center_crop = None
+    do_rescale = True
+    do_normalize = True
+    do_convert_rgb = True
+    do_sample_frames = False  # Set to False for BC, recommended to set `True` in new models
+    valid_kwargs = LlavaOnevisionFastVideoProcessorInitKwargs
+    model_input_names = ["pixel_values_videos"]
+
+    def __init__(self, **kwargs: Unpack[LlavaOnevisionFastVideoProcessorInitKwargs]):
+        super().__init__(**kwargs)
+
+
+__all__ = ["LlavaOnevisionVideoProcessor"]