Upload folder using huggingface_hub

config.json

@@ -0,0 +1,24 @@
+{
+  "architectures": [
+    "ViSCoP_VisionEncoderModel"
+  ],
+  "attention_dropout": 0.0,
+  "auto_map": {
+    "AutoConfig": "configuration_viscop_vision_encoder.ViSCoP_VisionEncoderConfig",
+    "AutoModel": "modeling_viscop_vision_encoder.ViSCoP_VisionEncoderModel"
+  },
+  "hidden_act": "gelu_pytorch_tanh",
+  "hidden_size": 1152,
+  "interaction_module": "cross_attention",
+  "interaction_module_layers": null,
+  "intermediate_size": 4304,
+  "layer_norm_eps": 1e-06,
+  "model_type": "viscop_vision_encoder",
+  "num_attention_heads": 16,
+  "num_channels": 3,
+  "num_hidden_layers": 27,
+  "num_visual_probes": 16,
+  "patch_size": 14,
+  "torch_dtype": "bfloat16",
+  "transformers_version": "4.46.3"
+}

configuration_viscop_vision_encoder.py

@@ -0,0 +1,49 @@
+# Adopted from https://github.com/DAMO-NLP-SG/VideoLLaMA3/blob/main/videollama3/model/videollama3_encoder/configuration_videollama3_encoder.py
+# Adopted from https://github.com/huggingface/transformers/blob/main/src/transformers/models/siglip/configuration_siglip.py.
+# Below is the original copyright:
+# 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.
+"""ViSCoP vision encoder model configuration."""
+
+from transformers import PretrainedConfig
+
+
+class ViSCoP_VisionEncoderConfig(PretrainedConfig):
+
+    model_type = "viscop_vision_encoder"
+
+    def __init__(
+        self,
+        hidden_size=768,
+        intermediate_size=3072,
+        num_hidden_layers=12,
+        num_attention_heads=12,
+        num_channels=3,
+        patch_size=16,
+        hidden_act="gelu_pytorch_tanh",
+        layer_norm_eps=1e-6,
+        attention_dropout=0.0,
+        **kwargs,
+    ):
+        super().__init__(**kwargs)
+        self.hidden_size = hidden_size
+        self.intermediate_size = intermediate_size
+        self.num_hidden_layers = num_hidden_layers
+        self.num_attention_heads = num_attention_heads
+        self.num_channels = num_channels
+        self.patch_size = patch_size
+        self.attention_dropout = attention_dropout
+        self.layer_norm_eps = layer_norm_eps
+        self.hidden_act = hidden_act

image_processing_viscop.py

@@ -0,0 +1,488 @@
+# Adopted from https://github.com/DAMO-NLP-SG/VideoLLaMA3/blob/main/videollama3/model/videollama3_encoder/image_processing_videollama3.py
+# Adopted from https://github.com/huggingface/transformers/blob/main/src/transformers/models/qwen2_vl/image_processing_qwen2_vl.py.
+# Below is the original copyright:
+# Copyright 2024 The Qwen team, Alibaba Group and the HuggingFace Inc. team. All rights reserved.
+#
+# This code is based on EleutherAI's GPT-NeoX library and the GPT-NeoX
+# and OPT implementations in this library. It has been modified from its
+# original forms to accommodate minor architectural differences compared
+# to GPT-NeoX and OPT used by the Meta AI team that trained the model.
+#
+# 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 ViSCoP."""
+
+import math
+from typing import Dict, List, Optional, Union
+
+import numpy as np
+
+import torch
+from transformers.image_processing_utils import BaseImageProcessor, BatchFeature
+from transformers.image_utils import ImageInput
+from transformers.image_transforms import (
+    convert_to_rgb,
+    resize,
+    to_channel_dimension_format,
+)
+from transformers.image_utils import (
+    OPENAI_CLIP_MEAN,
+    OPENAI_CLIP_STD,
+    ChannelDimension,
+    ImageInput,
+    PILImageResampling,
+    VideoInput,
+    get_image_size,
+    infer_channel_dimension_format,
+    is_scaled_image,
+    is_valid_image,
+    make_list_of_images,
+    to_numpy_array,
+)
+from transformers.utils import TensorType, is_vision_available, logging
+
+
+logger = logging.get_logger(__name__)
+
+
+if is_vision_available():
+    from PIL import Image
+
+
+def is_valid_video(video) -> bool:
+    if isinstance(video, (list, tuple)):
+        return all(is_valid_image(frame) for frame in video)
+    elif isinstance(video, np.ndarray):
+        return video.ndim == 4
+    elif isinstance(video, torch.Tensor):
+        return video.ndim == 4
+    return False
+
+
+def make_batched_images(images) -> List[List[ImageInput]]:
+    """
+    Accepts images in list or nested list format, and makes a list of images for preprocessing.
+
+    Args:
+        images (`Union[List[List[ImageInput]], List[ImageInput], ImageInput]`):
+            The input image.
+
+    Returns:
+        list: A list of images.
+    """
+    if isinstance(images, (list, tuple)):
+        # list of images/videos
+        if not all(is_valid_video(image) or is_valid_image(image) for image in images):
+            raise ValueError(f"Could not make batched images from {images}")
+        return images
+    elif is_valid_video(images) or is_valid_image(images):
+        # single image/video
+        return [images]
+
+    raise ValueError(f"Could not make batched images from {images}")
+
+
+def simple_batched_resize(
+    images, factor: int = 28, min_tokens: int = 4 * 4, max_tokens: int = 16384, input_data_format: str = None
+):
+    min_pixels = min_tokens * factor * factor
+    max_pixels = max_tokens * factor * factor
+
+    num_images = 0
+    for image in images:
+        if is_valid_video(image):
+            num_images += len(image)
+        else:
+            num_images += 1
+
+    image_sizes = []
+    for image in images:
+        if is_valid_video(image):
+            image = image[0]
+        if isinstance(image, Image.Image):
+            width, height = image.size
+        else:
+            height, width = get_image_size(image, channel_dim=input_data_format)
+        image_sizes.append([height, width])
+
+    tmp_image_sizes = []
+    for height, width in image_sizes:
+        h_bar = round(height / factor) * factor
+        w_bar = round(width / factor) * factor
+        if h_bar * w_bar > (max_pixels // num_images):
+            beta = math.sqrt((height * width) / (max_pixels // num_images))
+            h_bar = math.floor(height / beta / factor) * factor
+            w_bar = math.floor(width / beta / factor) * factor
+        # per image min_pixels
+        if h_bar * w_bar < min_pixels:
+            beta = math.sqrt(min_pixels / (height * width))
+            h_bar = math.ceil(height * beta / factor) * factor
+            w_bar = math.ceil(width * beta / factor) * factor
+        tmp_image_sizes.append((h_bar, w_bar))
+    image_sizes = tmp_image_sizes
+    return image_sizes
+
+
+def batched_resize(
+    images, factors: List[int], min_tokens: int = 4 * 4, max_tokens: int = 16384, input_data_format: str = None
+):
+    image_sizes = []
+    for image in images:
+        if is_valid_video(image):
+            num_frame = len(image)
+            image = image[0]
+        else:
+            num_frame = 1
+        if isinstance(image, Image.Image):
+            width, height = image.size
+        else:
+            height, width = get_image_size(image, channel_dim=input_data_format)
+        image_sizes.append([num_frame, height, width])
+
+    # global max_pixels
+    smart_scale_factors = 1.0
+    total_tokens = 0
+    for (num_frame, height, width), factor in zip(image_sizes, factors):
+        total_tokens += num_frame * math.ceil(height / factor) * math.ceil(width / factor)
+
+    # TODO: add min_pixels
+    if total_tokens > max_tokens:
+        beta = math.sqrt(total_tokens / max_tokens)
+        tmp_image_sizes = []
+        for (_, height, width), factor in zip(image_sizes, factors):
+            h_bar = math.floor(height / beta / factor) * factor
+            w_bar = math.floor(width / beta / factor) * factor
+            tmp_image_sizes.append((h_bar, w_bar))
+        image_sizes = tmp_image_sizes
+    else:
+        tmp_image_sizes = []
+        for (_, height, width), factor in zip(image_sizes, factors):
+            height = round(height / factor) * factor
+            width = round(width / factor) * factor
+            tmp_image_sizes.append((height, width))
+        image_sizes = tmp_image_sizes
+
+    return image_sizes
+
+
+class ViSCoP_ImageProcessor(BaseImageProcessor):
+    r"""
+    Constructs a DAMO-VL (https://huggingface.co/DAMO-NLP-SG/VL3-SigLIP-NaViT) image processor that dynamically resizes images based on the original images.
+
+    Args:
+        do_resize (`bool`, *optional*, defaults to `True`):
+            Whether to resize the image's (height, width) dimensions.
+        resample (`PILImageResampling`, *optional*, defaults to `Resampling.BICUBIC`):
+            Resampling filter to use when resizing the image.
+        do_rescale (`bool`, *optional*, defaults to `True`):
+            Whether to rescale the image by the specified scale `rescale_factor`.
+        rescale_factor (`int` or `float`, *optional*, defaults to `1/255`):
+            Scale factor to use if rescaling the image.
+        do_normalize (`bool`, *optional*, defaults to `True`):
+            Whether to normalize the image.
+        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 for each channel in the image.
+        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 for each channel in the image.
+        do_convert_rgb (`bool`, *optional*, defaults to `True`):
+            Whether to convert the image to RGB.
+        min_pixels (`int`, *optional*, defaults to `56 * 56`):
+            The min pixels of the image to resize the image.
+        max_pixels (`int`, *optional*, defaults to `28 * 28 * 1280`):
+            The max pixels of the image to resize the image.
+        patch_size (`int`, *optional*, defaults to 14):
+            The spacial patch size of the vision encoder.
+    """
+
+    model_input_names = ["pixel_values", "grid_sizes", "merge_sizes"]
+
+    def __init__(
+        self,
+        do_resize: bool = True,
+        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_convert_rgb: bool = True,
+        min_tokens: int = 4 * 4,
+        max_tokens: int = 16384,
+        patch_size: int = 14,
+        **kwargs,
+    ) -> None:
+        super().__init__(**kwargs)
+        self.do_resize = do_resize
+        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.min_tokens = min_tokens
+        self.max_tokens = max_tokens
+        self.patch_size = patch_size
+        self.do_convert_rgb = do_convert_rgb
+
+    def _preprocess(
+        self,
+        images: Union[ImageInput, VideoInput],
+        target_size: List[int],
+        merge_size: int = 1,
+        do_resize: bool = None,
+        resample: PILImageResampling = None,
+        do_rescale: bool = None,
+        rescale_factor: float = None,
+        do_normalize: bool = None,
+        image_mean: Optional[Union[float, List[float]]] = None,
+        image_std: Optional[Union[float, List[float]]] = None,
+        do_convert_rgb: bool = None,
+        data_format: Optional[ChannelDimension] = ChannelDimension.FIRST,
+        input_data_format: Optional[Union[str, ChannelDimension]] = None,
+    ):
+        """
+        Preprocess an image or batch of images. Copy of the `preprocess` method from `CLIPImageProcessor`.
+
+        Args:
+            images (`ImageInput`):
+                Image or batch of images to preprocess. Expects pixel values ranging from 0 to 255. If pixel values range from 0 to 1, set `do_rescale=False`.
+            target_size (`List[int]`):
+                The target size to resize the image to. Should be a list of two integers: [target_height, target_width].
+            merge_size (`int`, *optional*, defaults to `1`):
+                The merge size after the vision encoder.
+            do_resize (`bool`, *optional*, defaults to `self.do_resize`):
+                Whether to resize the image.
+            resample (`PILImageResampling`, *optional*, defaults to `self.resample`):
+                Resampling filter to use if resizing the image. This can be one of the `PILImageResampling` enums.
+            do_rescale (`bool`, *optional*, defaults to `self.do_rescale`):
+                Whether to rescale the image.
+            rescale_factor (`float`, *optional*, defaults to `self.rescale_factor`):
+                Scale factor to use if rescaling the image.
+            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`):
+                Mean to use if normalizing the image. Can be a float or a list of floats corresponding to the number of channels in the image.
+            image_std (`float` or `List[float]`, *optional*, defaults to `self.image_std`):
+                Standard deviation to use if normalizing the image. Can be a float or a list of floats corresponding to the number of channels in the image.
+            do_convert_rgb (`bool`, *optional*, defaults to `self.do_convert_rgb`):
+                Whether to convert the image to RGB.
+            data_format (`ChannelDimension`, *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. 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.   - `"none"` or `ChannelDimension.NONE`: image in (height, width) format.
+        """
+        images = make_list_of_images(images)
+
+        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 is_scaled_image(images[0]) and do_rescale:
+            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])
+
+        height, width = get_image_size(images[0], channel_dim=input_data_format)
+        resized_height, resized_width = height, width
+        processed_images = []
+        for image in images:
+            if do_resize:
+                resized_height, resized_width = target_size
+                image = resize(
+                    image, size=(resized_height, resized_width), resample=resample, input_data_format=input_data_format
+                )
+
+            if do_rescale:
+                image = self.rescale(image, scale=rescale_factor, input_data_format=input_data_format)
+
+            if do_normalize:
+                image = self.normalize(
+                    image=image, mean=image_mean, std=image_std, input_data_format=input_data_format
+                )
+
+            image = to_channel_dimension_format(image, data_format, input_channel_dim=input_data_format)
+            processed_images.append(image)
+
+        patches = np.array(processed_images)
+        if data_format == ChannelDimension.LAST:
+            patches = patches.transpose(0, 3, 1, 2)
+        t = patches.shape[0]
+        channel = patches.shape[1]
+        grid_h, grid_w = resized_height // self.patch_size, resized_width // self.patch_size
+        patches = patches.reshape(
+            t,
+            channel,
+            grid_h // merge_size,
+            merge_size,
+            self.patch_size,
+            grid_w // merge_size,
+            merge_size,
+            self.patch_size,
+        )
+        patches = patches.transpose(0, 2, 5, 3, 6, 1, 4, 7)
+        flatten_patches = patches.reshape(
+            t * grid_h * grid_w, channel * self.patch_size * self.patch_size
+        )
+
+        return flatten_patches, (t, grid_h, grid_w)
+
+    def preprocess(
+        self,
+        images: ImageInput,
+        do_resize: bool = None,
+        resample: PILImageResampling = None,
+        do_rescale: bool = None,
+        rescale_factor: float = None,
+        do_normalize: bool = None,
+        image_mean: Optional[Union[float, List[float]]] = None,
+        image_std: Optional[Union[float, List[float]]] = None,
+        do_convert_rgb: bool = None,
+        merge_size: Optional[Union[int, List[int]]] = None,
+        return_tensors: Optional[Union[str, TensorType]] = None,
+        data_format: Optional[ChannelDimension] = ChannelDimension.FIRST,
+        input_data_format: Optional[Union[str, ChannelDimension]] = None,
+    ):
+        """
+        Args:
+            images (`ImageInput`):
+                Image to preprocess. Expects a single or batch of images 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.
+            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_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
+        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
+        merge_size = merge_size if merge_size is not None else self.merge_size
+        do_convert_rgb = do_convert_rgb if do_convert_rgb is not None else self.do_convert_rgb
+
+        images = make_batched_images(images)
+
+        if isinstance(merge_size, (list, tuple)):
+            assert len(merge_size) == len(images), "Merge size must be the same length as images."
+            merge_sizes = merge_size
+        else:
+            merge_sizes = [merge_size for _ in images]
+
+        if all(merge_size == merge_sizes[0] for merge_size in merge_sizes):
+            target_sizes = simple_batched_resize(
+                images,
+                factor=self.patch_size * merge_sizes[0],
+                min_tokens=self.min_tokens,
+                max_tokens=self.max_tokens,
+                input_data_format=input_data_format,
+            )
+            # if target_sizes[0][0] != 448:
+            #     for i in range(len(images[0]), 0, -1):
+            #         target_sizes = simple_batched_resize(
+            #             [images[0][:i]],
+            #             factor=self.patch_size * merge_sizes[0],
+            #             min_tokens=self.min_tokens,
+            #             max_tokens=self.max_tokens,
+            #             input_data_format=input_data_format,
+            #         )
+
+            #         if target_sizes[0][0] == 448:
+            #             print(f'Num frames to maintain 448x448: {i}')
+        else:
+            target_sizes = batched_resize(
+                images,
+                factors=[self.patch_size * merge_size for merge_size in merge_sizes],
+                min_tokens=self.min_tokens,
+                max_tokens=self.max_tokens,
+                input_data_format=input_data_format,
+            )
+
+        pixel_values, grid_sizes = [], []
+        for image, merge_size, target_size in zip(images, merge_sizes, target_sizes):
+            patches, grid_size = self._preprocess(
+                image,
+                target_size=target_size,
+                merge_size=merge_size,
+                do_resize=do_resize,
+                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,
+                do_convert_rgb=do_convert_rgb,
+                input_data_format=input_data_format,
+            )
+            pixel_values.append(patches)
+            grid_sizes.append(grid_size)
+
+        pixel_value_shapes = [x.shape for x in pixel_values]
+        pixel_values = np.concatenate(pixel_values, axis=0)
+        grid_sizes = np.array(grid_sizes)
+        merge_sizes = np.array(merge_sizes)
+
+        data = {
+            "pixel_values": pixel_values,
+            "grid_sizes": grid_sizes,
+            "merge_sizes": merge_sizes,
+            "pixel_value_shapes": pixel_value_shapes
+        }
+
+        return BatchFeature(data=data, tensor_type=return_tensors)

model.safetensors

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:45c1890c91ad9b2aea588035bd528c968702ce4f9ff1fcff3b70cd05e16ff7c9
+size 824342816

modeling_viscop_vision_encoder.py

@@ -0,0 +1,661 @@
+# Adopted from https://github.com/DAMO-NLP-SG/VideoLLaMA3/blob/main/videollama3/model/videollama3_encoder/modeling_videollama3_encoder.py
+# Adopted from https://github.com/huggingface/transformers/blob/main/src/transformers/models/qwen2_vl/modeling_qwen2_vl.py.
+# Below is the original copyright:
+# Copyright 2024 The Qwen team, Alibaba Group and the HuggingFace Inc. team. All rights reserved.
+#
+# This code is based on EleutherAI's GPT-NeoX library and the GPT-NeoX
+# and OPT implementations in this library. It has been modified from its
+# original forms to accommodate minor architectural differences compared
+# to GPT-NeoX and OPT used by the Meta AI team that trained the model.
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+"""PyTorch ViSCoP vision encoder model. This file contains the implementation of visual probes and interaction modules"""
+
+import importlib.util
+import os.path as osp
+import math
+import warnings
+
+from einops import rearrange
+
+import torch
+import torch.nn as nn
+import torch.nn.functional as F
+import torch.utils.checkpoint
+from torch.nn.init import _calculate_fan_in_and_fan_out
+
+from transformers.activations import ACT2FN
+from transformers.modeling_utils import PreTrainedModel
+from transformers.utils import is_flash_attn_2_available
+
+if is_flash_attn_2_available():
+    from flash_attn import flash_attn_varlen_func
+else:
+    flash_attn_varlen_func = None
+
+from .configuration_viscop_vision_encoder import ViSCoP_VisionEncoderConfig
+
+# try:
+#     from .configuration_viscop_encoder import ViSCoP_VisionEncoderConfig
+# except ImportError:
+#     spec = importlib.util.spec_from_file_location(
+#         "configuration_videollama3_encoder",
+#         osp.join(osp.dirname(__file__), "configuration_videollama3_encoder.py"),
+#     )
+#     configuration_videollama3_encoder = importlib.util.module_from_spec(spec)
+#     spec.loader.exec_module(configuration_videollama3_encoder)
+#     Videollama3VisionEncoderConfig = getattr(
+#         configuration_videollama3_encoder,
+#         "Videollama3VisionEncoderConfig",
+#     )
+
+
+def _trunc_normal_(tensor, mean, std, a, b):
+    # Cut & paste from PyTorch official master until it's in a few official releases - RW
+    # Method based on https://people.sc.fsu.edu/~jburkardt/presentations/truncated_normal.pdf
+    def norm_cdf(x):
+        # Computes standard normal cumulative distribution function
+        return (1.0 + math.erf(x / math.sqrt(2.0))) / 2.0
+
+    if (mean < a - 2 * std) or (mean > b + 2 * std):
+        warnings.warn(
+            "mean is more than 2 std from [a, b] in nn.init.trunc_normal_. "
+            "The distribution of values may be incorrect.",
+            stacklevel=2,
+        )
+
+    # Values are generated by using a truncated uniform distribution and
+    # then using the inverse CDF for the normal distribution.
+    # Get upper and lower cdf values
+    l = norm_cdf((a - mean) / std)
+    u = norm_cdf((b - mean) / std)
+
+    # Uniformly fill tensor with values from [l, u], then translate to
+    # [2l-1, 2u-1].
+    tensor.uniform_(2 * l - 1, 2 * u - 1)
+
+    # Use inverse cdf transform for normal distribution to get truncated
+    # standard normal
+    tensor.erfinv_()
+
+    # Transform to proper mean, std
+    tensor.mul_(std * math.sqrt(2.0))
+    tensor.add_(mean)
+
+    # Clamp to ensure it's in the proper range
+    tensor.clamp_(min=a, max=b)
+
+
+def trunc_normal_tf_(
+    tensor: torch.Tensor, mean: float = 0.0, std: float = 1.0, a: float = -2.0, b: float = 2.0
+) -> torch.Tensor:
+    """Fills the input Tensor with values drawn from a truncated
+    normal distribution. The values are effectively drawn from the
+    normal distribution :math:`\\mathcal{N}(\text{mean}, \text{std}^2)`
+    with values outside :math:`[a, b]` redrawn until they are within
+    the bounds. The method used for generating the random values works
+    best when :math:`a \\leq \text{mean} \\leq b`.
+
+    NOTE: this 'tf' variant behaves closer to Tensorflow / JAX impl where the
+    bounds [a, b] are applied when sampling the normal distribution with mean=0, std=1.0
+    and the result is subsequently scaled and shifted by the mean and std args.
+
+    Args:
+        tensor: an n-dimensional `torch.Tensor`
+        mean: the mean of the normal distribution
+        std: the standard deviation of the normal distribution
+        a: the minimum cutoff value
+        b: the maximum cutoff value
+    """
+    with torch.no_grad():
+        _trunc_normal_(tensor, 0, 1.0, a, b)
+        tensor.mul_(std).add_(mean)
+
+
+def variance_scaling_(tensor, scale=1.0, mode="fan_in", distribution="normal"):
+    fan_in, fan_out = _calculate_fan_in_and_fan_out(tensor)
+    if mode == "fan_in":
+        denom = fan_in
+    elif mode == "fan_out":
+        denom = fan_out
+    elif mode == "fan_avg":
+        denom = (fan_in + fan_out) / 2
+
+    variance = scale / denom
+
+    if distribution == "truncated_normal":
+        # constant is stddev of standard normal truncated to (-2, 2)
+        trunc_normal_tf_(tensor, std=math.sqrt(variance) / 0.87962566103423978)
+    elif distribution == "normal":
+        with torch.no_grad():
+            tensor.normal_(std=math.sqrt(variance))
+    elif distribution == "uniform":
+        bound = math.sqrt(3 * variance)
+        with torch.no_grad():
+            tensor.uniform_(-bound, bound)
+    else:
+        raise ValueError(f"invalid distribution {distribution}")
+
+
+def lecun_normal_(tensor):
+    variance_scaling_(tensor, mode="fan_in", distribution="truncated_normal")
+
+
+def default_flax_embed_init(tensor):
+    variance_scaling_(tensor, mode="fan_in", distribution="normal")
+
+
+# Copied from transformers.models.llama.modeling_llama.rotate_half
+def rotate_half(x):
+    """Rotates half the hidden dims of the input."""
+    x1 = x[..., : x.shape[-1] // 2]
+    x2 = x[..., x.shape[-1] // 2 :]
+    return torch.cat((-x2, x1), dim=-1)
+
+
+def apply_rotary_pos_emb_vision(tensor: torch.Tensor, freqs: torch.Tensor) -> torch.Tensor:
+    orig_dtype = tensor.dtype
+    tensor = tensor.float()
+    cos = freqs.cos()
+    sin = freqs.sin()
+    cos = cos.unsqueeze(1).repeat(1, 1, 2).unsqueeze(0).float()
+    sin = sin.unsqueeze(1).repeat(1, 1, 2).unsqueeze(0).float()
+    output = (tensor * cos) + (rotate_half(tensor) * sin)
+    output = output.to(orig_dtype)
+    return output
+
+
+class VisionRotaryEmbedding(nn.Module):
+
+    def __init__(self, dim: int, theta: float = 10000.0) -> None:
+        super().__init__()
+        inv_freq = 1.0 / (theta ** (torch.arange(0, dim, 2, dtype=torch.float) / dim))
+        self.register_buffer("inv_freq", inv_freq, persistent=False)
+
+    def forward(self, seqlen: int) -> torch.Tensor:
+        seq = torch.arange(seqlen, device=self.inv_freq.device, dtype=self.inv_freq.dtype)
+        freqs = torch.outer(seq, self.inv_freq)
+        return freqs
+    
+
+class VisionEmbeddings(nn.Module):
+
+    def __init__(self, config: ViSCoP_VisionEncoderConfig):
+        super().__init__()
+        self.config = config
+        self.embed_dim = config.hidden_size
+        self.patch_size = config.patch_size
+
+        self.patch_embedding = nn.Conv2d(
+            in_channels=config.num_channels,
+            out_channels=self.embed_dim,
+            kernel_size=self.patch_size,
+            stride=self.patch_size,
+            padding="valid",
+        )
+
+    def forward(self, hidden_states: torch.Tensor) -> torch.Tensor:
+        hidden_states = hidden_states.view(
+            -1, self.config.num_channels, self.patch_size, self.patch_size
+        )
+        patch_embeds = self.patch_embedding(hidden_states)  # shape = [*, width, grid, grid]
+        # embeddings = patch_embeds.flatten(2).transpose(1, 2)
+        embeddings = patch_embeds.view(-1, self.embed_dim)
+
+        return embeddings
+
+
+class VisionAttention(nn.Module):
+    """Multi-headed attention from 'Attention Is All You Need' paper"""
+
+    # Copied from transformers.models.clip.modeling_clip.CLIPAttention.__init__
+    def __init__(self, config):
+        super().__init__()
+        self.config = config
+        self.embed_dim = config.hidden_size
+        self.num_heads = config.num_attention_heads
+        self.head_dim = self.embed_dim // self.num_heads
+        if self.head_dim * self.num_heads != self.embed_dim:
+            raise ValueError(
+                f"embed_dim must be divisible by num_heads (got `embed_dim`: {self.embed_dim} and `num_heads`:"
+                f" {self.num_heads})."
+            )
+        self.scale = self.head_dim**-0.5
+        self.dropout = config.attention_dropout
+
+        self.k_proj = nn.Linear(self.embed_dim, self.embed_dim)
+        self.v_proj = nn.Linear(self.embed_dim, self.embed_dim)
+        self.q_proj = nn.Linear(self.embed_dim, self.embed_dim)
+        self.out_proj = nn.Linear(self.embed_dim, self.embed_dim)
+
+    def forward(
+        self,
+        hidden_states: torch.Tensor,
+        cu_seqlens: torch.Tensor,
+        rotary_pos_emb: torch.Tensor = None,
+    ) -> torch.Tensor:
+        """Input shape: Time x Channel"""
+
+        q_len, _ = hidden_states.size()
+
+        query_states = self.q_proj(hidden_states)
+        key_states = self.k_proj(hidden_states)
+        value_states = self.v_proj(hidden_states)
+
+        query_states = query_states.view(q_len, self.num_heads, self.head_dim)
+        key_states = key_states.view(q_len, self.num_heads, self.head_dim)
+        value_states = value_states.view(q_len, self.num_heads, self.head_dim)
+
+        query_states = apply_rotary_pos_emb_vision(query_states.unsqueeze(0), rotary_pos_emb).squeeze(0)
+        key_states = apply_rotary_pos_emb_vision(key_states.unsqueeze(0), rotary_pos_emb).squeeze(0)
+
+        attention_mask = torch.zeros([1, q_len, q_len], device=query_states.device, dtype=torch.bool)
+        for i in range(1, len(cu_seqlens)):
+            attention_mask[..., cu_seqlens[i - 1] : cu_seqlens[i], cu_seqlens[i - 1] : cu_seqlens[i]] = True
+
+        query_states = query_states.transpose(0, 1)
+        key_states = key_states.transpose(0, 1)
+        value_states = value_states.transpose(0, 1)
+
+        attn_weights = torch.matmul(query_states, key_states.transpose(1, 2)) / math.sqrt(self.head_dim)
+        attn_weights = attn_weights + attention_mask
+
+        # upcast attention to fp32
+        attn_weights = nn.functional.softmax(attn_weights, dim=-1, dtype=torch.float32).to(query_states.dtype)
+        attn_weights = nn.functional.dropout(attn_weights, p=self.dropout, training=self.training)
+        attn_output = torch.matmul(attn_weights, value_states)
+
+        attn_output = attn_output.transpose(0, 1)
+        attn_output = attn_output.reshape(q_len, -1)
+        attn_output = self.out_proj(attn_output)
+
+        return attn_output
+
+
+class VisionFlashAttention2(VisionAttention):
+
+    def __init__(self, *args, **kwargs):
+        super().__init__(*args, **kwargs)
+
+    # Adapted from transformers.models.llama.modeling_llama.LlamaFlashAttention2.forward
+    def forward(
+        self,
+        hidden_states: torch.Tensor,
+        cu_seqlens: torch.Tensor,
+        rotary_pos_emb: torch.Tensor = None,
+    ) -> torch.Tensor:
+        q_len, _ = hidden_states.size()
+
+        query_states = self.q_proj(hidden_states)
+        key_states = self.k_proj(hidden_states)
+        value_states = self.v_proj(hidden_states)
+
+        # Flash attention requires the input to have the shape
+        # batch_size x seq_length x head_dim x hidden_dim
+        # therefore we just need to keep the original shape
+        query_states = query_states.view(q_len, self.num_heads, self.head_dim)
+        key_states = key_states.view(q_len, self.num_heads, self.head_dim)
+        value_states = value_states.view(q_len, self.num_heads, self.head_dim)
+        query_states = apply_rotary_pos_emb_vision(query_states.unsqueeze(0), rotary_pos_emb).squeeze(0)
+        key_states = apply_rotary_pos_emb_vision(key_states.unsqueeze(0), rotary_pos_emb).squeeze(0)
+        
+        max_seqlen = (cu_seqlens[1:] - cu_seqlens[:-1]).max().item()
+        attn_output = flash_attn_varlen_func(query_states, key_states, value_states, cu_seqlens, cu_seqlens, max_seqlen, max_seqlen).reshape(
+            q_len, -1
+        )
+        attn_output = self.out_proj(attn_output)
+        
+        return attn_output
+    
+class InteractionModule_CrossAttention_FA2(VisionAttention):
+
+    def __init__(self, *args, **kwargs):
+        super().__init__(*args, **kwargs)
+
+    # Adapted from transformers.models.llama.modeling_llama.LlamaFlashAttention2.forward
+    def forward(
+        self,
+        visual_probes: torch.Tensor,
+        hidden_states: torch.Tensor,
+        cu_seqlensq: torch.Tensor,
+        cu_seqlenskv: torch.Tensor,
+        rotary_pos_emb: torch.Tensor = None,
+    ) -> torch.Tensor:
+        q_len, _ = visual_probes.size()
+        kv_len, _ = hidden_states.size()
+
+        query_states = self.q_proj(visual_probes)
+        key_states = self.k_proj(hidden_states)
+        value_states = self.v_proj(hidden_states)
+
+        # Flash attention requires the input to have the shape
+        # batch_size x seq_length x head_dim x hidden_dim
+        # therefore we just need to keep the original shape
+        query_states = query_states.view(q_len, self.num_heads, self.head_dim)
+        key_states = key_states.view(kv_len, self.num_heads, self.head_dim)
+        value_states = value_states.view(kv_len, self.num_heads, self.head_dim)
+        # query_states = apply_rotary_pos_emb_vision(query_states.unsqueeze(0), rotary_pos_emb).squeeze(0)
+        # key_states = apply_rotary_pos_emb_vision(key_states.unsqueeze(0), rotary_pos_emb).squeeze(0)
+
+        max_seqlen_q = (cu_seqlensq[1:] - cu_seqlensq[:-1]).max().item()
+        max_seqlen_kv = (cu_seqlenskv[1:] - cu_seqlenskv[:-1]).max().item()
+
+        attn_output = flash_attn_varlen_func(query_states, key_states, value_states, cu_seqlensq, cu_seqlenskv, max_seqlen_q, max_seqlen_kv).reshape(
+            q_len, -1
+        )
+        attn_output = self.out_proj(attn_output)
+        
+        return attn_output
+
+
+class VisionSdpaAttention(VisionAttention):
+
+    def forward(
+        self,
+        hidden_states: torch.Tensor,
+        cu_seqlens: torch.Tensor,
+        rotary_pos_emb: torch.Tensor = None,
+    ) -> torch.Tensor:
+        seq_length = hidden_states.shape[0]
+        query_states = self.q_proj(hidden_states)
+        key_states = self.k_proj(hidden_states)
+        value_states = self.v_proj(hidden_states)
+
+        query_states = query_states.view(seq_length, self.num_heads, self.head_dim)
+        key_states = key_states.view(seq_length, self.num_heads, self.head_dim)
+        value_states = value_states.view(seq_length, self.num_heads, self.head_dim)
+
+        query_states = apply_rotary_pos_emb_vision(query_states.unsqueeze(0), rotary_pos_emb).squeeze(0)
+        key_states = apply_rotary_pos_emb_vision(key_states.unsqueeze(0), rotary_pos_emb).squeeze(0)
+
+        attention_mask = torch.zeros([1, seq_length, seq_length], device=query_states.device, dtype=torch.bool)
+        for i in range(1, len(cu_seqlens)):
+            attention_mask[..., cu_seqlens[i - 1] : cu_seqlens[i], cu_seqlens[i - 1] : cu_seqlens[i]] = True
+
+        query_states = query_states.transpose(0, 1)
+        key_states = key_states.transpose(0, 1)
+        value_states = value_states.transpose(0, 1)
+        attn_output = F.scaled_dot_product_attention(query_states, key_states, value_states, attention_mask, dropout_p=0.0)
+        attn_output = attn_output.transpose(0, 1)
+        attn_output = attn_output.reshape(seq_length, -1)
+        attn_output = self.out_proj(attn_output)
+        return attn_output
+
+
+VISION_ATTENTION_CLASSES = {
+    "eager": VisionAttention,
+    "flash_attention_2": VisionFlashAttention2,
+    "sdpa": VisionSdpaAttention,
+}
+
+class SigLIPVisionMLP(nn.Module):
+
+    def __init__(self, config):
+        super().__init__()
+        self.config = config
+        self.activation_fn = ACT2FN[config.hidden_act]
+        self.fc1 = nn.Linear(config.hidden_size, config.intermediate_size)
+        self.fc2 = nn.Linear(config.intermediate_size, config.hidden_size)
+
+    def forward(self, hidden_states: torch.Tensor) -> torch.Tensor:
+        hidden_states = self.fc1(hidden_states)
+        hidden_states = self.activation_fn(hidden_states)
+        hidden_states = self.fc2(hidden_states)
+        return hidden_states
+
+
+class SigLIPVisionEncoderLayer(nn.Module):
+
+    def __init__(self, config: ViSCoP_VisionEncoderConfig):
+        super().__init__()
+        self.embed_dim = config.hidden_size
+        self.self_attn = VISION_ATTENTION_CLASSES[config._attn_implementation](config=config)
+        self.layer_norm1 = nn.LayerNorm(self.embed_dim, eps=config.layer_norm_eps)
+        self.mlp = SigLIPVisionMLP(config)
+        self.layer_norm2 = nn.LayerNorm(self.embed_dim, eps=config.layer_norm_eps)
+
+    # Ignore copy
+    def forward(self, hidden_states, cu_seqlens, rotary_pos_emb) -> torch.Tensor:
+        # hidden_states = hidden_states + self.self_attn(
+        #     self.layer_norm1(hidden_states), cu_seqlens=cu_seqlens, rotary_pos_emb=rotary_pos_emb
+        # )
+        hidden_states = hidden_states + self.self_attn(
+            self.layer_norm1(hidden_states), cu_seqlens, rotary_pos_emb
+        )
+        hidden_states = hidden_states + self.mlp(self.layer_norm2(hidden_states))
+        return hidden_states
+
+# Alternative where each interaction module is a copy of the entire vision transformer layer.
+# Empirically we found that simple cross-attention works better.
+class InteractionModule_CrossAttentionEncoderLayer(nn.Module):
+    def __init__(self, config):
+        super().__init__()
+        embed_dim = config.hidden_size
+        self.cross_attn = InteractionModule_CrossAttention_FA2(config)
+
+        # Separate norms for queries and visual features
+        self.q_ln1 = nn.LayerNorm(embed_dim, eps=config.layer_norm_eps)
+        self.kv_ln1 = nn.LayerNorm(embed_dim, eps=config.layer_norm_eps)
+        
+        self.q_ln2 = nn.LayerNorm(embed_dim, eps=config.layer_norm_eps)
+        self.mlp = SigLIPVisionMLP(config)
+
+    def forward(self, visual_probes, hidden_states, query_seqlens, ca_kv_seqlens, rotary_pos_emb):
+        visual_probes = visual_probes + self.cross_attn(
+            self.q_ln1(visual_probes), self.kv_ln1(hidden_states), query_seqlens, ca_kv_seqlens, rotary_pos_emb
+        )
+        visual_probes = visual_probes + self.mlp(self.q_ln2(visual_probes))
+
+        return visual_probes
+
+    
+INTERACTION_MODULES = {
+    'cross_attention': InteractionModule_CrossAttention_FA2,
+    'cross_attention_transformer': InteractionModule_CrossAttentionEncoderLayer
+}
+
+
+class ViSCoP_VisionTransformerEncoder(nn.Module):
+
+    def __init__(self, config: ViSCoP_VisionEncoderConfig):
+        super().__init__()
+        self.config = config
+        head_dim = config.hidden_size // config.num_attention_heads
+        self.rotary_pos_emb = VisionRotaryEmbedding(head_dim // 2)
+        self.layers = nn.ModuleList([SigLIPVisionEncoderLayer(config) for _ in range(config.num_hidden_layers)])
+        self.gradient_checkpointing = False
+        
+        assert config.interaction_module in INTERACTION_MODULES, f'interaction module must be in: {INTERACTION_MODULES.keys()}'
+        num_interaction_modules = config.num_hidden_layers if config.interaction_module_layers is None else len(config.interaction_module_layers)
+        self.interaction_modules = nn.ModuleList([INTERACTION_MODULES[config.interaction_module](config) for _ in range(num_interaction_modules)]) # > Create interaction modules
+        self.interaction_module_layers = [_ for _ in range(config.num_hidden_layers)] if config.interaction_module_layers is None else config.interaction_module_layers
+        self.interaction_module_mapping = {k: v for v, k in enumerate(self.interaction_module_layers)}
+
+        self.num_visual_probes = config.num_visual_probes
+        self.visual_probes = nn.Parameter(torch.zeros(1, self.num_visual_probes, self.config.hidden_size)) # > Create the visual probes
+
+    def rot_pos_emb(self, grid_sizes, merge_sizes):
+        pos_ids = []
+        for (t, h, w), merge_size in zip(grid_sizes, merge_sizes):
+            hpos_ids = torch.arange(h).unsqueeze(1).expand(-1, w)
+            hpos_ids = hpos_ids.reshape(
+                h // merge_size,
+                merge_size,
+                w // merge_size,
+                merge_size,
+            )
+            hpos_ids = hpos_ids.permute(0, 2, 1, 3)
+            hpos_ids = hpos_ids.flatten()
+
+            wpos_ids = torch.arange(w).unsqueeze(0).expand(h, -1)
+            wpos_ids = wpos_ids.reshape(
+                h // merge_size,
+                merge_size,
+                w // merge_size,
+                merge_size,
+            )
+            wpos_ids = wpos_ids.permute(0, 2, 1, 3)
+            wpos_ids = wpos_ids.flatten()
+            pos_ids.append(torch.stack([hpos_ids, wpos_ids], dim=-1).repeat(t, 1))
+
+        pos_ids = torch.cat(pos_ids, dim=0)
+        max_grid_size = grid_sizes[:, 1:].max()
+        rotary_pos_emb_full = self.rotary_pos_emb(max_grid_size)
+        rotary_pos_emb = rotary_pos_emb_full[pos_ids].flatten(1)
+
+        return rotary_pos_emb
+
+    def forward(self, hidden_states, grid_sizes, merge_sizes) -> torch.Tensor:
+        rotary_pos_emb = self.rot_pos_emb(grid_sizes, merge_sizes)
+
+        # mask for self-attention (per-grid attention)
+        cu_seqlens = torch.repeat_interleave(grid_sizes[:, 1] * grid_sizes[:, 2], grid_sizes[:, 0]).cumsum(dim=0, dtype=torch.int32)
+        cu_seqlens = F.pad(cu_seqlens, (1, 0), value=0)
+
+        # > repeat probes
+        batch_size = grid_sizes.shape[0]
+        visual_probes = self.visual_probes.repeat(batch_size, 1, 1)
+        visual_probes = visual_probes.view(batch_size*self.num_visual_probes, self.config.hidden_size)
+
+        # > mask for interaction module (cross-attend probes to all grids)
+        visual_probe_seqlens = torch.tensor([self.num_visual_probes]*batch_size).cumsum(dim=0, dtype=torch.int32)
+        visual_probe_seqlens = F.pad(visual_probe_seqlens, (1, 0), value=0).to(cu_seqlens.device)
+
+        ca_kv_seqlens = (grid_sizes[:, 0] * (grid_sizes[:, 1] * grid_sizes[:, 2])).cumsum(dim=0, dtype=torch.int32)
+        ca_kv_seqlens = F.pad(ca_kv_seqlens, (1, 0), value=0)
+
+        for i, blk in enumerate(self.layers):
+            if self.gradient_checkpointing and self.training:
+                hidden_states = self._gradient_checkpointing_func(
+                    blk.__call__,
+                    hidden_states,
+                    cu_seqlens,
+                    rotary_pos_emb
+                )
+
+                if i in self.interaction_module_layers:
+                    visual_probes = self._gradient_checkpointing_func(
+                        self.interaction_modules[self.interaction_module_mapping[i]].__call__,
+                        visual_probes,
+                        hidden_states,
+                        visual_probe_seqlens,
+                        ca_kv_seqlens,
+                        rotary_pos_emb
+                    )
+            else:
+                hidden_states = blk(hidden_states, cu_seqlens=cu_seqlens, rotary_pos_emb=rotary_pos_emb)
+
+                if i in self.interaction_module_layers:
+                    visual_probes = self.interaction_modules[self.interaction_module_mapping[i]](
+                        visual_probes,
+                        hidden_states,
+                        visual_probe_seqlens,
+                        ca_kv_seqlens,
+                        rotary_pos_emb
+                    )
+
+        return hidden_states, visual_probes
+
+
+class ViSCoP_VisionEncoderModel(PreTrainedModel):
+
+    config_class = ViSCoP_VisionEncoderConfig
+    base_model_prefix = "viscop"
+    main_input_name = "pixel_values"
+    supports_gradient_checkpointing = True
+    _no_split_modules = [
+        "SigLIPVisionEncoderLayer",
+        "VisionEmbeddings",
+    ]
+    _supports_flash_attn_2 = True
+    _supports_sdpa = True
+
+    def __init__(self, config: ViSCoP_VisionEncoderConfig, **kwargs):
+        super().__init__(config=config)
+        embed_dim = config.hidden_size
+        config.num_visual_probes = kwargs['model_cfg'].num_visual_probes
+        config.interaction_module_layers = kwargs['model_cfg'].interaction_module_layers
+        config.interaction_module = kwargs['model_cfg'].interaction_module
+
+        self.embeddings = VisionEmbeddings(config)
+        self.encoder = ViSCoP_VisionTransformerEncoder(config)
+        self.post_layernorm = nn.LayerNorm(embed_dim, eps=config.layer_norm_eps)
+        # if config.interaction_module != 'cross_attention':
+        #     self.post_layernorm_ca = nn.LayerNorm(embed_dim, eps=config.layer_norm_eps)
+
+        self.post_init()
+
+    def forward(self, pixel_values, grid_sizes, merge_sizes=None) -> torch.Tensor:
+        hidden_states = self.embeddings(pixel_values)
+        hidden_states, visual_probes = self.encoder(hidden_states, grid_sizes, merge_sizes)
+        hidden_states = self.post_layernorm(hidden_states)
+        # if self.config.interaction_module != 'cross_attention':
+        #     visual_probes = self.post_layernorm_ca(visual_probes)
+
+        hidden_states_chunks = hidden_states.split(grid_sizes.prod(dim=1).tolist(), dim=0)
+        outputs = []
+
+        for hidden_states, grid_size, merge_size in zip(hidden_states_chunks, grid_sizes, merge_sizes):
+            # NOTE: previous implementation, which supports downsampling with any factor
+            c = hidden_states.shape[-1]
+            hidden_states = hidden_states.view(
+                grid_size[0], grid_size[1] // merge_size, grid_size[2] // merge_size, merge_size, merge_size,  c
+            ).permute(0, 1, 3, 2, 4, 5)
+            hidden_states = hidden_states.reshape(
+                grid_size[0], grid_size[1], grid_size[2], c
+            ).permute(0, 3, 1, 2)
+            hidden_states = torch.nn.functional.interpolate(
+                hidden_states,
+                size=(grid_size[1] // merge_size, grid_size[2] // merge_size),
+                mode='bilinear'
+            )
+            hidden_states = hidden_states.permute(0, 2, 3, 1).view(-1, c)
+
+            # NOTE: simplified implementation, which only supports downsampling with integer factor
+            # NOTE: this implementation is mathematically equivalent to the previous one when merge_size is 1 or 2 but may cause slightly different results
+            # hidden_states = hidden_states.view(-1, merge_size * merge_size, hidden_states.size(-1))
+            # hidden_states = hidden_states.mean(dim=1)
+
+            outputs.append(hidden_states)
+
+        return torch.cat(outputs, dim=0), visual_probes
+
+    def _init_weights(self, module):
+        """Initialize the weights"""
+        if isinstance(module, nn.Embedding):
+            default_flax_embed_init(module.weight)
+        elif isinstance(module, VisionAttention):
+            if isinstance(module, InteractionModule_CrossAttention_FA2): # for deepspeed-zero3
+                return
+            nn.init.xavier_uniform_(module.q_proj.weight)
+            nn.init.xavier_uniform_(module.k_proj.weight)
+            nn.init.xavier_uniform_(module.v_proj.weight)
+            nn.init.xavier_uniform_(module.out_proj.weight)
+            nn.init.zeros_(module.q_proj.bias)
+            nn.init.zeros_(module.k_proj.bias)
+            nn.init.zeros_(module.v_proj.bias)
+            nn.init.zeros_(module.out_proj.bias)
+        elif isinstance(module, SigLIPVisionMLP):
+            nn.init.xavier_uniform_(module.fc1.weight)
+            nn.init.xavier_uniform_(module.fc2.weight)
+            nn.init.normal_(module.fc1.bias, std=1e-6)
+            nn.init.normal_(module.fc2.bias, std=1e-6)
+        elif isinstance(module, (nn.Linear, nn.Conv2d)):
+            lecun_normal_(module.weight)
+            if module.bias is not None:
+                nn.init.zeros_(module.bias)
+        elif isinstance(module, nn.LayerNorm):
+            module.bias.data.zero_()
+            module.weight.data.fill_(1.0)
+        elif isinstance(module, ViSCoP_VisionTransformerEncoder):
+            nn.init.normal_(module.visual_probes, std=.02)

preprocessor_config.json

@@ -0,0 +1,25 @@
+{
+  "auto_map": {
+    "AutoImageProcessor": "image_processing_viscop.ViSCoP_ImageProcessor"
+  },
+  "do_convert_rgb": null,
+  "do_normalize": true,
+  "do_rescale": true,
+  "do_resize": true,
+  "image_mean": [
+    0.5,
+    0.5,
+    0.5
+  ],
+  "image_processor_type": "ViSCoP_ImageProcessor",
+  "image_std": [
+    0.5,
+    0.5,
+    0.5
+  ],
+  "max_tokens": 16384,
+  "min_tokens": 16,
+  "patch_size": 14,
+  "resample": 3,
+  "rescale_factor": 0.00392156862745098
+}