Transfer from pg-team/pg-vision-encoder

README.md

@@ -0,0 +1,87 @@
+<p align="center">
+  <img src="assets/logo.png" width="160" />
+</p>
+
+<h2 align="center">Vision Encoder of PenguinVL</h2>
+<h4 align="center">
+Exploring the Efficiency Limits of VLM with LLM-based Vision Encoders
+</h4>
+
+---
+
+## 📰 News
+
+* **2025.03** — PenguinVL-Encoder now available for general use.
+* **2025.03** — Released PenguinVL-2B, PenguinVL-8B.
+
+---
+
+## 🌟 Model Overview
+
+PenguinVL is a compact Vision-Language Model, designed to explore the efficiency limits of small-scale VLMs.
+
+Unlike most existing VLMs that rely on contrastive-pretrained vision encoders (e.g., CLIP/SigLIP), PG-VL initializes its vision encoder directly from a **text-only LLM**. This design avoids the objective mismatch between contrastive learning and autoregressive language modeling, enabling tighter alignment between visual representations and the language backbone.
+
+### Key Characteristics
+
+- 🧠 **LLM-based Vision Encoder**  
+  The vision encoder is adapted from a pretrained text LLM (Qwen3-0.6B), modified with bidirectional attention and 2D-RoPE for spatial modeling.  
+  This provides strong semantic priors and native compatibility with the downstream LLM.
+
+- 🎥 **Efficient Video Understanding**  
+  A Temporal Redundancy-Aware (TRA) token compression strategy dynamically allocates token budgets across frames, enabling long-video reasoning within a limited context window.
+
+- 🏗 Unified Architecture  
+  The model consists of:
+  1. LLM-initialized vision encoder  
+  2. Lightweight MLP projector  
+  3. Qwen3 language backbone  
+
+- 📊 Compact but Strong  
+  At 2B scale, PG-VL achieves competitive performance across image, document, OCR, math, and video benchmarks while remaining deployment-friendly.
+
+---
+
+## 🧪 Quick Start — Transformers Inference
+
+```python
+import torch
+from transformers import AutoModel, AutoImageProcessor
+from transformers.image_utils import load_image
+
+model_name = "pg-team/pg-vision-encoder"
+image_path = "xxx"
+images = load_image(image_path)
+
+model = AutoModel.from_pretrained(
+    model_name,
+    trust_remote_code=True,
+    device_map="auto",
+    torch_dtype=torch.bfloat16,
+    attn_implementation="flash_attention_2",
+)
+processor = AutoImageProcessor.from_pretrained(model_name, trust_remote_code=True)
+
+inputs = processor(images=images, merge_size=1)
+inputs = {k: torch.tensor(v).cuda() for k, v in inputs.items()}
+if "pixel_values" in inputs:
+    inputs["pixel_values"] = inputs["pixel_values"].to(torch.bfloat16)
+image_features = model(**inputs)
+```
+
+## 🌎 Model Zoo
+| Model                | Base Model   | HF Link                                                      |
+| -------------------- | ------------ | ------------------------------------------------------------ |
+| PenguinVL-8B         | Qwen3-8B     | [pg-team/pg-vl-8b-hf](https://huggingface.co/pg-team/pg-vl-8b-hf) |
+| PenguinVL-2B         | Qwen3-1.7B   | [pg-team/pg-vl-2b-hf](https://huggingface.co/pg-team/pg-vl-2b-hf) |
+| PenguinVL-Encoder    | Qwen3-0.6B   | [pg-team/pg-vision-encoder](https://huggingface.co/pg-team/pg-vision-encoder) |
+
+## 🚀 Main Results
+xxx
+
+## Citation
+
+If you find PenguinVL useful for your research and applications, please cite using this BibTeX:
+```bibtex
+...
+```

config.json

@@ -0,0 +1,37 @@
+{
+  "architectures": [
+    "PenguinVLVisionEncoderModel"
+  ],
+  "auto_map": {
+    "AutoConfig": "configuration_penguinvl_encoder.PenguinVLVisionEncoderConfig",
+    "AutoModel": "modeling_penguinvl_encoder.PenguinVLVisionEncoderModel"
+  },
+  "attention_bias": false,
+  "attention_dropout": 0.0,
+  "bos_token_id": 151643,
+  "eos_token_id": 151645,
+  "head_dim": 128,
+  "hidden_act": "silu",
+  "hidden_size": 1024,
+  "initializer_range": 0.02,
+  "intermediate_size": 3072,
+  "layer_norm_eps": 1e-06,
+  "max_position_embeddings": 40960,
+  "max_window_layers": 28,
+  "model_type": "penguinvl_vision_encoder",
+  "num_attention_heads": 16,
+  "num_channels": 3,
+  "num_hidden_layers": 28,
+  "num_key_value_heads": 8,
+  "patch_size": 14,
+  "rms_norm_eps": 1e-06,
+  "rope_scaling": null,
+  "rope_theta": 1000000,
+  "sliding_window": null,
+  "tie_word_embeddings": true,
+  "torch_dtype": "bfloat16",
+  "transformers_version": "4.51.3",
+  "use_cache": true,
+  "use_sliding_window": false,
+  "vocab_size": 151936
+}

configuration_penguinvl_encoder.py

@@ -0,0 +1,33 @@
+"""PenguinVL vision encoder model configuration."""
+
+from transformers import Qwen3Config
+
+
+class PenguinVLVisionEncoderConfig(Qwen3Config):
+
+    model_type = "penguinvl_vision_encoder"
+
+    def __init__(
+        self,
+        hidden_size=1536,
+        intermediate_size=8960,
+        num_hidden_layers=12,
+        num_attention_heads=12,
+        num_channels=3,
+        patch_size=14,
+        layer_norm_eps=1e-6,
+        attention_dropout=0.0,
+        num_key_value_heads=2,
+        **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.num_key_value_heads = num_key_value_heads
+        self.layer_norm_eps = layer_norm_eps

image_processing_penguinvl.py

@@ -0,0 +1,548 @@
+# 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 PenguinVL."""
+
+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,
+    get_image_size,
+    infer_channel_dimension_format,
+    is_scaled_image,
+    is_valid_image,
+    make_list_of_images,
+    to_numpy_array,
+)
+try:
+    from transformers.image_utils import VideoInput
+except:
+    from transformers.video_utils import VideoInput
+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]]:
+    """
+    Normalize visual inputs to ``List[List[ImageInput]]`` – a list of *clips*,
+    where each clip is a list of frames.
+
+    Supported input formats::
+
+        Nested clips : [[image], [f1, f2, ...], ...]   → returned as-is
+        Flat frames  : [f1, f2, ...]                    → [[f1, f2, ...]]
+        Single image : image                            → [[image]]
+
+    Returns:
+        List of clips, where each clip is a list of valid images / frames.
+    """
+    if isinstance(images, (list, tuple)) and len(images) > 0:
+        if isinstance(images[0], (list, tuple)):
+            return [list(clip) for clip in images]
+        if all(is_valid_image(f) for f in images):
+            return [list(images)]
+    if is_valid_image(images):
+        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, 
+    frame_types=None
+):
+    """
+    Compute per-frame target (h, w) for a video frame list under a token budget (key/intermediate may differ).
+
+    Uses the Temporal Redundancy-Aware (TRA) token compression strategy: key and intermediate frames
+    can have different target areas (e.g. 1:16 ratio when compressing) to stay within max_tokens.
+
+    Args:
+        images: List of video frames (each PIL Image or ndarray).
+        factor: Alignment granularity (height and width are multiples of factor), default 28.
+        min_tokens: Minimum tokens per frame (used to derive min_pixels), default 16.
+        max_tokens: Token cap for total pixel budget, default 16384.
+        input_data_format: Channel format when not PIL, e.g. "channels_first".
+        frame_types: Per-frame type list, 0=key, 1=intermediate; None means all key.
+
+    Returns:
+        image_sizes: List of (h, w) per frame, one-to-one with images.
+    """
+    min_pixels = min_tokens * factor * factor * 1.5
+    max_pixels = max_tokens * factor * factor * 0.95
+
+    # --- Base info ---
+    first_image = images[0]
+    if isinstance(first_image, Image.Image):
+        width, height = first_image.size
+    else:
+        height, width = get_image_size(first_image, channel_dim=input_data_format)
+
+    aspect_ratio = height / width
+    raw_area = height * width
+
+    num_frames = len(images)
+    if frame_types is not None:
+        ft_list = frame_types.tolist() if hasattr(frame_types, 'tolist') else frame_types
+        num_intermediate = ft_list.count(1)
+        num_key = ft_list.count(0)
+    else:
+        num_key = num_frames
+        num_intermediate = 0
+        ft_list = [0] * num_frames
+
+    def get_dims_from_area(target_area, ar, fac):
+        """Compute aligned (h, w) from target area and aspect ratio; area = w²·ar => w = sqrt(area/ar)."""
+        w_new = math.sqrt(target_area / ar)
+        h_new = w_new * ar
+
+        h_bar = round(h_new / fac) * fac
+        w_bar = round(w_new / fac) * fac
+        h_bar = max(h_bar, fac)
+        w_bar = max(w_bar, fac)
+
+        return h_bar, w_bar
+
+    # --- Stage 1: No-downscale check ---
+    # If total pixels within budget, keep original size for both key and intermediate frames.
+    total_raw_pixels = num_frames * raw_area
+    target_key_area = raw_area
+    target_intermediate_area = raw_area
+
+    if total_raw_pixels > max_pixels:
+        # --- Stage 2: Sync compression ---
+        # Over budget: compress with 1:16 area ratio, intermediate_area = key_area / 16.
+        # Constraint: N_key·A_key + N_intermediate·(A_key/16) = max_pixels => A_key = max_pixels / (N_key + N_intermediate/16).
+        effective_count = num_key + (num_intermediate / 16.0)
+        calc_key_area = max_pixels / effective_count
+        calc_intermediate_area = calc_key_area / 16.0
+
+        # --- Stage 3: Intermediate-frame floor ---
+        # If computed intermediate area is below min_pixels, pin intermediate to min_pixels and give remaining budget to key.
+        if calc_intermediate_area >= min_pixels:
+            target_key_area = calc_key_area
+            target_intermediate_area = calc_intermediate_area
+        else:
+            target_intermediate_area = min_pixels
+            pixels_taken_by_intermediate = num_intermediate * min_pixels
+            remaining_for_key = max_pixels - pixels_taken_by_intermediate
+            target_key_area = remaining_for_key / num_key
+
+        # --- Stage 4: Key-frame hard floor ---
+        if target_key_area < min_pixels:
+            target_key_area = min_pixels
+
+    # --- Area to aligned dimensions ---
+    k_h, k_w = get_dims_from_area(target_key_area, aspect_ratio, factor)
+    if num_intermediate > 0:
+        i_h, i_w = get_dims_from_area(target_intermediate_area, aspect_ratio, factor)
+    else:
+        i_h, i_w = 0, 0
+
+    def ensure_min_hw(h, w, min_p, raw_ar):
+        """If area still below min_pixels after alignment (rounding), recompute from min area and align upward."""
+        if h * w < min_p:
+            w = math.sqrt(min_p / raw_ar)
+            h = w * raw_ar
+            h = math.ceil(h / factor) * factor
+            w = math.ceil(w / factor) * factor
+        return h, w
+
+    k_h, k_w = ensure_min_hw(k_h, k_w, min_pixels, aspect_ratio)
+    if num_intermediate > 0:
+        i_h, i_w = ensure_min_hw(i_h, i_w, min_pixels, aspect_ratio)
+
+    image_sizes = [
+        (i_h, i_w) if ft_list[i] == 1 else (k_h, k_w)
+        for i in range(num_frames)
+    ]
+    return image_sizes
+
+
+class PenguinVLImageProcessor(BaseImageProcessor):
+    r"""
+    Constructs a PenguinVL 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 _allocate_token_budget(self, clips, clip_merge_sizes, input_data_format):
+        """Distribute self.max_tokens across clips proportionally to their raw token counts."""
+        clip_raw_tokens = []
+        for clip, ms in zip(clips, clip_merge_sizes):
+            first_frame = clip[0]
+            if isinstance(first_frame, Image.Image):
+                w, h = first_frame.size
+            else:
+                h, w = get_image_size(first_frame, channel_dim=input_data_format)
+            factor = self.patch_size * ms
+            clip_raw_tokens.append(len(clip) * h * w / (factor * factor))
+
+        total_raw_tokens = sum(clip_raw_tokens)
+        if total_raw_tokens <= self.max_tokens:
+            return [self.max_tokens] * len(clips)
+
+        return [
+            max(self.min_tokens * len(clip), raw * self.max_tokens / total_raw_tokens)
+            for clip, raw in zip(clips, clip_raw_tokens)
+        ]
+
+    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,
+        frame_types: 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
+
+        clips = make_batched_images(images)
+        num_clips = len(clips)
+
+        if isinstance(merge_size, (list, tuple)):
+            assert len(merge_size) == num_clips, (
+                f"merge_size length ({len(merge_size)}) must match number of clips ({num_clips})"
+            )
+            clip_merge_sizes = list(merge_size)
+        else:
+            clip_merge_sizes = [merge_size] * num_clips
+
+        if frame_types is None:
+            clip_frame_types = [None] * num_clips
+        elif isinstance(frame_types, (list, tuple)) and len(frame_types) > 0:
+            if isinstance(frame_types[0], (list, tuple)) or frame_types[0] is None:
+                assert len(frame_types) == num_clips, (
+                    f"frame_types length ({len(frame_types)}) must match number of clips ({num_clips})"
+                )
+                clip_frame_types = list(frame_types)
+            else:
+                assert num_clips == 1, "Flat frame_types is only supported for a single clip"
+                clip_frame_types = [frame_types]
+        else:
+            clip_frame_types = [None] * num_clips
+
+        pixel_values, grid_sizes, per_frame_merge_sizes = [], [], []
+
+        clip_max_tokens_list = self._allocate_token_budget(
+            clips, clip_merge_sizes, input_data_format,
+        )
+
+        for clip, ms, ft, clip_max_tokens in zip(clips, clip_merge_sizes, clip_frame_types, clip_max_tokens_list):
+            target_sizes = simple_batched_resize(
+                clip,
+                factor=self.patch_size * ms,
+                min_tokens=self.min_tokens,
+                max_tokens=clip_max_tokens,
+                input_data_format=input_data_format,
+                frame_types=ft,
+            )
+
+            for frame, target_size in zip(clip, target_sizes):
+                patches, grid_size = self._preprocess(
+                    frame,
+                    target_size=target_size,
+                    merge_size=ms,
+                    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)
+                per_frame_merge_sizes.append(ms)
+
+        pixel_values = np.concatenate(pixel_values, axis=0)
+        grid_sizes = np.array(grid_sizes)
+        merge_sizes = np.array(per_frame_merge_sizes)
+
+        data = {
+            "pixel_values": pixel_values,
+            "grid_sizes": grid_sizes,
+            "merge_sizes": merge_sizes,
+        }
+
+        return BatchFeature(data=data, tensor_type=return_tensors)

model.safetensors

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:8c12c1beb59b8437b833884ab866628f5ddcf8fc83e3da2f7e10a9189c8d8aec
+size 882176944

modeling_penguinvl_encoder.py

@@ -0,0 +1,548 @@
+from torch import nn
+import torch
+import math
+import warnings
+from functools import partial
+from .configuration_penguinvl_encoder import PenguinVLVisionEncoderConfig
+from transformers.modeling_utils import PreTrainedModel
+from transformers.models.qwen3.modeling_qwen3 import Qwen3Model, Qwen3Attention, rotate_half, Qwen3DecoderLayer
+from typing import List, Optional, Tuple, Union
+from transformers.modeling_outputs import BaseModelOutputWithPast
+from transformers.processing_utils import Unpack
+from transformers.modeling_flash_attention_utils import FlashAttentionKwargs
+from transformers.cache_utils import Cache, DynamicCache
+from transformers.utils import logging, is_flash_attn_greater_or_equal_2_10, is_flash_attn_2_available
+from transformers.modeling_rope_utils import ROPE_INIT_FUNCTIONS
+from torch.nn.init import _calculate_fan_in_and_fan_out
+import torch.nn.functional as F
+if is_flash_attn_2_available():
+    from transformers.modeling_flash_attention_utils import _flash_attention_forward
+    from flash_attn import flash_attn_varlen_func
+
+logger = logging.get_logger(__name__)
+
+class PenguinVLVisionEncoderEmbeddings(nn.Module):
+
+    def __init__(self, config: PenguinVLVisionEncoderConfig):
+        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)
+        embeddings = patch_embeds.view(-1, self.embed_dim)
+
+        return embeddings
+    
+
+# Adapted from Qwen2VLRotaryEmbedding in transformers/models/qwen2/modeling_qwen2.py
+class VisualRotaryEmbedding(nn.Module):
+    def __init__(
+        self,
+        dim=None,
+        max_position_embeddings=2048,
+        base=10000,
+        device=None,
+        scaling_factor=1.0,
+        rope_type="default",
+        config = None,
+    ):
+        super().__init__()
+        # TODO (joao): remove the `if` below, only used for BC
+        self.rope_kwargs = {}
+        if config is None:
+            logger.warning_once(
+                "`Qwen2VLRotaryEmbedding` can now be fully parameterized by passing the model config through the "
+                "`config` argument. All other arguments will be removed in v4.46"
+            )
+            self.rope_kwargs = {
+                "rope_type": rope_type,
+                "factor": scaling_factor,
+                "dim": dim,
+                "base": base,
+                "max_position_embeddings": max_position_embeddings,
+            }
+            self.rope_type = rope_type
+            self.max_seq_len_cached = max_position_embeddings
+            self.original_max_seq_len = max_position_embeddings
+        else:
+            # BC: "rope_type" was originally "type"
+            if config.rope_scaling is not None:
+                self.rope_type = config.rope_scaling.get("rope_type", config.rope_scaling.get("type"))
+            else:
+                self.rope_type = "default"
+            self.max_seq_len_cached = config.max_position_embeddings
+            self.original_max_seq_len = config.max_position_embeddings
+
+        self.config = config
+        self.rope_init_fn = ROPE_INIT_FUNCTIONS[self.rope_type]
+
+        inv_freq, self.attention_scaling = self.rope_init_fn(self.config, device, **self.rope_kwargs)
+        self.register_buffer("inv_freq", inv_freq, persistent=False)
+        self.original_inv_freq = self.inv_freq
+
+    def _dynamic_frequency_update(self, position_ids, device):
+        """
+        dynamic RoPE layers should recompute `inv_freq` in the following situations:
+        1 - growing beyond the cached sequence length (allow scaling)
+        2 - the current sequence length is in the original scale (avoid losing precision with small sequences)
+        """
+        seq_len = torch.max(position_ids) + 1
+        if seq_len > self.max_seq_len_cached:  # growth
+            inv_freq, self.attention_scaling = self.rope_init_fn(
+                self.config, device, seq_len=seq_len, **self.rope_kwargs
+            )
+            self.register_buffer("inv_freq", inv_freq, persistent=False)  # TODO joao: may break with compilation
+            self.max_seq_len_cached = seq_len
+
+        if seq_len < self.original_max_seq_len and self.max_seq_len_cached > self.original_max_seq_len:  # reset
+            self.register_buffer("inv_freq", self.original_inv_freq, persistent=False)
+            self.max_seq_len_cached = self.original_max_seq_len
+
+    @torch.no_grad()
+    def forward(self, x, position_ids):
+        if "dynamic" in self.rope_type:
+            self._dynamic_frequency_update(position_ids, device=x.device)
+
+        inv_freq_expanded = self.inv_freq[None, None, :, None].float().expand(2, position_ids.shape[1], -1, 1)
+        position_ids_expanded = position_ids[:, :, None, :].float()  # shape (2, bs, 1, positions)
+        # Force float32 (see https://github.com/huggingface/transformers/pull/29285)
+        device_type = x.device.type
+        device_type = device_type if isinstance(device_type, str) and device_type != "mps" else "cpu"
+        with torch.autocast(device_type=device_type, enabled=False):
+            freqs = (inv_freq_expanded.float() @ position_ids_expanded.float()).transpose(2, 3)
+            emb = torch.cat((freqs, freqs), dim=-1)
+            cos = emb.cos()
+            sin = emb.sin()
+
+        # Advanced RoPE types (e.g. yarn) apply a post-processing scaling factor, equivalent to scaling attention
+        cos = cos * self.attention_scaling
+        sin = sin * self.attention_scaling
+
+        return cos.to(dtype=x.dtype), sin.to(dtype=x.dtype)
+
+
+def apply_multimodal_rotary_pos_emb(q, k, cos, sin, unsqueeze_dim=1):
+    rope_section = [cos.shape[-1] // 2, cos.shape[-1] // 2]
+    cos = torch.cat([m[i % 2] for i, m in enumerate(cos.split(rope_section, dim=-1))], dim=-1).unsqueeze(unsqueeze_dim)
+    sin = torch.cat([m[i % 2] for i, m in enumerate(sin.split(rope_section, dim=-1))], dim=-1).unsqueeze(unsqueeze_dim)
+
+    q_embed = (q * cos) + (rotate_half(q) * sin)
+    k_embed = (k * cos) + (rotate_half(k) * sin)
+    return q_embed, k_embed
+
+
+class PenguinVLAttention(Qwen3Attention):
+    """Multi-headed attention from 'Attention Is All You Need' paper"""
+
+    # Copied from transformers.models.llama.modeling_llama.LlamaFlashAttention2.__init__
+    def __init__(self, *args, **kwargs):
+        super().__init__(*args, **kwargs)
+        self.is_causal = False
+
+    def forward(
+        self,
+        hidden_states: torch.Tensor,
+        position_embeddings: Tuple[torch.Tensor, torch.Tensor],
+        attention_mask: Optional[torch.Tensor],
+        past_key_value: Optional[Cache] = None,
+        cache_position: Optional[torch.LongTensor] = None,
+        cu_seqlens: Optional[torch.Tensor] = None,
+        **kwargs: Unpack[FlashAttentionKwargs],
+    ) -> Tuple[torch.Tensor, Optional[torch.Tensor], Optional[Tuple[torch.Tensor]]]:
+        input_shape = hidden_states.shape[:-1]
+        hidden_shape = (*input_shape, -1, self.head_dim)
+
+        query_states = self.q_norm(self.q_proj(hidden_states).view(hidden_shape)).transpose(1, 2)
+        key_states = self.k_norm(self.k_proj(hidden_states).view(hidden_shape)).transpose(1, 2)
+        value_states = self.v_proj(hidden_states).view(hidden_shape).transpose(1, 2)
+
+        cos, sin = position_embeddings
+        query_states, key_states = apply_multimodal_rotary_pos_emb(query_states, key_states, cos, sin)
+
+        if past_key_value is not None:
+            # sin and cos are specific to RoPE models; cache_position needed for the static cache
+            cache_kwargs = {"sin": sin, "cos": cos, "cache_position": cache_position}
+            key_states, value_states = past_key_value.update(key_states, value_states, self.layer_idx, cache_kwargs)
+
+        # This is before the transpose
+        seq_len = query_states.shape[2]
+
+        # In PEFT, usually we cast the layer norms in float32 for training stability reasons
+        # therefore the input hidden states gets silently casted in float32. Hence, we need
+        # cast them back in the correct dtype just to be sure everything works as expected.
+        # This might slowdown training & inference so it is recommended to not cast the LayerNorms
+        # in fp32. (usually our RMSNorm modules handle it correctly)
+        target_dtype = None
+        if query_states.dtype == torch.float32:
+            if torch.is_autocast_enabled():
+                target_dtype = torch.get_autocast_gpu_dtype()
+            # Handle the case where the model is quantized
+            elif hasattr(self.config, "_pre_quantization_dtype"):
+                target_dtype = self.config._pre_quantization_dtype
+            else:
+                target_dtype = next(layer for layer in self.modules() if isinstance(layer, torch.nn.Linear)).weight.dtype
+
+        # FA2 always relies on the value set in the module, so remove it if present in kwargs to avoid passing it twice
+        kwargs.pop("is_causal", None)
+
+        # Reashape to the expected shape for Flash Attention
+        query_states = query_states.transpose(1, 2).squeeze(0)
+        key_states = key_states.transpose(1, 2).squeeze(0)
+        value_states = value_states.transpose(1, 2).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_q=cu_seqlens,
+            cu_seqlens_k=cu_seqlens,
+            max_seqlen_q=max_seqlen,
+            max_seqlen_k=max_seqlen,
+            dropout_p=0.0 if not self.training else self.attention_dropout,
+            causal=self.is_causal
+        )
+
+        attn_output = attn_output.reshape(*input_shape, -1).contiguous()
+        attn_output = self.o_proj(attn_output)
+        return attn_output, None
+    
+
+class PenguinVLDecoderLayer(Qwen3DecoderLayer):
+    def __init__(self, config: PenguinVLVisionEncoderConfig, layer_idx: int):
+        super(PenguinVLDecoderLayer, self).__init__(config, layer_idx)
+        self.self_attn = PenguinVLAttention(config, layer_idx)
+
+    def forward(
+        self,
+        hidden_states: torch.Tensor,
+        attention_mask: Optional[torch.Tensor] = None,
+        position_ids: Optional[torch.LongTensor] = None,
+        past_key_value: Optional[Cache] = None,
+        output_attentions: Optional[bool] = False,
+        use_cache: Optional[bool] = False,
+        cache_position: Optional[torch.LongTensor] = None,
+        position_embeddings: Optional[Tuple[torch.Tensor, torch.Tensor]] = None,  # necessary, but kept here for BC
+        cu_seqlens: Optional[torch.Tensor] = None,
+        **kwargs: Unpack[FlashAttentionKwargs],
+    ) -> Tuple[torch.FloatTensor, Optional[Tuple[torch.FloatTensor, torch.FloatTensor]]]:
+        residual = hidden_states
+
+        hidden_states = self.input_layernorm(hidden_states)
+
+        # Self Attention
+        hidden_states, self_attn_weights = self.self_attn(
+            hidden_states=hidden_states,
+            attention_mask=attention_mask,
+            position_ids=position_ids,
+            past_key_value=past_key_value,
+            output_attentions=output_attentions,
+            use_cache=use_cache,
+            cache_position=cache_position,
+            position_embeddings=position_embeddings,
+            cu_seqlens=cu_seqlens,
+            **kwargs,
+        )
+        hidden_states = residual + hidden_states
+
+        # Fully Connected
+        residual = hidden_states
+        hidden_states = self.post_attention_layernorm(hidden_states)
+        hidden_states = self.mlp(hidden_states)
+        hidden_states = residual + hidden_states
+
+        outputs = (hidden_states,)
+        if output_attentions:
+            outputs += (self_attn_weights,)
+
+        return outputs
+    
+    
+class PenguinVLVisionEncoderFromQwen3Model(Qwen3Model):
+    def __init__(self, config: PenguinVLVisionEncoderConfig):
+        super().__init__(config)
+        self.layers = nn.ModuleList(
+            [PenguinVLDecoderLayer(config, layer_idx) for layer_idx in range(config.num_hidden_layers)]
+        )
+        self.rotary_emb = VisualRotaryEmbedding(config=config)
+        del self.embed_tokens
+
+    @staticmethod
+    def _prepare_4d_causal_attention_mask_with_cache_position(
+        attention_mask: torch.Tensor,
+        sequence_length: int,
+        target_length: int,
+        dtype: torch.dtype,
+        device: torch.device,
+        cache_position: torch.Tensor,
+        batch_size: int,
+        config: PenguinVLVisionEncoderConfig,
+        past_key_values: Cache,
+    ):
+        """
+        Override the original causal mask method to create full attention mask instead.
+        Creates a full attention 4D mask of shape `(batch_size, 1, query_length, key_value_length)` 
+        from a 2D mask of shape `(batch_size, key_value_length)`.
+        
+        For vision encoding, we want full attention between all patches, not causal attention.
+        """
+        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.
+            full_attention_mask = attention_mask
+        else:
+            # Create full attention mask (all zeros, meaning attend to all positions)
+            # We only mask based on the provided attention_mask for padding
+            if attention_mask is not None:
+                # Use the provided attention_mask to handle padding
+                min_dtype = torch.finfo(dtype).min
+                full_attention_mask = torch.zeros(
+                    (sequence_length, target_length), dtype=dtype, device=device
+                )
+                # Expand to 4D
+                full_attention_mask = full_attention_mask[None, None, :, :].expand(batch_size, 1, -1, -1)
+                
+                # Apply padding mask if provided
+                full_attention_mask = full_attention_mask.clone()  # copy to contiguous memory for in-place edit
+                if attention_mask.shape[-1] > target_length:
+                    attention_mask = attention_mask[:, :target_length]
+                mask_length = attention_mask.shape[-1]
+                padding_mask = attention_mask[:, None, None, :] == 0
+                full_attention_mask[:, :, :, :mask_length] = full_attention_mask[:, :, :, :mask_length].masked_fill(
+                    padding_mask, min_dtype
+                )
+            else:
+                # No attention mask provided, create all-zeros mask (full attention)
+                full_attention_mask = torch.zeros(
+                    (batch_size, 1, sequence_length, target_length), dtype=dtype, device=device
+                )
+        return full_attention_mask
+
+    def get_rope_index(self, grid_sizes, merge_sizes, position_ids):
+        position_ids = position_ids.contiguous()
+        batch_size = grid_sizes.shape[0]
+
+        # Vision Part: Generate 2D position indices for vision tokens
+        vision_pos_ids = []
+        for (t, h, w), merge_size in zip(grid_sizes, merge_sizes):
+            # Generate height position indices
+            hpos_ids = torch.arange(h).unsqueeze(1).expand(-1, w).to(position_ids.device)
+            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()
+
+            # Generate width position indices  
+            wpos_ids = torch.arange(w).unsqueeze(0).expand(h, -1).to(position_ids.device)
+            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()
+            
+            # Stack height and width to create 2D positions
+            vision_pos_ids.append(torch.stack([hpos_ids, wpos_ids], dim=-1).repeat(t, 1))
+
+        num_start_idx = 0
+        for batch_idx in range(batch_size):
+            pos_len = vision_pos_ids[batch_idx].shape[0]
+            position_ids[:, 0, num_start_idx: num_start_idx+pos_len] = vision_pos_ids[batch_idx].permute(1, 0)
+            num_start_idx += pos_len
+        
+        return position_ids
+
+
+    def forward(
+        self,
+        input_ids: 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,
+        use_cache: Optional[bool] = None,
+        output_attentions: Optional[bool] = None,
+        output_hidden_states: Optional[bool] = None,
+        cache_position: Optional[torch.LongTensor] = None,
+        grid_sizes: Optional[torch.Tensor] = None,
+        merge_sizes: Optional[torch.Tensor] = None,
+        **flash_attn_kwargs: Unpack[FlashAttentionKwargs],
+    ) -> BaseModelOutputWithPast:
+        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
+        )
+        use_cache = use_cache if use_cache is not None else self.config.use_cache
+
+        if (input_ids is None) ^ (inputs_embeds is not None):
+            raise ValueError("You must specify exactly one of input_ids or inputs_embeds")
+
+        if self.gradient_checkpointing and self.training and use_cache:
+            logger.warning_once(
+                "`use_cache=True` is incompatible with gradient checkpointing. Setting `use_cache=False`."
+            )
+            use_cache = False
+
+        # TODO (joao): remove this exception in v4.56 -- it exists for users that try to pass a legacy cache
+        if not isinstance(past_key_values, (type(None), Cache)):
+            raise ValueError("The `past_key_values` should be either a `Cache` object or `None`.")
+
+        if inputs_embeds is None:
+            inputs_embeds = self.embed_tokens(input_ids)
+
+        if use_cache and past_key_values is None:
+            past_key_values = DynamicCache()
+
+        if cache_position is None:
+            past_seen_tokens = past_key_values.get_seq_length() if past_key_values is not None else 0
+            cache_position = torch.arange(
+                past_seen_tokens, past_seen_tokens + inputs_embeds.shape[1], device=inputs_embeds.device
+            )
+
+        # the hard coded `2` is for temporal, height and width.
+        if position_ids is None:
+            position_ids = cache_position.view(1, 1, -1).expand(2, inputs_embeds.shape[0], -1)
+        elif position_ids.dim() == 2:
+            position_ids = position_ids[None, ...].expand(2, position_ids.shape[0], -1)
+        position_ids = self.get_rope_index(grid_sizes, merge_sizes, position_ids)
+
+        causal_mask = None
+
+        hidden_states = inputs_embeds
+
+        # create position embeddings to be shared across the decoder layers
+        position_embeddings = self.rotary_emb(hidden_states, position_ids)
+
+        # decoder layers
+        all_hidden_states = () if output_hidden_states else None
+        all_self_attns = () if output_attentions else None
+
+        # Calculate cumulative sequence lengths for the grid sizes
+        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)
+
+        for decoder_layer in self.layers[: self.config.num_hidden_layers]:
+            if output_hidden_states:
+                all_hidden_states += (hidden_states,)
+
+            if self.gradient_checkpointing and self.training:
+                layer_outputs = self._gradient_checkpointing_func(
+                    partial(decoder_layer.__call__, **flash_attn_kwargs),
+                    hidden_states,
+                    causal_mask,
+                    position_ids,
+                    past_key_values,
+                    output_attentions,
+                    use_cache,
+                    cache_position,
+                    position_embeddings,
+                    cu_seqlens,
+                )
+            else:
+                layer_outputs = decoder_layer(
+                    hidden_states,
+                    attention_mask=causal_mask,
+                    position_ids=position_ids,
+                    past_key_value=past_key_values,
+                    output_attentions=output_attentions,
+                    use_cache=use_cache,
+                    cache_position=cache_position,
+                    position_embeddings=position_embeddings,
+                    cu_seqlens=cu_seqlens,
+                    **flash_attn_kwargs,
+                )
+
+            hidden_states = layer_outputs[0]
+
+            if output_attentions:
+                all_self_attns += (layer_outputs[1],)
+
+        hidden_states = self.norm(hidden_states)
+
+        # add hidden states from the last decoder layer
+        if output_hidden_states:
+            all_hidden_states += (hidden_states,)
+
+        return BaseModelOutputWithPast(
+            last_hidden_state=hidden_states,
+            past_key_values=past_key_values if use_cache else None,
+            hidden_states=all_hidden_states,
+            attentions=all_self_attns,
+        )
+
+
+class PenguinVLVisionEncoderModel(PreTrainedModel):
+
+    config_class = PenguinVLVisionEncoderConfig
+    base_model_prefix = "penguinvl_vision_encoder"
+    main_input_name = "pixel_values"
+    supports_gradient_checkpointing = True
+    _no_split_modules = [
+        "PenguinVLVisionEncoderEmbeddings",
+    ]
+    _supports_flash_attn_2 = True
+    _supports_sdpa = True
+
+    def __init__(self, config: PenguinVLVisionEncoderConfig):
+        super().__init__(config=config)
+        self.embeddings = PenguinVLVisionEncoderEmbeddings(config)
+        self.encoder = PenguinVLVisionEncoderFromQwen3Model(config)
+
+        self.post_init()
+
+
+    def forward(self, pixel_values, grid_sizes, merge_sizes=None) -> torch.Tensor:
+        hidden_states = self.embeddings(pixel_values)
+        encoder_output = self.encoder(
+            inputs_embeds=hidden_states[None, ...],
+            grid_sizes=grid_sizes,
+            merge_sizes=merge_sizes,
+            output_hidden_states=True,
+        )
+        hidden_states = encoder_output.hidden_states
+        hidden_states = hidden_states[-1].squeeze(0)
+
+        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)

preprocessor_config.json

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