File size: 7,192 Bytes

---
license: cc-by-4.0
datasets:
- jihyoung/AQuA
language:
- en
base_model:
- OpenGVLab/InternVL3-2B-Instruct
---

# AQuA-Qwen

<p align="center">
  <a href="https://arxiv.org/abs/2603.07394"><img src="https://img.shields.io/badge/arXiv-2603.07394-b31b1b?logo=arxiv&logoColor=white" alt="arXiv"/></a>
  &nbsp;
  <a href="https://huggingface.co/datasets/jihyoung/AQuA"><img src="https://img.shields.io/badge/🤗-Dataset-FFD21E?logoColor=black" alt="Dataset"/></a>
  &nbsp;
  <a href="https://huggingface.co/jihyoung/AQuA-Qwen"><img src="https://img.shields.io/badge/🤗-AQuA--Qwen-FFD21E?logoColor=black" alt="AQuA-Qwen"/></a>
  &nbsp;
  <a href="https://huggingface.co/jihyoung/AQuA-InternVL"><img src="https://img.shields.io/badge/🤗-AQuA--InternVL-FFD21E?logoColor=black" alt="AQuA-InternVL"/></a>
</p>

This model is a fine-tuned version of [InternVL3-2B-Instruct](https://huggingface.co/OpenGVLab/InternVL3-2B-Instruct) on the [AQuA dataset](https://huggingface.co/datasets/jihyoung/AQuA), trained to generate strategic responses to ambiguous visual questions.
 
## Usage
 
```python
import math
import numpy as np
import torch
import torchvision.transforms as T
from decord import VideoReader, cpu
import requests
from io import BytesIO
from PIL import Image
from torchvision.transforms.functional import InterpolationMode
from transformers import AutoModel, AutoTokenizer, AutoConfig

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


def build_transform(input_size):
    MEAN, STD = IMAGENET_MEAN, IMAGENET_STD
    transform = T.Compose([
        T.Lambda(lambda img: img.convert('RGB') if img.mode != 'RGB' else img),
        T.Resize((input_size, input_size), interpolation=InterpolationMode.BICUBIC),
        T.ToTensor(),
        T.Normalize(mean=MEAN, std=STD)
    ])
    return transform


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


def dynamic_preprocess(image, min_num=1, max_num=12, image_size=448, use_thumbnail=False):
    orig_width, orig_height = image.size
    aspect_ratio = orig_width / orig_height

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

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

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

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


def load_image(image_file, input_size=448, max_num=12):
    if isinstance(image_file, str) and image_file.startswith(('http://', 'https://')):
        response = requests.get(image_file, timeout=10)
        response.raise_for_status()
        image = Image.open(BytesIO(response.content)).convert('RGB')
    else:
        image = Image.open(image_file).convert('RGB')

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


def split_model(model_name):
    device_map = {}
    world_size = torch.cuda.device_count()
    config = AutoConfig.from_pretrained(model_name, trust_remote_code=True)
    num_layers = config.llm_config.num_hidden_layers
    # Since the first GPU will be used for ViT, treat it as half a GPU.
    num_layers_per_gpu = math.ceil(num_layers / (world_size - 0.5))
    num_layers_per_gpu = [num_layers_per_gpu] * world_size
    num_layers_per_gpu[0] = math.ceil(num_layers_per_gpu[0] * 0.5)
    layer_cnt = 0
    for i, num_layer in enumerate(num_layers_per_gpu):
        for j in range(num_layer):
            device_map[f'language_model.model.layers.{layer_cnt}'] = i
            layer_cnt += 1
    device_map['vision_model'] = 0
    device_map['mlp1'] = 0
    device_map['language_model.model.tok_embeddings'] = 0
    device_map['language_model.model.embed_tokens'] = 0
    device_map['language_model.output'] = 0
    device_map['language_model.model.norm'] = 0
    device_map['language_model.model.rotary_emb'] = 0
    device_map['language_model.lm_head'] = 0
    device_map[f'language_model.model.layers.{num_layers - 1}'] = 0

    return device_map


# If you set `load_in_8bit=True`, you will need two 80GB GPUs.
# If you set `load_in_8bit=False`, you will need at least three 80GB GPUs.
path = 'jihyoung/AQuA-InternVL'
device_map = split_model(path)
model = AutoModel.from_pretrained(
    path,
    torch_dtype=torch.bfloat16,
    load_in_8bit=False,
    low_cpu_mem_usage=True,
    use_flash_attn=True,
    trust_remote_code=True,
    device_map=device_map).eval()
tokenizer = AutoTokenizer.from_pretrained(path, trust_remote_code=True, use_fast=False)

# set the max number of tiles in `max_num`
pixel_values = load_image('http://images.cocodataset.org/train2017/000000539311.jpg', max_num=12).to(torch.bfloat16).cuda()
generation_config = dict(max_new_tokens=1024, do_sample=True)

question = '<image>\nWhat color is this bat?'
response = model.chat(tokenizer, pixel_values, question, generation_config)
print(f'User: {question}\nAssistant: {response}')
```
 
## Citation
 
If you use this model in your research, please cite:
 
```bibtex
@inproceedings{
  jang2026aqua,
  title={{AQ}uA: Toward Strategic Response Generation for Ambiguous Visual Questions},
  author={Jihyoung Jang and Hyounghun Kim},
  booktitle={The Fourteenth International Conference on Learning Representations},
  year={2026},
  url={https://openreview.net/forum?id=7b1MpD6IF8}
}
```