Qwen2.5-VL-3B-Instruct-GPTQ-Int4

This is an UNOFFICIAL GPTQ-Int4 quantized version of the Qwen2.5-VL model using gptqmodel library.

The model is compatible with the latest transformers library (which can run non-quantized Qwen2.5-VL models).

Performance

Model	Size (Disk)	ChartQA (test)	OCRBench
Qwen2.5-VL-3B-Instruct	7.1 GB	83.48	791
Qwen2.5-VL-3B-Instruct-AWQ	3.2 GB	82.52	786
Qwen2.5-VL-3B-Instruct-GPTQ-Int4	3.2 GB	82.56	784
Qwen2.5-VL-3B-Instruct-GPTQ-Int3	2.9 GB	76.68	742
Qwen2.5-VL-7B-Instruct	16.0 GB	83.2	846
Qwen2.5-VL-7B-Instruct-AWQ	6.5 GB	79.68	837
Qwen2.5-VL-7B-Instruct-GPTQ-Int4	6.5 GB	81.48	845
Qwen2.5-VL-7B-Instruct-GPTQ-Int3	5.8 GB	78.56	823

Note

• Evaluations are performed using lmms-eval with default setting.
• GPTQ models are computationally more effective (fewer VRAM usage, faster inference speed) than AWQ series in these evaluations.
• We recommend use gptqmodel instead of autogptq library, as autogptq is no longer maintained.

Quick Tour

Install the required libraries:

pip install git+https://github.com/huggingface/transformers accelerate qwen-vl-utils
pip install git+https://github.com/huggingface/optimum.git
pip install gptqmodel 

Optionally, you may need to install:

pip install tokenicer device_smi logbar

Sample code:

from transformers import Qwen2_5_VLForConditionalGeneration, AutoTokenizer, AutoProcessor
from qwen_vl_utils import process_vision_info

model = Qwen2_5_VLForConditionalGeneration.from_pretrained(
    "hfl/Qwen2.5-VL-3B-Instruct-GPTQ-Int4", 
    attn_implementation="flash_attention_2",
    device_map="auto"
)
processor = AutoProcessor.from_pretrained("hfl/Qwen2.5-VL-3B-Instruct-GPTQ-Int4")

messages = [{
    "role": "user",
    "content": [
        {"type": "image", "image": "https://raw.githubusercontent.com/ymcui/Chinese-LLaMA-Alpaca-3/refs/heads/main/pics/banner.png"},
        {"type": "text", "text": "请你描述一下这张图片。"},
    ],
}]

text = processor.apply_chat_template(messages, tokenize=False, add_generation_prompt=True)
image_inputs, video_inputs = process_vision_info(messages)
inputs = processor(
    text=[text], images=image_inputs, videos=video_inputs,
    padding=True, return_tensors="pt",
).to("cuda")

generated_ids = model.generate(**inputs, max_new_tokens=512)
generated_ids_trimmed = [out_ids[len(in_ids) :] for in_ids, out_ids in zip(inputs.input_ids, generated_ids)]
output_text = processor.batch_decode(generated_ids_trimmed, skip_special_tokens=True, clean_up_tokenization_spaces=False)
print(output_text[0])

Response:

这张图片展示了一个中文和英文的标志，内容为“中文LLaMA & Alpaca大模型”和“Chinese LLaMA & Alpaca Large Language Models”。标志左侧有两个卡通形象，一个是红色围巾的羊驼，另一个是白色毛发的羊驼，背景是一个绿色的草地和一座红色屋顶的建筑。标志右侧有一个数字3，旁边有一些电路图案。整体设计简洁明了，使用了明亮的颜色和可爱的卡通形象来吸引注意力。

Disclaimer

• This is NOT an official model by Qwen. Use at your own risk.
• For detailed usage, please check Qwen2.5-VL's page.