File size: 4,264 Bytes
247943c |
1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25 26 27 28 29 30 31 32 33 34 35 36 37 38 39 40 41 42 43 44 45 46 47 48 49 50 51 52 53 54 55 56 57 58 59 60 61 62 63 64 65 66 67 68 69 70 71 72 73 74 75 76 77 78 79 80 81 82 83 84 85 86 87 88 89 |
---
license: apache-2.0
datasets:
- lmms-lab/LLaVA-OneVision-Data
- BAAI/Infinity-MM
language:
- en
- zh
base_model:
- google/siglip2-so400m-patch16-512
- Qwen/Qwen2-1.5B-Instruct
pipeline_tag: image-text-to-text
library_name: transformers
---
# FlashVL-2B-Static-GRPO
[\[📜 FlashVL\]](https://www.arxiv.org/abs/2505.09498)
## Introduction
We are excited to introduce **FlashVL**, a novel approach to optimizing Vision-Language Models (VLMs) for real-time applications, targeting ultra-low latency and high throughput without sacrificing accuracy. Leveraging advanced architectural enhancements and efficient computational strategies, Flash-VL 2B is designed to maximize throughput by reducing processing time while maintaining competitive performance across multiple vision-language benchmarks. Our approach includes tailored architectural choices, token compression mechanisms, data curation, training schemes, and a novel image processing technique called implicit semantic stitching that effectively balances computational load and model performance. Through extensive evaluations on 11 standard VLM benchmarks, we demonstrate that Flash-VL 2B achieves state-of-the-art results in both speed and accuracy, making it a promising solution for deployment in resource-constrained environments and large-scale real-time applications.
### Environment Setup
```bash
pip install torch==2.1.2
pip install transformers==4.50.0.dev0
```
### How to use it?
```python
import torch
from PIL import Image
import requests
from io import BytesIO
from transformers import AutoModel, AutoTokenizer, SiglipProcessor
model_path = "Flash-VL/FlashVL-2B-Static-GRPO"
model = AutoModel.from_pretrained(model_path, torch_dtype=torch.bfloat16,trust_remote_code=True,device_map='cuda')
model.tokenizer = AutoTokenizer.from_pretrained(model_path,device_map='cuda')
model.im_trans = SiglipProcessor.from_pretrained(model_path).image_processor
image_url ="https://s3plus.meituan.net/automl-datasets/mlm/3FF4.png"
response = requests.get(image_url)
image_data = BytesIO(response.content)
pil_image = Image.open(image_data).convert('RGB')
messages = [{'role': 'user', 'content': "说说图中第一行第二列是什么蔬菜，买一斤多少钱"}]
answer = model.chat(pil_image, messages, do_sample=False, max_new_tokens=256)
print(answer)
# 图片中第一行第二列的蔬菜是**荷兰豆**，买一斤的价格是**￥16.8**。
```
### Evaluation
| Method/model | Average | DynaMath | MathVision | MathVerse | MMMU Pro | WeMath |
| :--------------------: | :----------------: | :----------------: | :----------------: | :----------------: | :----------------: | :----------------: |
| Flash-VL-2B<sub>s</sub> | 23.80 | 23.19 | 26.72 | 16.84 | 16.24 | 36.03 |
| InternVL3-2B | 27.03 | 32.55 | 26.49 | 17 | 22.56 | 36.55 |
| + SFT | 26.08 (+2.28) | 28.28 | 31.06 | 16.97 | 15.95 | 38.16 |
| + RL | 27.23 (+3.43) | 26.94 | 27.94 | 17.73 | 16.99 | 46.55 |
| FlashVL-2B-Static-GRPO| 29.05 (+5.25) | 30.61 | 32.48 | 18.45 | 16.53 | 47.18 |
Note: FlashVL-2B-Static-GRPO applies both SFT and RL.
We use [VLMEvalKit](https://github.com/open-compass/VLMEvalKit) to evaluate FlashVL-2B-Static.
## Citation
If you find this project useful in your research, please consider citing:
```BibTeX
@misc{zhang2025flashvl2boptimizingvisionlanguage,
title={Flash-VL 2B: Optimizing Vision-Language Model Performance for Ultra-Low Latency and High Throughput},
author={Bo Zhang and Shuo Li and Runhe Tian and Yang Yang and Jixin Tang and Jinhao Zhou and Lin Ma},
year={2025},
eprint={2505.09498},
archivePrefix={arXiv},
primaryClass={cs.CV},
url={https://arxiv.org/abs/2505.09498},
}
``` |