Qwen-VL-Series-Finetune

Why we included this project

This repository brings Alibaba Cloud's Qwen vision-language models together in a single fine-tuning script, spanning Qwen2-VL, Qwen3-VL, and Qwen3.5. Training runs on HuggingFace and Liger-Kernel instead of a proprietary stack, and supports SFT, DPO, and GRPO for images and video. Hardware-limited teams get concrete help: selective layer freezing, a window-attention patch that cuts memory usage, and kernel-level speedups that reduce VRAM demands. If you want to adapt a Qwen model to your own data rather than just call it through an API, this gives you one coherent place to start instead of a scattered set of notebooks.