#4343rd of 53 in Training and fine-tuning
qwen-vl-series-finetune
Fine-tuning scripts for Qwen2-VL, Qwen2.5-VL, Qwen3-VL and Qwen3.5
- Stars
- 2k
- License
- Apache-2.0
- Last commit
- Sep 2026
- Language
- Python
1,807 on 21 Apr 20261,983 on 10 Nov 2026
From GitHub's daily star counts.
Checked on GitHub.
Overview
Training scripts built on HuggingFace and Liger-Kernel for fine-tuning Qwen vision-language models. They cover full fine-tuning, LoRA/QLoRA/DoRA, DPO, GRPO and classification, with DeepSpeed, multi-image and video inputs, and a Gradio inference UI. Data uses the LLaVA JSON format, with an optional reasoning field for Qwen3-VL-Thinking and Qwen3.5.
Who it is for: Engineers fine-tuning Qwen vision-language models on their own GPUs
Strengths
- Covers SFT, DPO, GRPO and classification in one repo
- Supports Qwen2-VL through Qwen3.5, including Qwen3-VL-MoE
- Separate learning rates for vision tower, merger and LLM; top-k layer unfreezing
- Image, multi-image, video and mixed-modality data with DeepSpeed zero2/zero3
Weaknesses
- Qwen3.5 needs Flash Attention 2 disabled; it raised CUDA errors in testing
- QLoRA cannot be combined with vision training or Liger kernel
- Full fine-tuning of dense Qwen3-VL with Liger can be slow
- Limited to Qwen-VL family; other models live in separate repos
What it needs
- GPU required
- Needs PyTorch, HuggingFace Transformers 5.3.0, Liger-Kernel 0.8.0, DeepSpeed, flash-attn, qwen-vl-utils, CUDA 12.8
- Models: Qwen2-VL, Qwen2.5-VL, Qwen3-VL, Qwen3-VL-MoE, Qwen3.5
Quick answers
Is qwen-vl-series-finetune open source?
Yes. Its code is on GitHub under the Apache-2.0 license.
Can you run qwen-vl-series-finetune with Docker?
Yes. Its repository has a Docker setup: a Dockerfile, a Compose file or a published image.
Does qwen-vl-series-finetune need a GPU?
Yes, it needs a GPU.
Also in Training and fine-tuning
See all 53| Rank | Project | Score |
|---|---|---|
| 1 | unslothDesktop and web app to run and fine-tune local models | 88 out of 100 |
| 2 | peftPython library for LoRA and other parameter-efficient fine-tuning of Transformers models | 88 out of 100 |
| 3 | deepspeedPyTorch library for distributed training of very large models with ZeRO | 86 out of 100 |
| #4 | verlReinforcement learning training library for LLM post-training with PPO and GRPO | 86 out of 100 |
| #5 | trlPost-training library with SFT, DPO, GRPO and KTO trainers | 86 out of 100 |