DuQuant: Distributing Outliers via Dual Transformation Makes Stronger Quantized LLMs

Welcome to the official code repository for "DuQuant: Distributing Outliers via Dual Transformation Makes Stronger Quantized LLMs (NeurIPS 2024, Oral)".

🔍 For more details, please refer to the project page: https://duquant.github.io/.

📰 News

• [2024/09/26] 🌟 Our DuQuant paper has been accepted for a Oral presentation at NeurIPS 2024 (only top 1% out of 15,671 submissions)! 🎉 Cheers!
• [2024/09/06] 🔥 We release the code!
• [2024/06/03] 🚀 Our paper is available on arXiv!

👀 Introduction

• We firstly identify Massive Outliers existence at the down_proj layer of FFN module in recent LLMs.
• DuQuant proposes to use Rotation transformation and Permutation transformation to effectively eliminate both massive and normal outliers.
• DuQuant establishs new state-of-the-art baselines for 4-bit weight-activation quantization across various model types and downstream tasks.

🔧 Installation

conda create -n duquant python=3.10 -y
conda activate duquant
git clone https://github.com/Hsu1023/DuQuant.git
pip install --upgrade pip 
pip install -r requirements.txt

⚙️ Usage

1. Preprocessing

python get_rot.py # need to be run only once for all models
python generate_act_scale_shift.py --model PATH_OF_MODEL # need to be run only once for each model (path can be hugging-face hub path or relative path)

2. Quantization

The bash script for DuQuant can be found in run.sh. You can choose the model to be quantized by providing model path after --model order. To evaluate DuQuant + lwc method, you can run run_lwc.sh script. In addition, you can add --save_dir to save the quantized models, and use --resume to reload the saved models.

Explanation of arguments:

• --model: the local model path or huggingface format.
• --wbits: weight quantization bits.
• --abits: activation quantization bits.
• --block_size: the block size of rotation matrices.
• --max_rotation_step: the max greedy search steps of rotation transformation.
• --permutation_times: the time of permutation transformation.
• --swc: the ratio of weight clipping (enable without LWC operation).
• --lac: the ratio of activation clipping.
• --lwc: activate the Learnable Weight Clipping (LWC).
• --epochs: the training epochs of LWC.
• --resume: loading pre-trained DuQuant parameters.
• --multigpu: to inference larger network on multiple GPUs.
• --save_dir: saving the quantization model for further exploration.
• --eval_ppl: evaluating the perplexity of quantized models.
• --tasks: evaluating on the zero-shot tasks.
• --eval_mmlu: evaluating on the MMLU benchmarks.
• --mmlu_data_dir: data path of the MMLU benchmarks.
• --eval_mtbench: evaluating on the MT-Bench.

3. Model Zoo

Currently, we support LLaMA series (LLaMA 1, 2 and 3), Vicuna series, and Mistral models.

Models	7B/8B	13B	30B	65B/70B
LLaMA1	✅	✅	✅	✅
LLaMA2	✅	✅	---	✅
LLaMA3	✅	---	---	✅
Vicuna-v1.5	✅	✅	---	---
Mistral	✅	---	---	---

📜 Result

• DuQuant achieves SoTA performance in PPL evaluation under W4A4 quantization.

• DuQuant showcases robustness towards LLaMA3-8B quantization.

📂 Contact

For immediate queries or further information, please open an issue or contact xuhb20@mails.tsinghua.edu.cn or haokun.lin@cripac.ia.ac.cn.

🙏 Acknowledgement

This repo is built upon the following projects:

• OmniQuant
• IntactKV
• EAGLE
• FastChat

We thank the authors for their code.

📝 Citation

We kindly request that you cite our work if you utilize the code or reference our findings in your research:

@article{lin2024duquant,
  title={DuQuant: Distributing Outliers via Dual Transformation Makes Stronger Quantized LLMs},
  author={Lin, Haokun and Xu, Haobo and Wu, Yichen and Cui, Jingzhi and Zhang, Yingtao and Mou, Linzhan and Song, Linqi and Sun, Zhenan and Wei, Ying},
  journal={arXiv preprint arXiv:2406.01721},
  year={2024}
}