Agile Quant

Official repo for paper: Agile-Quant: Activation-Guided Quantization for Faster Inference of LLMs on the Edge

This paper is accepted by AAAI 2024

Usage

1. Replace llama model files in transformers package with transformers/models/llama
2. Download models sh download.sh
3. Use GPTQ to quantize weights sh run-gptq-llama.sh
4. Quantize activation with gptq_fq_quant_llama.py

Citation

@inproceedings{
    shen2024agile,
    title     = {Agile-Quant: Activation-Guided Quantization for Faster Inference of LLMs on the Edge},
    author    = {Shen, Xuan and Dong, Peiyan and Lu, Lei and Kong, Zhenglun and Li, Zhengang and Lin, Ming and Wu, Chao and Wang, Yanzhi},
    booktitle = {AAAI},
    year      = {2024},
}

Acknowledgment

The code is mainly based on the quantization works GPTQ and FQ-ViT.