We introduce Integer Scale, a novel post-training quantization scheme for large language models that effectively resolves the inference bottleneck in current fine-grained quantization approaches while maintaining similar accuracies. Integer Scale is a free lunch as it requires no extra calibration or fine-tuning which will otherwise incur additional costs. It can be used plug-and-play for most fine-grained quantization methods. Its integration results in at most 1.85x end-to-end speed boost over the original counterpart with comparable accuracy. Additionally, due to the orchestration of the proposed Integer Scale and fine-grained quantization, we resolved the quantization difficulty for Mixtral-8x7B and LLaMA-3 models with negligible performance degradation, and it comes with an end-to-end speed boost of 2.13x, and 2.31x compared with their FP16 versions respectively.
Cite
@article{arxiv.2405.14597,
title = {Integer Scale: A Free Lunch for Faster Fine-grained Quantization of LLMs},
author = {Qingyuan Li and Ran Meng and Yiduo Li and Bo Zhang and Yifan Lu and Yerui Sun and Lin Ma and Yuchen Xie},
journal= {arXiv preprint arXiv:2405.14597},
year = {2024}
}