GPTQT: Quantize Large Language Models Twice to Push the Efficiency

TL;DR

GPTQT is a novel two-step quantization method that significantly reduces memory and increases speed of large language models by converting weights into 3-bit and 2-bit binary codes, outperforming existing methods.

Contribution

The paper introduces GPTQT, a new progressive quantization approach that effectively minimizes quantization error and enhances efficiency for large language models.

Findings

Reduces perplexity by 4.01 on opt-66B

Increases inference speed by 1.24 times on opt-30b

Outperforms existing binary coding quantization methods on Llama2

Abstract

Due to their large size, generative Large Language Models (LLMs) require significant computing and storage resources. This paper introduces a new post-training quantization method, GPTQT, to reduce memory usage and enhance processing speed by expressing the weight of LLM in 3bit/2bit. Practice has shown that minimizing the quantization error of weights is ineffective, leading to overfitting. Therefore, GPTQT employs a progressive two-step approach: initially quantizing weights using Linear quantization to a relatively high bit, followed by converting obtained int weight to lower bit binary coding. A re-explore strategy is proposed to optimize initial scaling factor. During inference, these steps are merged into pure binary coding, enabling efficient computation. Testing across various models and datasets confirms GPTQT's effectiveness. Compared to the strong 3-bit quantization baseline, GPTQT further reduces perplexity by 4.01 on opt-66B and increases speed by 1.24 times on opt-30b. The results on Llama2 show that GPTQT is currently the best binary coding quantization method for such kind of LLMs.