Extremely Low Bit Transformer Quantization for On-Device Neural Machine Translation

TL;DR

This paper introduces a mixed precision quantization method for Transformer models, significantly reducing model size and memory usage, enabling efficient on-device neural machine translation with minimal BLEU score loss.

Contribution

It proposes a novel mixed precision quantization strategy tailored for Transformer weights, achieving extremely low bit representation and substantial efficiency gains.

Findings

11.8× smaller model size with less than -0.5 BLEU

8.3× reduction in run-time memory footprint

3.5× speedup on Galaxy N10+

Abstract

The deployment of widely used Transformer architecture is challenging because of heavy computation load and memory overhead during inference, especially when the target device is limited in computational resources such as mobile or edge devices. Quantization is an effective technique to address such challenges. Our analysis shows that for a given number of quantization bits, each block of Transformer contributes to translation quality and inference computations in different manners. Moreover, even inside an embedding block, each word presents vastly different contributions. Correspondingly, we propose a mixed precision quantization strategy to represent Transformer weights by an extremely low number of bits (e.g., under 3 bits). For example, for each word in an embedding block, we assign different quantization bits based on statistical property. Our quantized Transformer model achieves 11.8$\times$ smaller model size than the baseline model, with less than -0.5 BLEU. We achieve 8.3$\times$ reduction in run-time memory footprints and 3.5$\times$ speed up (Galaxy N10+) such that our proposed compression strategy enables efficient implementation for on-device NMT.