Fully Quantized Transformer for Machine Translation

TL;DR

This paper introduces FullyQT, a fully quantized Transformer model for machine translation that maintains or improves translation quality while significantly reducing computational costs, achieving state-of-the-art quantization results.

Contribution

It presents the first fully quantized Transformer that avoids performance loss, demonstrating that 8-bit models can match or surpass full-precision models in translation quality.

Findings

8-bit models achieve equal or higher BLEU scores than full-precision models

FullyQT outperforms previous quantization methods in machine translation tasks

The approach reduces computational costs without sacrificing translation quality

Abstract

State-of-the-art neural machine translation methods employ massive amounts of parameters. Drastically reducing computational costs of such methods without affecting performance has been up to this point unsuccessful. To this end, we propose FullyQT: an all-inclusive quantization strategy for the Transformer. To the best of our knowledge, we are the first to show that it is possible to avoid any loss in translation quality with a fully quantized Transformer. Indeed, compared to full-precision, our 8-bit models score greater or equal BLEU on most tasks. Comparing ourselves to all previously proposed methods, we achieve state-of-the-art quantization results.