Abstract
Transformer is being widely used in Neural Machine Translation (NMT). Deploying Transformer models to mobile or edge devices with limited resources is challenging because of heavy computation and memory overhead during inference. 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 accuracy 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 with lower 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 11.8x smaller model size than the baseline model, with less than -0.5 BLEU. We achieve 8.3x reduction in run-time memory footprints and 3.5x speed up (Galaxy N10+) such that our proposed compression strategy enables efficient implementation for on-device NMT.
Abstract (translated)
URL
https://arxiv.org/abs/2009.07453
