Self-Distilled Quantization: Achieving High Compression Rates in Transformer-Based Language Models

TL;DR

This paper introduces self-distilled quantization (SDQ), a novel method for reducing transformer language models from 32-bit to 8-bit weights, maintaining performance and addressing multilingual generalization challenges.

Contribution

The paper proposes SDQ, a new quantization technique that minimizes errors and improves compression of multilingual transformer models without significant performance loss.

Findings

SDQ outperforms baseline quantization methods.

Models compressed from 32-bit to 8-bit retain high performance.

Multilingual models face unique quantization challenges.

Abstract

We investigate the effects of post-training quantization and quantization-aware training on the generalization of Transformer language models. We present a new method called self-distilled quantization (SDQ) that minimizes accumulative quantization errors and outperforms baselines. We apply SDQ to multilingual models XLM-R-Base and InfoXLM-Base and demonstrate that both models can be reduced from 32-bit floating point weights to 8-bit integer weights while maintaining a high level of performance on the XGLUE benchmark. Our results also highlight the challenges of quantizing multilingual models, which must generalize to languages they were not fine-tuned on.