Zero-Shot Dynamic Quantization for Transformer Inference

TL;DR

This paper presents a run-time zero-shot quantization method for BERT-like models that reduces accuracy loss during 8-bit integer quantization without additional training or calibration, enabling efficient NLP inference.

Contribution

The proposed method allows quantization of transformer models at run-time without training modifications or calibration, simplifying deployment.

Findings

Effective on multiple NLP tasks

Reduces accuracy loss compared to existing methods

No additional calibration needed

Abstract

We introduce a novel run-time method for significantly reducing the accuracy loss associated with quantizing BERT-like models to 8-bit integers. Existing methods for quantizing models either modify the training procedure,or they require an additional calibration step to adjust parameters that also requires a selected held-out dataset. Our method permits taking advantage of quantization without the need for these adjustments. We present results on several NLP tasks demonstrating the usefulness of this technique.