Q8BERT: Quantized 8bit BERT

Ofir Zafrir, Guy Boudoukh, Peter Izsak, Moshe Wasserblat. 2019 Fifth Workshop on Energy Efficient Machine Learning and Cognitive Computing - NeurIPS Edition (EMC2-NIPS) 2019 – 183 citations

[Paper]    
Model Architecture GPT Transformer Fine-Tuning Quantization Reinforcement Learning BERT Efficiency and Optimization Training Techniques

Recently, pre-trained Transformer based language models such as BERT and GPT, have shown great improvement in many Natural Language Processing (NLP) tasks. However, these models contain a large amount of parameters. The emergence of even larger and more accurate models such as GPT2 and Megatron, suggest a trend of large pre-trained Transformer models. However, using these large models in production environments is a complex task requiring a large amount of compute, memory and power resources. In this work we show how to perform quantization-aware training during the fine-tuning phase of BERT in order to compress BERT by \(4\times\) with minimal accuracy loss. Furthermore, the produced quantized model can accelerate inference speed if it is optimized for 8bit Integer supporting hardware.

Similar Work