Created by: kuke

When saving inference model of BERT after fine-tuning, there is a bug that the model is not saved completely. It is a known bug that will be fixed soon in the forthcoming Paddle Fluid 1.3.1