Abstract: Model quantization can reduce the memory footprint of the neural network and improve the computing efficiency. However, the sparse attention in Transformer models is difficult to quantize, ...