Towards Effective and Efficient Non-autoregressive Decoding Using Block-based Attention Mask | Read Paper on Bytez