Efficient Inference For Neural Machine Translation