It was used in https://ai.googleblog.com/2020/01/reformer-efficient-transfo... for faster and more memory-efficient attention.