It is being used in the reformer architecture to relate the dot product attention layer. This enables much longer input sequences and reduces computational complexity from O(n^2) to O(nlog(n)), where n in the length of the sequence.

https://huggingface.co/transformers/model_doc/reformer.html