Native Sparse Attention: Hardware-Aligned, Natively Trainable Sparse Attention

This paper is looking to be pretty significant. Was hoping to find some discussion here as well!