3 ms·Native Sparse Attention: Hardware-Aligned, Natively Trainable Sparse Attention15 points by mfiguiere 2y agorkwasny 2y agoI am not sure why there are no comments, no one in AI here? They show 10x improvement in inference speed!adlumal 2y agoThis paper is looking to be pretty significant. Was hoping to find some discussion here as well!