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

I am not sure why there are no comments, no one in AI here?<p>They show 10x improvement in inference speed!