Fetching the paper…

SpAtten: Efficient Sparse Attention Architecture with Cascade Token and Head Pruning · Around