3 ms·Twilight: Adaptive Attention Sparsity with Hierarchical Top-$P$ Pruning2 points by PaulHoule 2y ago