Tag
Sol-Attn introduces a training-free method to sparsify attention for video generation inference, achieving over 2x speedup by dynamically selecting key-value blocks during online softmax with minimal quality loss.