view article Article Trainable Dynamic Mask Sparse Attention: Bridging Efficiency and Effectiveness in Long-Context Language Models Aug 5 โข 7