Efficient Sparse Attention needs Adaptive Token Release

Open in new window