For readers who write attention kernels for long sequences.