It adds a very local sliding window attention, the context is only 3 adjacent frames per step. They need the access to adjacent frames to show the implicit mode

It adds a very local sliding window attention, the context is only 3 adjacent frames per step. They need the access to adjacent frames to show the implicit model gradient computation but I didn't yet follow the derivation for why this is so.