@tri_dao
This was a wild bug hunt, weeks of effort from @MayankMish98 to track down. The wrong init of Mamba2 in many reimplementations causes the layer to decay its states too quickly, focusing in short context instead. Pretraining is mostly about getting these little things right