It has to be 2^n nodes and limited to one per attention head that the model has.