Possibly nomenclature is tripping you up here as it has done so repeatedly for me (see here for my most recent self-clarifying post).
Using the terms of the SUG 1.13, think first at the individual level, where there are K predictors; then each of the L group-level predictors serve to shove around each of these K predictors, yielding a K\times L matrix of group-level coefficients, encoded in the model code by the gamma variable.