You mean log-deviations vs. absolute deviations? The answer is: put all variables to be log-linearized (i.e. in percentage deviations from steady state) in exp() to perform a variable substitution, see [Model in exp-logs). Variables in absoluted deviations from steady state are not put into exp()