What is going on is that before the fitting stuff starts, the df1b2 verzion of the user function gets called once to determine whcih group of random effects get used in each separable function call so that the correct data structures can get set up. The actual value of the random effects is irrelevant at this point.