New bug in fast sample posterior predictive

Perhaps this should be reported as a bug on github instead?