My suggestion is to add a user-definable prefilter for the "–models_to_relax=Best" command, based on prefiltering using the iPTM scores (or other scores).
i.e. the models_to_relax=best protocol is only activated on protein pairs where the best model obtained a iPTM score of or above a user-definable number e.g. 0.6.
This way GPU processing time is not wasted on relaxing the best model from each protein pair, but is instead focused on relaxing the best model of prefiltered protein pairs only.