In case Llama 3 dynamically adjust the number of active parameters when running, then we resolve to the most number that still makes sense (i.e. with the best performance)