The current job limit will be multiplied by num_seeds, which is quite unintuitive. Max_parallel needs to be divided by num_seeds on init to fix this.