Larger circuits use more GPU-MEM than smaller circuits even though we only train on k-hop neighborhood.