Since 897e52f ansible runs cuda_init.sh when the cuda_gpu_name0 device does not eixst. Can we remove the entry for cuda_init.sh in rc.local? ( I think we can )
As the ansible in rc.local takes a long time and slurm starts without waiting for rc.local then we'd still need to run cuda_init before ansible. Added as a systemd.service in f93fd8a (not tagged as a release or added to fgci-ansible's requirements yet)