nvidia-fabricmanager.yaml

# https://docs.nvidia.com/datacenter/tesla/pdf/fabric-manager-user-guide.pdf
name: nvidia-fabricmanager
container:
  entrypoint: /usr/local/bin/nv-fabricmanager
  args:
   - --config
   - /usr/local/share/nvidia/nvswitch/fabricmanager.cfg
  mounts:
    # device files
    - source: /dev
      destination: /dev
      type: bind
      options:
        - rshared
        - rbind
        - rw
    # shared libraries
    - source: /lib64
      destination: /lib64
      type: bind
      options:
        - bind
        - ro
    # shared libraries
    - source: /usr/local/glibc
      destination: /usr/local/glibc
      type: bind
      options:
        - bind
        - ro
    # nvidia libraries
    - source: /usr/local/lib
      destination: /usr/local/lib
      type: bind
      options:
        - bind
        - ro
    # service state file
    - source: /var/run/nvidia-fabricmanager
      destination: /var/run/nvidia-fabricmanager
      type: bind
      options:
        - rshared
        - rbind
        - rw
    # log files
    - source: /var/log
      destination: /var/log
      type: bind
      options:
        - rshared
        - rbind
        - rw
    # fabric topology files
    - source: /usr/local/share/nvidia/nvswitch
      destination: /usr/local/share/nvidia/nvswitch
      type: bind
      options:
        - rshared
        - rbind
        - ro
    # binaries
    - source: /usr/local/bin
      destination: /usr/local/bin
      type: bind
      options:
        - bind
        - ro
depends:
  - service: cri
  # we need to depend on udevd so that the nvidia device files are created
  - service: udevd
  - path: /sys/bus/pci/drivers/nvidia
restart: always