docker-compose.arcane.gpu.yml

# honeypot-arcane — GPU-monitoring overlay (#2950).
#
# Arcane's GPU panel needs an NVIDIA device reserved for the container. That
# reservation is a HARD requirement: Docker refuses to start the container at
# all when no nvidia device driver is present —
#
#   Error response from daemon: could not select device driver "nvidia" with capabilities: [[gpu]]
#
# — and `count: 0` does not soften it (measured on a driverless host: both
# `count: all` and `count: 0` fail identically). Since Arcane is the control
# plane that materializes every other stack via its gitops syncs, having that
# requirement in the base file meant one optional monitoring panel could take
# out the entire install on a host whose GPU driver was not up yet. That is
# exactly what happened on the 2026-09-03 Rocky 10 rebuild.
#
# So the base docker-compose.arcane.yml is GPU-free and always startable, and
# install-homeserver.sh merges this overlay into the rendered compose.yml only
# after it has confirmed a working NVIDIA runtime on the host. Nothing here is
# required for Arcane to manage stacks.
services:
  arcane:
    environment:
      GPU_MONITORING_ENABLED: "true"
      GPU_TYPE: nvidia
      NVIDIA_VISIBLE_DEVICES: all
    deploy:
      resources:
        reservations:
          devices:
            - driver: nvidia
              count: all
              capabilities: [gpu]