mirror of
https://github.com/kevinveenbirkenbach/computer-playbook.git
synced 2026-11-09 12:17:13 +00:00
A device reservation is not conditional in compose, and a count of 0 does not disable it: the daemon still selects the nvidia driver and fails where there is none. Every GPU-less host therefore lost compose up with 'could not select device driver', which is what broke test-dns on all five CI distros. Move the reservation out of compose.yml and i18n/compose.override.yml into compose/gpu.override.yml, layered by compose_file_args() only when Profile.gpu_reservable() sees a non-zero INFINITO_GPU_COUNT. A GPU host renders exactly what it rendered before, so the passthrough sys-svc-container needs to find /dev/nvidia0 and emit runtime: nvidia for a gpu:true service stays intact. Co-Authored-By: Claude Opus 5 (1M context) <noreply@anthropic.com>
Compose Configuration 🧩
This directory holds the static configuration assets consumed by the top-level compose.yml of the local development and CI testing stack.
Scope 📋
- The directory MUST contain only files that are bind-mounted into a
service defined in
compose.yml. - The directory MUST NOT host application configuration that belongs to
an Ansible role. Per-role compose payloads live under
roles/<role>/templates/. - The directory MUST NOT host workflow scripts, test fixtures, or
build helpers. Those belong under
scripts/ortests/.
Layout 🗂️
Each compose service that needs static configuration owns a
sub-directory whose name matches the service. New service-specific
config MUST go into the matching sub-directory rather than at the
compose/ root.
For the rules that govern compose.yml itself (env-variable contract,
profile gating, resource caps), see
compose.yml.md.