sparkup
sparkup is an Ansible playbook with sane defaults and observability.
Defaults: Docker with the NVIDIA runtime, users and GitHub SSH keys, signed kernel, firmware staged, ufw and optional, the spbm firmware module for the whole-system power.
Observability: node and GPU exporters, Prometheus with a year of retention, Grafana on http://spark.local with a provisioned dashboard.
nvidia-smi sees the GPU rail, about half the wall draw, but I want the whole-system power and energy into Prometheus. That’s why I’ve added the spbmfirmware module (OFF by default).
Driving it with an agent by pointing it at INSTALL_CLAUDE.md