GPU Runners

Caution

GPU passthrough is not a first-class Runaway feature today. The scale set settings don’t expose a per-runner --gpus flag, so there’s no turnkey “enable GPU” toggle. This is an advanced, operator-side topic — read it as “here’s the realistic path and its limits,” not “here’s a supported flow.”

What’s missing and why it matters

Docker grants a container GPU access via the --gpus flag (or the equivalent device-request API). Runaway spawns runner containers itself and doesn’t surface a way to set that per runner, so you can’t declare “this scale set’s runners get a GPU” from the UI.

The realistic path

Getting GPU access onto runner containers is an agent-host and Docker-daemon exercise:

1. Install the NVIDIA Container Toolkit on the agent host. This is what lets Docker expose host GPUs to containers. Follow NVIDIA’s Container Toolkit install guide for your distro.
2. Configure the Docker daemon for GPU access. The toolkit registers a runtime; whether spawned containers get GPUs without an explicit per-container flag depends on the host’s daemon defaults (for example, setting the NVIDIA runtime as the default runtime in daemon.json). If the daemon hands GPUs to containers by default, runner containers on that host inherit them.

Note

Whether this actually works end-to-end depends entirely on your host and daemon configuration. Because Runaway doesn’t pass --gpus per runner, you’re relying on daemon-level defaults to grant access — verify it with a probe job (nvidia-smi in a step) before trusting it for real workloads.

Related

• Scale sets — what the scale set settings do expose.
• Multi-host — adding a dedicated agent host (e.g. a GPU box).