Installing the NVIDIA device plugin

Why this matters

Before the GPU Operator automates everything, it is worth installing the bare device plugin once by hand, because it teaches you the prerequisites the operator otherwise hides: the NVIDIA driver must already be on the node, the NVIDIA container runtime must be configured as the default, and only then does the plugin DaemonSet have anything to advertise. Doing this manually once means that when the operator later breaks, you can reason about which of those three layers failed. This is the foundational install every other module assumes is working.

Demo

With the driver and container runtime already present on the node, installing the device plugin is a single DaemonSet apply. The DaemonSet lands on every node; on GPU nodes it discovers devices, on CPU-only nodes it finds none and idles.

Try it yourself

1. Apply the DaemonSet and confirm rollout status reaches 'desired number scheduled'.
2. Verify nvidia.com/gpu appears on GPU nodes within a minute of the rollout completing.
3. Confirm the toleration lets the plugin land on tainted GPU nodes (you'll add that taint in Module 3).
4. On a node missing the driver, read the plugin logs and identify the exact failure message before it can advertise.

Prompt your AI

Use these three in order. Each builds on the one before.

1. Basics & terminology

In one paragraph, what are the prerequisites for the NVIDIA device plugin to work, beyond just applying its DaemonSet?

2. Why it works (the mechanism)

Walk me through what has to be true on a node — driver, container runtime, plugin — for nvidia.com/gpu to appear, and the order those layers depend on each other.

3. Advanced — application & what's next

Given the manual device plugin install, what failure modes does it expose that the GPU Operator would otherwise hide, and how does that knowledge help debugging later?

References

• NVIDIA — k8s-device-plugin install
• Kubernetes — Schedule GPUs