GPU Setup Script

Automated, reproducible installation of either NVIDIA drivers or the NVIDIA GPU Operator on Kubernetes worker nodes.

---

Supported Linux distributions

• Ubuntu 22.04 LTS
• SUSE Linux Enterprise Server (SLES) 15 SP6
• SUSE Linux Enterprise Micro 6.0 or newer

---

Key capabilities

• One‑command GPU enablement
• Mutually exclusive driver / operator paths – avoid partial or conflicting configurations.
• Dry‑run mode – print every command without touching the host.
• CUDA sample workload validation – optional VectorAdd pod to prove everything works.

---

Quick start

Install the latest driver and keep the node running (no reboot):

curl -fsSL https://raw.githubusercontent.com/victorstevansuse/nvidia-drivers/latest/install.sh | sudo bash

Download + tweak flags locally:

curl -fsSL -o gpu-setup.sh https://raw.githubusercontent.com/victorstevansuse/nvidia-drivers/latest/install.sh
chmod +x gpu-setup.sh
sudo ./gpu-setup.sh [flags]

---

CLI flags

Short	Long option	Default	Purpose
-d	--skip-driver	false	Skip NVIDIA driver installation.
-o	--install-gpu-operator	false	Deploy the NVIDIA GPU Operator via Helm. Cannot be used together with driver installation – make sure to pass --skip-driver.
-r	--enable-reboot	false	Reboot immediately after successfully installing the driver.
-w	--deploy-sample-workload	false	Run the CUDA vectoradd sample pod after setup to validate GPU functionality.
	--dry-run	false	Echo every shell command instead of executing it.
-h	--help	–	Show usage information.

---

Examples

Scenario	Command
Driver only (default path)	sudo ./gpu-setup.sh
GPU Operator only (driver already present)	./gpu-setup.sh --skip-driver --install-gpu-operator
Dry‑run – explore without changes	sudo ./gpu-setup.sh --dry-run

---

Workflow overview

1. Pre‑flight checks • Ensures required utilities (bash, kubectl, helm, etc.). • Verifies root privileges when the driver will be installed. • Detects OS and GPU presence. • Confirms cluster health when the GPU Operator path is chosen.

2. Driver installation path (default) • Adds the correct NVIDIA repository and installs the newest production branch: ‑ nvidia-driver-535 on Ubuntu. ‑ nvidia-open-driver-G06 on SLES / SLE Micro. • Optionally reboots the host if -r is set.

3. GPU Operator installation path (--install-gpu-operator) • Creates the gpu-operator namespace with the required security label. • Adds the NVIDIA Helm repo and deploys the operator. • Honours existing Node Feature Discovery (NFD) deployment if found.

4. Optional sample workload (-w) • Launches a CUDA VectorAdd pod. • Waits for completion and prints the container logs. • Cleans up the pod when it succeeds.

---

Prerequisites

Requirement	Needed for	Notes
Root privileges	Driver installation	Operator installation can run as an unprivileged user, provided that kubectl and helm are on the user’s login PATH.
kubectl & helm	GPU Operator	helm 3.0+ required. Control‑plane must be Ready.
Physical NVIDIA GPU	Driver installation & sample workload	The script aborts early if no GPU (nvidia PCI device) is detected.

---

Troubleshooting

• “NVIDIA driver already present” – rerun with --skip-driver if you intentionally keep the existing driver.
• “GPUs not found” – validate hardware passthrough (for VMs) or physical GPU presence.
• Operator deploy hangs – check cluster state (kubectl get pods -A) and verify Node Feature Discovery labels.
• "Control-plane not healthy" when cluster is available - make sure to run the script without sudo when installing GPU Operator or deploying the sample workload.