NVIDIA Driver + CUDA Toolkit Installer

Interactive, APT-first installer for NVIDIA drivers and the optional CUDA Toolkit on Debian 12 (Bookworm) and Debian 13 (Trixie).

The recommended paths use Debian's own packages. The script does not add NVIDIA repositories, force an Xorg configuration, kill package-manager processes, or delete APT lock files.

Supported systems

• Debian 12 or 13
• amd64 or arm64
• An NVIDIA PCI display device
• Root access
• APT sources with main contrib non-free non-free-firmware

The advanced NVIDIA .run installer is available only on amd64 and is intentionally not recommended for normal Debian installations.

Features

• Installs nvidia-driver from Debian stable or an already configured backports suite
• Installs the Debian nvidia-cuda-toolkit package on request
• Installs both the architecture header metapackage and exact running-kernel headers when available
• Understands classic .list files and modern deb822 .sources files
• Detects Secure Boot when mokutil is available and explains MOK enrollment
• Can switch back to nouveau or remove installed nvidia-* packages through APT
• Fetches current .run installer metadata from NVIDIA instead of using a stale hardcoded version
• Logs output to /var/log/nvidia-optimizer.log

Installation

git clone https://github.com/dennishilk/nvidia-driver.git
cd nvidia-driver
chmod +x install-nvidia-cuda.sh
sudo ./install-nvidia-cuda.sh

Choose one of the five menu actions:

1. Install the Debian stable driver (recommended)
2. Install the driver from an already enabled Debian backports suite
3. Remove NVIDIA packages and enable nouveau
4. Remove NVIDIA packages and clean unused dependencies
5. Install NVIDIA's .run driver (advanced, amd64 only)

After a driver change, reboot and verify:

nvidia-smi
nvcc --version  # only when the CUDA Toolkit was installed

APT source requirements

The script accepts both /etc/apt/sources.list entries and deb822 files such as /etc/apt/sources.list.d/debian.sources.

For example, a classic Debian 13 source line contains:

deb http://deb.debian.org/debian trixie main contrib non-free non-free-firmware

Replace trixie with bookworm on Debian 12. Backports option 2 additionally expects ${VERSION_CODENAME}-backports, such as trixie-backports.

Secure Boot

With Secure Boot enabled, the NVIDIA DKMS module may not load until its Machine Owner Key is enrolled. If Debian created /var/lib/dkms/mok.pub, enroll it with:

sudo mokutil --import /var/lib/dkms/mok.pub

Then reboot and complete enrollment in the firmware's MOK Manager. If mokutil is missing, install the Debian package of the same name first.

Troubleshooting

Unable to locate package nvidia-cuda-toolkit

The Debian package is named nvidia-cuda-toolkit, not cuda. Confirm that non-free is enabled, then refresh APT:

sudo apt update
apt-cache policy nvidia-cuda-toolkit

DKMS build fails

Check whether headers exist for the running kernel:

uname -r
apt-cache policy "linux-headers-$(uname -r)"

If only the newer architecture header metapackage was available, reboot into the newly installed Debian kernel and run the driver installation again.

Black screen or login loop

From a recovery shell or TTY, inspect:

cat /var/log/nvidia-optimizer.log
dkms status
sudo mokutil --sb-state

Secure Boot, a failed DKMS build, or an old manually installed .run driver are the usual causes.

Uninstall or return to nouveau

Run the script again and choose option 3 or 4. Cleanup is performed through APT; the script does not manually erase /var/lib/dkms or user-owned NVIDIA configuration files.

License and warranty

MIT License. See LICENSE.

The software is provided "as is", without warranty of any kind. Use it at your own risk; the author is not responsible for damage, data loss, or other issues caused by its use.