Skip to main content
Version: 3.0.0-alpha (Diátaxis)

How to install CUDA and GPU drivers

Hikube VMs with an attached GPU do not come with pre-installed NVIDIA drivers. This guide details the installation of NVIDIA drivers and the CUDA toolkit on an Ubuntu VM to leverage the GPU.

Prerequisites​

  • A Hikube VMInstance with an attached GPU (see the GPU API reference)
  • VM based on Ubuntu 24.04 (commands are tailored for this version)
  • SSH or console access to the VM
  • root or sudo privileges
No pre-installed drivers

Hikube golden images do not include NVIDIA GPU drivers. You must install them manually or via cloud-init after creating the VM.

Steps​

1. Connect to the VM​

virtctl ssh -i ~/.ssh/id_ed25519 ubuntu@my-gpu-vm

Or via direct SSH if the VM is exposed:

ssh -i ~/.ssh/hikube-vm ubuntu@<IP-EXTERNE>

2. Verify the GPU is detected​

Before installing the drivers, verify that the GPU is properly detected by the system:

lspci | grep -i nvidia

Expected output:

06:00.0 3D controller: NVIDIA Corporation ...

If no GPU appears, verify that your VMInstance has a GPU attached in its configuration.

3. Install NVIDIA drivers and CUDA​

Add the NVIDIA repository and install the drivers:

# Télécharger et installer le keyring NVIDIA
wget https://developer.download.nvidia.com/compute/cuda/repos/ubuntu2404/x86_64/cuda-keyring_1.1-1_all.deb
sudo dpkg -i cuda-keyring_1.1-1_all.deb

# Mettre à jour les paquets
sudo apt-get update

# Installer le toolkit CUDA et le driver
sudo apt-get install -y cuda-toolkit nvidia-driver-560
Reboot required

A reboot is required after installing the drivers to load the NVIDIA kernel modules.

Reboot the VM:

sudo reboot

Reconnect after the reboot (wait approximately 1 minute):

virtctl ssh -i ~/.ssh/id_ed25519 ubuntu@my-gpu-vm

4. Verify the driver installation​

nvidia-smi

Expected output:

+---------------------------------------------------------------------------------------+
| NVIDIA-SMI 560.xx.xx    Driver Version: 560.xx.xx    CUDA Version: 12.x              |
|-----------------------------------------+----------------------+----------------------+
| GPU  Name                 Persistence-M | Bus-Id        Disp.A | Volatile Uncorr. ECC |
| Fan  Temp   Perf          Pwr:Usage/Cap |         Memory-Usage | GPU-Util  Compute M. |
|=========================================+======================+======================|
|   0  NVIDIA L40S                    Off | 00000000:06:00.0 Off |                    0 |
| N/A   30C    P8              20W / 350W |      0MiB / 46068MiB |      0%      Default |
+-----------------------------------------+----------------------+----------------------+

Check the CUDA version:

nvcc --version

5. Configure environment variables (optional)​

Add CUDA to PATH for permanent access:

echo 'export PATH=/usr/local/cuda/bin:$PATH' >> ~/.bashrc
echo 'export LD_LIBRARY_PATH=/usr/local/cuda/lib64:$LD_LIBRARY_PATH' >> ~/.bashrc
source ~/.bashrc

6. Test with PyTorch​

Install PyTorch with CUDA support to validate that the GPU is fully functional:

pip3 install torch --index-url https://download.pytorch.org/whl/cu124

Test GPU detection:

python3 -c "
import torch
print(f'CUDA disponible : {torch.cuda.is_available()}')
print(f'GPU détecté : {torch.cuda.get_device_name(0)}')
print(f'Mémoire GPU : {torch.cuda.get_device_properties(0).total_mem / 1e9:.1f} Go')
"

Expected output:

CUDA disponible : True
GPU détecté : NVIDIA L40S
Mémoire GPU : 46.1 Go

Verification​

Run a simple computation on the GPU to confirm proper operation:

python3 -c "
import torch
# Créer un tenseur sur le GPU
x = torch.randn(1000, 1000, device='cuda')
y = torch.randn(1000, 1000, device='cuda')
z = torch.mm(x, y)
print(f'Calcul GPU réussi, taille du résultat : {z.shape}')
"

Expected output:

Calcul GPU réussi, taille du résultat : torch.Size([1000, 1000])
Automate installation via cloud-init

To automate GPU driver installation at startup, use the cloudInit parameter of the VMInstance. See the guide How to configure cloud-init for a complete example.

Going further​