Installation des pilotes NVIDIA GPU et CUDA sur Ubuntu Linux¶

Nous supportons officiellement les cartes Nvidia uniquement sur les versions LTS d'Ubuntu 22.04/24.04/26.04. Pour installer les pilotes sur d'autres distributions, veuillez vous référer aux instructions officielles des développeurs :

Information

Attention

Pour un fonctionnement correct des GPU de la série Tesla (ex: NVIDIA Tesla T4), assurez-vous que le paramètre 'above 4G decoding', 'large/64bit BARs' ou 'Above 4G MMIO BIOS assignment' est activé dans le BIOS du serveur.

Ouvrez la console du serveur (via la console native ou SSH), connectez-vous en tant que root, copiez le script ci-dessous, collez-le dans la ligne de commande et appuyez sur Enter pour installer automatiquement les pilotes et CUDA. Si vous avez besoin de docker ou docker compose pour faire fonctionner votre logiciel, installez-les AVANT l'installation des pilotes afin que le script puisse assurer le "pass-through" du support GPU vers les conteneurs.

Remarque

Si vous utilisez un shell en ligne de commande autre que bash (ex: zsh ou autre), passez à bash pour exécuter le script en exécutant exec bash.

Attention

Pendant l'installation, il se peut que vous deviez appuyer sur Enter pour confirmer l'installation de nouveaux modules du noyau ou le redémarrage de services.

Script d'installation¶

#!/bin/bash
#===============================================================================
# Universal NVIDIA Driver + CUDA Installer for Ubuntu 22.04/24.04/26.04 LTS
# - Legacy GPUs (CC < 7.5): pinned driver + CUDA 12.9 (last branch supporting Pascal/Volta)
# - Blackwell GPUs: open kernel modules + latest CUDA
# - All other GPUs (incl. RTX A5000, CC 8.6): proprietary driver 580 + latest CUDA
# - cuda-drivers meta-package is NEVER installed (avoids pulling open modules)
#===============================================================================
set -euo pipefail

#-------------------------------------------------------------------------------
# CONFIGURATION FLAGS
#-------------------------------------------------------------------------------
DO_OS_POLICY_CHECK=1
ALLOWED_UBUNTU_VERSIONS=("22.04" "24.04" "26.04")

DO_APT_UPGRADE=1
DO_INSTALL_HWE=1
DO_INSTALL_KERNEL_HEADERS=1
DO_INSTALL_BUILD_TOOLS=1

DO_PURGE_OLD_PACKAGES=1
DO_SETUP_CUDA_REPO=1
DO_INSTALL_CUDA_STACK=1

# Legacy GPUs: pinned driver per card (CC < 7.5, CUDA 12.x only)
LEGACY_GPUS=(
  "10de:1b06|GeForce GTX 1080 Ti|nvidia-driver-535"
  "10de:1db1|Tesla V100|nvidia-driver-535"
  "10de:1b80|GeForce GTX 1070|nvidia-driver-535"
  "10de:1c03|GeForce GTX 1060|nvidia-driver-535"
)

# Blackwell GPUs: require open kernel modules
BLACKWELL_GPUS=(
  "10de:2b85|GeForce RTX 5090"
  "10de:2b87|GeForce RTX 5090 D"
  "10de:2b84|GeForce RTX 5080"
  "10de:2b83|GeForce RTX 5070 Ti"
  "10de:2b82|GeForce RTX 5070"
  "10de:2f80|GeForce RTX 5070 Laptop GPU"
  "10de:2bb5|RTX PRO 6000 Blackwell Server Edition"
  "10de:2bb4|RTX PRO 6000 Blackwell Workstation Edition"
  "10de:2bb1|RTX PRO 6000 Blackwell Max-Q"
  "10de:2bb3|RTX PRO 5000 Blackwell"
  "10de:2c31|RTX PRO 4500 Blackwell"
  "10de:2c37|RTX PRO 4000 Blackwell"
  "10de:2c36|RTX PRO 3000 Blackwell"
  "10de:2d30|RTX PRO 2000 Blackwell"
  "10de:2c35|RTX PRO 2000 Blackwell (Alternative)"
  "10de:2c3a|RTX PRO 4500 Blackwell Server Edition"
  "10de:2c77|RTX PRO 5000 Blackwell Embedded GPU"
  "10de:2c38|RTX PRO 5000 Blackwell Generation Laptop GPU"
  "10de:2f38|RTX PRO 3000 Blackwell Generation Laptop GPU"
  "10de:2db8|RTX PRO 1000 Blackwell Generation Laptop GPU"
)

DO_BLACKLIST_NOUVEAU=1
DO_TRY_RMMOD_NOUVEAU=1
DO_USER_GROUPS=0
DO_BASHRC_CUDA_PATHS=1
DO_VERIFY_NVIDIA_SMI=1
DO_VERIFY_NVCC=1
DO_INSTALL_NVIDIA_CONTAINER_TOOLKIT=1
DO_CONFIGURE_DOCKER_RUNTIME=1
DO_RESTART_DOCKER=1

# ---- Start ----
echo "Starting NVIDIA driver + CUDA installation (Non-Interactive)..."

#-------------------------------------------------------------------------------
# DEPENDENCY CHECK
#-------------------------------------------------------------------------------
for cmd in lspci wget gpg curl sed awk uname; do
  command -v "$cmd" >/dev/null 2>&1 || { echo "Missing dependency: $cmd"; exit 1; }
done

#-------------------------------------------------------------------------------
# OS DETECTION
#-------------------------------------------------------------------------------
osr="/etc/os-release"
[[ -r "$osr" ]] || osr="/usr/lib/os-release"
[[ -r "$osr" ]] || { echo "Cannot read os-release file. Exiting."; exit 1; }

set -a
. "$osr"
set +a

[[ "${ID:-}" != "ubuntu" ]] && { echo "This script is intended for Ubuntu only. Exiting."; exit 1; }

UBUNTU_VERSION="${VERSION_ID:-}"
UBUNTU_CODENAME="${VERSION_CODENAME:-}"

if [[ -z "${UBUNTU_CODENAME}" ]]; then
  case "${UBUNTU_VERSION}" in
    "22.04") UBUNTU_CODENAME="jammy" ;;
    "24.04") UBUNTU_CODENAME="noble" ;;
    "26.04") UBUNTU_CODENAME="resolute" ;;
    *) UBUNTU_CODENAME="unknown" ;;
  esac
fi

[[ -z "${UBUNTU_VERSION}" ]] && { echo "Cannot detect Ubuntu VERSION_ID. Exiting."; exit 1; }

if [[ "${DO_OS_POLICY_CHECK:-0}" -eq 1 ]]; then
  ok=0
  for v in "${ALLOWED_UBUNTU_VERSIONS[@]}"; do
    [[ "${UBUNTU_VERSION}" == "${v}" ]] && { ok=1; break; }
  done
  [[ "${ok}" -ne 1 ]] && { echo "Unsupported Ubuntu version: ${UBUNTU_VERSION}"; exit 1; }
fi

RELEASE_VERSION="$(echo "${UBUNTU_VERSION}" | sed 's/\([0-9]\+\)\.\([0-9]\+\)/\1\2/')"
echo "OS: Ubuntu ${UBUNTU_VERSION} (${UBUNTU_CODENAME})"

#-------------------------------------------------------------------------------
# GPU DETECTION
#-------------------------------------------------------------------------------
NVIDIA_GPU_LINES="$(lspci -nn | grep -iE 'vga|3d' | grep -i '10de:' || true)"
[[ -z "${NVIDIA_GPU_LINES}" ]] && { echo "No NVIDIA GPU detected (vendor 10de). Exiting."; exit 1; }

echo "NVIDIA GPUs detected:"
echo "${NVIDIA_GPU_LINES}"

DETECTED_GPUS="$(lspci -nn)"
REBOOT_REQUIRED=0

#-------------------------------------------------------------------------------
# SYSTEM PREPARATION
#-------------------------------------------------------------------------------
export DEBIAN_FRONTEND=noninteractive
export NEEDRESTART_MODE=a

if [[ "${DO_APT_UPGRADE}" -eq 1 ]]; then
  echo "Updating system packages..."
  sudo -E apt-get update
  sudo -E apt-get upgrade -y \
    -o Dpkg::Options::="--force-confdef" \
    -o Dpkg::Options::="--force-confold"
fi

if [[ "${DO_INSTALL_HWE}" -eq 1 ]]; then
  case "${UBUNTU_VERSION}" in
    "22.04") HWE_PKG="linux-generic-hwe-22.04" ;;
    "24.04") HWE_PKG="linux-generic-hwe-24.04" ;;
    "26.04") HWE_PKG="" ;;
    *) HWE_PKG="" ;;
  esac
  if [[ -n "${HWE_PKG}" ]]; then
    echo "Installing HWE kernel: ${HWE_PKG}"
    sudo -E apt-get install -y \
      -o Dpkg::Options::="--force-confdef" \
      -o Dpkg::Options::="--force-confold" \
      "${HWE_PKG}"
    REBOOT_REQUIRED=1
  fi
fi

if [[ "${DO_INSTALL_KERNEL_HEADERS}" -eq 1 ]]; then
  CURRENT_KERNEL="$(uname -r)"
  echo "Installing kernel headers for: ${CURRENT_KERNEL}"
  sudo -E apt-get install -y \
    -o Dpkg::Options::="--force-confdef" \
    -o Dpkg::Options::="--force-confold" \
    "linux-headers-${CURRENT_KERNEL}" || {
    echo "Failed to install linux-headers. DKMS may fail."; exit 1; }
fi

if [[ "${DO_INSTALL_BUILD_TOOLS}" -eq 1 ]]; then
  case "${UBUNTU_VERSION}" in
    "22.04") GCC_PACKAGES=("gcc-12" "g++-12") ;;
    "24.04") GCC_PACKAGES=("gcc-13" "g++-13") ;;
    "26.04") GCC_PACKAGES=("gcc-15" "g++-15") ;;
    *) GCC_PACKAGES=("gcc" "g++") ;;
  esac
  echo "Installing build tools: ${GCC_PACKAGES[*]}"
  sudo -E apt-get install -y -o Dpkg::Options::="--force-confdef" "${GCC_PACKAGES[@]}"
fi

#-------------------------------------------------------------------------------
# PURGE OLD PACKAGES
#-------------------------------------------------------------------------------
if [[ "${DO_PURGE_OLD_PACKAGES}" -eq 1 ]]; then
  echo "Purging previous NVIDIA/CUDA installations..."
  sudo dpkg --configure -a || true
  sudo -E apt-get purge -y "nvidia-*" "libnvidia-*" "cuda-*" "nvidia-driver-*" "*cudnn*" "*nsight*" 2>/dev/null || true
  sudo -E apt-get remove --purge -y nvidia-cuda-toolkit nvidia-prime nvidia-settings 2>/dev/null || true
  sudo -E apt-get autoremove -y || true
  sudo -E apt-get --fix-broken install -y || true
  sudo -E apt-get clean -y || true
fi

#-------------------------------------------------------------------------------
# SETUP CUDA REPOSITORY
#-------------------------------------------------------------------------------
if [[ "${DO_SETUP_CUDA_REPO}" -eq 1 ]]; then
  echo "Setting up CUDA repo for ubuntu${RELEASE_VERSION}..."
  wget -q "https://developer.download.nvidia.com/compute/cuda/repos/ubuntu${RELEASE_VERSION}/x86_64/cuda-keyring_1.1-1_all.deb"
  sudo dpkg -i cuda-keyring_1.1-1_all.deb 2>/dev/null || true
  rm -f cuda-keyring_1.1-1_all.deb
  sudo -E apt-get update
fi

#-------------------------------------------------------------------------------
# BLACKLIST NOUVEAU
#-------------------------------------------------------------------------------
if [[ "${DO_BLACKLIST_NOUVEAU}" -eq 1 ]]; then
  BL_FILE="/etc/modprobe.d/blacklist-nouveau.conf"
  if [[ ! -f "${BL_FILE}" ]] || ! grep -q "^blacklist nouveau" "${BL_FILE}" 2>/dev/null; then
    echo "Blacklisting nouveau (requires reboot)..."
    sudo tee "${BL_FILE}" >/dev/null <<'EOF'
blacklist nouveau
options nouveau modeset=0
EOF
    sudo update-initramfs -u
    REBOOT_REQUIRED=1
  fi
fi

#-------------------------------------------------------------------------------
# INSTALL DRIVER
#-------------------------------------------------------------------------------
IS_BLACKWELL=0
IS_LEGACY=0
IS_PROPRIETARY=0

# 1) Blackwell: open modules
for gpu_spec in "${BLACKWELL_GPUS[@]}"; do
  IFS='|' read -r pci_id gpu_name <<< "$gpu_spec"
  if echo "$DETECTED_GPUS" | grep -q "$pci_id"; then
    echo "Detected Blackwell GPU: $gpu_name ($pci_id)"
    echo "Installing open kernel modules..."
    sudo -E apt-get install -y -o Dpkg::Options::="--force-confdef" \
      nvidia-open nvidia-driver-open nvidia-dkms-open
    IS_BLACKWELL=1
    break
  fi
done

# 2) Legacy GPUs: pinned driver
if [[ "$IS_BLACKWELL" -eq 0 ]]; then
  for gpu_spec in "${LEGACY_GPUS[@]}"; do
    IFS='|' read -r pci_id gpu_name driver_pkg <<< "$gpu_spec"
    if echo "$DETECTED_GPUS" | grep -q "$pci_id"; then
      echo "Detected legacy GPU: $gpu_name ($pci_id) -> $driver_pkg"
      sudo -E apt-get install -y -o Dpkg::Options::="--force-confdef" \
        "$driver_pkg" "nvidia-dkms-${driver_pkg#nvidia-driver-}"
      IS_LEGACY=1
      break
    fi
  done
fi

# 3) Everything else: PROPRIETARY driver 580 (RTX A5000 falls here)
if [[ "$IS_BLACKWELL" -eq 0 && "$IS_LEGACY" -eq 0 ]]; then
  echo "Detected non-Blackwell, non-legacy GPU (e.g., RTX A5000)."
  echo "Installing proprietary driver 580..."
  sudo -E apt-get install -y -o Dpkg::Options::="--force-confdef" \
    nvidia-driver-580 nvidia-dkms-580
  IS_PROPRIETARY=1
fi

#-------------------------------------------------------------------------------
# INSTALL CUDA TOOLKIT
# - Legacy (CC < 7.5): CUDA 12.9 (last branch supporting Pascal/Volta)
# - Blackwell / Proprietary (CC >= 7.5): latest cuda-toolkit meta-package
# - cuda-drivers is NEVER installed
#-------------------------------------------------------------------------------
if [[ "${DO_INSTALL_CUDA_STACK}" -eq 1 ]]; then
  if [[ "$IS_LEGACY" -eq 1 ]]; then
    echo "Installing CUDA Toolkit 12.9 (legacy path, last branch supporting Pascal/Volta)..."
    sudo -E apt-get install -y -o Dpkg::Options::="--force-confdef" cuda-toolkit-12-9
  else
    echo "Installing latest available CUDA Toolkit (Blackwell/proprietary path)..."
    sudo -E apt-get install -y -o Dpkg::Options::="--force-confdef" cuda-toolkit
  fi
fi

#-------------------------------------------------------------------------------
# POST-INSTALL STEPS
#-------------------------------------------------------------------------------
if [[ "${DO_TRY_RMMOD_NOUVEAU}" -eq 1 ]]; then
  sudo rmmod -f nouveau 2>/dev/null || true
fi

if [[ "${DO_USER_GROUPS}" -eq 1 ]]; then
  TARGET_USER="${SUDO_USER:-$USER}"
  sudo usermod -aG video,render "${TARGET_USER}" || true
  echo "User ${TARGET_USER} added to video/render groups."
  REBOOT_REQUIRED=1
fi

if [[ "${DO_BASHRC_CUDA_PATHS}" -eq 1 ]]; then
  TARGET_USER="${SUDO_USER:-$USER}"
  TARGET_HOME="$(getent passwd "${TARGET_USER}" | cut -d: -f6)"
  TARGET_BASHRC="${TARGET_HOME}/.bashrc"
  MARKER="NVIDIA CUDA Paths"

  [[ -f "${TARGET_BASHRC}" ]] || sudo -u "${TARGET_USER}" touch "${TARGET_BASHRC}" || true

  if ! grep -q "${MARKER}" "${TARGET_BASHRC}" 2>/dev/null; then
    cat >> "${TARGET_BASHRC}" <<'EOF'

# NVIDIA CUDA Paths
export PATH="/usr/local/cuda/bin:$PATH"
export LD_LIBRARY_PATH="/usr/local/cuda/lib64${LD_LIBRARY_PATH:+:${LD_LIBRARY_PATH}}"
EOF
    echo "Added CUDA PATH to ${TARGET_BASHRC}"
  fi
fi

#-------------------------------------------------------------------------------
# VERIFICATION
#-------------------------------------------------------------------------------
if [[ "${DO_VERIFY_NVIDIA_SMI}" -eq 1 ]]; then
  echo "Verifying nvidia-smi..."
  nvidia-smi || true
fi

if [[ "${DO_VERIFY_NVCC}" -eq 1 ]]; then
  if command -v nvcc >/dev/null 2>&1; then
    echo "Verifying nvcc..."
    nvcc -V || true
  else
    echo "nvcc not found in PATH yet (re-login/reboot may be needed)"
  fi
fi

#-------------------------------------------------------------------------------
# NVIDIA CONTAINER TOOLKIT
#-------------------------------------------------------------------------------
if [[ "${DO_INSTALL_NVIDIA_CONTAINER_TOOLKIT}" -eq 1 ]]; then
  if command -v docker >/dev/null 2>&1; then
    echo "Docker detected: installing NVIDIA Container Toolkit..."
    sudo -E apt-get update
    sudo -E apt-get install -y -o Dpkg::Options::="--force-confdef" curl gnupg2

    sudo install -d -m 0755 /usr/share/keyrings
    curl -fsSL https://nvidia.github.io/libnvidia-container/gpgkey \
          | sudo gpg --dearmor --yes -o /usr/share/keyrings/nvidia-container-toolkit-keyring.gpg

    curl -s -L https://nvidia.github.io/libnvidia-container/stable/deb/nvidia-container-toolkit.list \
      | sed 's#deb https://#deb [signed-by=/usr/share/keyrings/nvidia-container-toolkit-keyring.gpg] https://#g' \
      | sudo tee /etc/apt/sources.list.d/nvidia-container-toolkit.list >/dev/null

    sudo -E apt-get update
    sudo -E apt-get install -y -o Dpkg::Options::="--force-confdef" nvidia-container-toolkit

    if [[ "${DO_CONFIGURE_DOCKER_RUNTIME}" -eq 1 ]]; then
      command -v nvidia-ctk >/dev/null 2>&1 && sudo nvidia-ctk runtime configure --runtime=docker || true
    fi

    if [[ "${DO_RESTART_DOCKER}" -eq 1 ]]; then
      sudo systemctl restart docker || true
    fi
  else
    echo "Docker not found. Skipping container toolkit."
  fi
fi

#-------------------------------------------------------------------------------
# FINAL
#-------------------------------------------------------------------------------
echo ""
echo "========================================"
echo "Installation finished"
echo "========================================"

if [[ "${REBOOT_REQUIRED}" -eq 1 ]]; then
  echo ">> REBOOT REQUIRED: sudo reboot"
fi

echo "After reboot, verify with:"
echo "  nvidia-smi"
echo "  nvcc -V"

Attention

Après l'exécution du script, un redémarrage du serveur via la commande sudo reboot est obligatoire. Ceci est nécessaire pour activer les nouveaux modules du noyau.

Après le redémarrage, vérifiez l'installation des pilotes et de CUDA à l'aide des commandes nvidia-smi et nvcc -V. Vous devriez voir un résultat similaire au suivant :

+-----------------------------------------------------------------------------------------+
| NVIDIA-SMI 595.71.05              Driver Version: 595.71.05      CUDA Version: 13.2     |
+-----------------------------------------+------------------------+----------------------+
| GPU  Name                 Persistence-M | Bus-Id          Disp.A | Volatile Uncorr. ECC |
| Fan  Temp   Perf          Pwr:Usage/Cap |           Memory-Usage | GPU-Util  Compute M. |
|                                         |                        |               MIG M. |
|=========================================+========================+======================|
|   0  NVIDIA GeForce RTX 4090        On  |   00000000:07:00.0 Off |                  Off |
| 31%   27C    P8              7W /  450W |      33MiB /  24564MiB |      0%      Default |
|                                         |                        |                  N/A |
+-----------------------------------------+------------------------+----------------------+

+-----------------------------------------------------------------------------------------+
| Processes:                                                                              |
|  GPU   GI   CI              PID   Type   Process name                        GPU Memory |
|        ID   ID                                                               Usage      |
|=========================================================================================|
|  No running processes found                                                             |
+-----------------------------------------------------------------------------------------+

et

nvcc: NVIDIA (R) Cuda compiler driver
Copyright (c) 2005-2026 NVIDIA Corporation
Built on Thu_Mar_19_11:12:51_PM_PDT_2026
Cuda compilation tools, release 13.2, V13.2.78
Build cuda_13.2.r13.2/compiler.37668154_0

2026-10-31