Install AMD ROCm on a GPU server

ROCm is AMD's open software stack for GPU computing: drivers, runtimes, compilers and libraries used by PyTorch, llama.cpp, Ollama and many other tools. This guide follows AMD's quick start for Ubuntu 24.04 on a GPU server with an AMD GPU.

Before you start

• A GPU VPS, GPU VDS or GPU dedicated server with an AMD GPU that AMD lists as supported by the ROCm release you install.
• Ubuntu 24.04 LTS, fully updated and rebooted. AMD also documents Debian 12 and 13 and RHEL-based systems; use the commands on AMD's page for those.
• Check that the GPU is visible: lspci | grep -iE "vga|display|3d".

Step 1: Register AMD's repository

AMD's quick start installs a small package that adds its repositories. At the time of writing the page uses this file for Ubuntu 24.04 (noble); copy the current file name from AMD's quick start page:

wget https://repo.radeon.com/amdgpu-install/7.2.4/ubuntu/noble/amdgpu-install_7.2.4.70204-1_all.deb
sudo apt install ./amdgpu-install_7.2.4.70204-1_all.deb
sudo apt update

Step 2: Install the kernel driver

sudo apt install "linux-headers-$(uname -r)" "linux-modules-extra-$(uname -r)"
sudo apt install amdgpu-dkms

Step 3: Install ROCm and set permissions

sudo apt install python3-setuptools python3-wheel
sudo usermod -a -G render,video $LOGNAME
sudo apt install rocm

The render and video groups give your user access to the GPU devices. Reboot to load the driver and apply the groups:

sudo reboot

Step 4: Verify

rocminfo | grep -i "Marketing Name:"
clinfo | grep -i "Board name:"
amd-smi version

Verify: rocminfo and clinfo name your GPU, and amd-smi version prints its versions.

Step 5: Libraries and paths (when needed)

AMD's post-installation steps add the ROCm libraries to the system linker:

echo -e "/opt/rocm/lib\n/opt/rocm/lib64" | sudo tee /etc/ld.so.conf.d/rocm.conf
sudo ldconfig

If a tool is not found, add the ROCm binaries to your path, for example export PATH=$PATH:/opt/rocm/bin in ~/.bashrc.

Troubleshooting

rocminfo shows only the CPU agent, or reports permission errors. Your user is not in the render and video groups yet, or you have not logged in again since adding them. Check with groups.

The amdgpu-dkms build fails. The headers for the running kernel are missing. Install linux-headers-$(uname -r) and reboot into the newest kernel before you try again.

Packages are not found after apt update. The repository for your Ubuntu release is not registered, or you used an installer file for another release. Use the file for noble (Ubuntu 24.04) from AMD's page.

The GPU is not supported. ROCm releases support a defined list of GPUs. Check AMD's compatibility list; for unsupported GPUs, tools with a Vulkan backend may still work.

More in GPU troubleshooting.

Next steps

• Containers: GPUs in Docker containers.
• Run models: install Ollama.