Rootless Podman GPU Passthrough for AMD ROCm

Q: Which devices does ROCm need in the container?

Two: /dev/kfd (the ROCm compute interface) and /dev/dri (the DRM render nodes). Add both with devices: in compose or --device on the CLI.

Q: I get 'permission denied' on /dev/kfd in a rootless container. Fix?

The container user must be in the host's render and video groups. Add their GIDs with group_add (find them via getent group render and getent group video ).

Q: What is HIP_VISIBLE_DEVICES for?

It selects which GPU ROCm uses. Set HIP_VISIBLE_DEVICES=0 to pin to the first/discrete GPU; check rocm-smi if an integrated GPU also appears and shifts indices.

Q: How do I verify the GPU is actually visible inside the container?

Run podman exec <name> rocminfo | grep gfx (should list your arch, e.g. gfx1201) or rocm-smi . If it errors, re-check the devices and group_add.

Getting ROCm working inside rootless Podman is mostly about two device nodes and the right group membership. Here's the complete checklist.

Minimal working service

devices: ["/dev/kfd", "/dev/dri"]
group_add: ["<render-gid>", "<video-gid>"]   # getent group render / video
security_opt: ["seccomp=unconfined"]
environment: ["HIP_VISIBLE_DEVICES=0"]

Verify from inside the container with rocminfo | grep gfx — on RDNA4 you should see gfx1201.

Technical FAQ

Which devices does ROCm need in the container?

Two: /dev/kfd (the ROCm compute interface) and /dev/dri (the DRM render nodes). Add both with devices: in compose or --device on the CLI.

I get 'permission denied' on /dev/kfd in a rootless container. Fix?

The container user must be in the host's render and video groups. Add their GIDs with group_add (find them via getent group render and getent group video).

What is HIP_VISIBLE_DEVICES for?

It selects which GPU ROCm uses. Set HIP_VISIBLE_DEVICES=0 to pin to the first/discrete GPU; check rocm-smi if an integrated GPU also appears and shifts indices.

Why do I need seccomp=unconfined?

Some ROCm/HIP ioctls are blocked by the default seccomp profile in rootless mode. security_opt: [seccomp=unconfined] is the simplest fix; a custom profile allowing the needed syscalls also works.

A bind-mounted volume is owned by root and my non-root container user can't write it. Why?

Rootless Podman maps your host user to container root by default, so a non-root in-container user can't write host-mounted paths. Use --userns=keep-id:uid=<uid>,gid=<gid> so your host user maps to that in-container user and owns the mount.

How do I verify the GPU is actually visible inside the container?

Run podman exec <name> rocminfo | grep gfx (should list your arch, e.g. gfx1201) or rocm-smi. If it errors, re-check the devices and group_add.

Do I need root or sudo to run GPU containers?

No — this is fully rootless. You only need your user in the render/video groups and the device/userns settings above. No daemon, no root.

Get the full stack on GitHub →