Docker GPU Fix (could not select device driver "nvidia")

If your containers fail to start with errors like:

could not select device driver "nvidia" with capabilities: [[gpu]]

or

could not select device driver "nvidia" with capabilities: [[gpu compute video]]

this is usually a Docker daemon/runtime issue (and in my case, contrary to what I initially assumed, an issue with the container I tried to run, i.e., Immich).

What was actually wrong

• Docker was running on a daemon/context without NVIDIA runtime support.
• The user running Docker did not have permission to access the native Docker socket (/var/run/docker.sock) and was not in the docker group.

Even with NVIDIA drivers and nvidia-ctk installed, Docker GPU requests fail until both are fixed.

What fixed it

1. Add your user to the docker group:

sudo usermod -aG docker $USER
newgrp docker

2. Switch to the native Docker Engine context:

docker context use default

3. Ensure NVIDIA runtime is configured for Docker:

sudo nvidia-ctk runtime configure --runtime=docker
sudo systemctl restart docker

4. Verify Docker can use the GPU before starting Immich (replace <cuda-tag> with a CUDA tag available for your setup, for example 12.4.1):

docker --context default run --rm --gpus all nvidia/cuda:<cuda-tag>-base-ubuntu22.04 nvidia-smi

If this command works, GPU passthrough is fixed.

5. Start:

docker --context default compose up -d
docker --context default compose ps

Quick sanity checks

• Confirm you are on the right context:

docker context ls

• Confirm Docker sees runtimes:

docker info | grep -i runtime -A2

• Confirm host GPU works outside Docker:

nvidia-smi

Notes

• If you are on Docker Desktop Linux, GPU support may not behave like native Docker Engine. Use the native daemon (default context) for NVIDIA workloads.
• Including --context default every single time is not necessary once you've set it.
• Tested on CachyOS, Nvidia Driver 590.48.01, CUDA 13.1 using the nvidia-container-toolkit 1.18.2-1 from the AUR.