I have LibreTranslate running fine with the official docker container but would like to enabled CUDA with it.

So I downloaded the libretranslate:latest-cuda container image instead and configured it to pass through the Nvidia GPU I have.

The container starts fine and I could confirm that the GPU passthrough works by executing nvidia-smi in the container, but I noticed that it doesn't actually start libretranslate itself in that container.

On investigation it seems like the latest-cuda image seems to have an issue with all the files in the /app folder being owned by the root user, instead of the libretranslate user like in the normal non-CUDA container image. It also seems to miss the venv setup.

When I try to execute the libretranslate file in the container manually it fails with an permission issue.

Looking at the docker.cuda file it seems like the container does not switch to another user and thus it should run as root, but maybe it can't access the /dev/nvidia0 devices or so?