How to run docker container in AMD GPU Droplet and get access to the GPU

Hey!

What base image are you using in your container? If it’s not one of the official ROCm supported images (like rocm/pytorch), that might be part of the issue.

You’re already passing the right flags with --device=/dev/kfd and /dev/dri, but sometimes it still won’t detect the GPU unless the container has everything set up for ROCm.

Might also be worth running rocminfo or hipinfo inside the container to see if anything shows up at all.

If it’s still not working, could be something subtle with the setup. I’d suggest reaching out to DigitalOcean support at do.co/support. They might be able to confirm whether ROCm is fully set up and available for your Droplet.

- Bobby

Heya, @stevemcatee

If your container isn’t using a ROCm-ready base image (like rocm/pytorch or an Ubuntu image with ROCm installed), PyTorch won’t see the GPU. --device=/dev/kfd and /dev/dri are needed, but sometimes /dev/dri/renderD128 helps too. Always test with rocminfo or rocm-smi inside the container — if those don’t see the GPU, PyTorch won’t either.

If you’ve got the right base image and devices but it still fails, it’s worth checking with DigitalOcean support to confirm the host’s ROCm drivers are set up correctly for your Droplet type.