Hmm I needed GPU support for a training loop too, docker-compose does not yet support the “–gpu all” flag equivalent, so I had to hack a bit around my local docker container. Not sure how this will play on Github workflow anyway as I’m just doing this locally for now. I’m guessing they don’t provide gpu machines there anyway, but I’m guessing you could develop your notebook on colab or locally (with those hacks) and just push with the regular flow.
Reference issue for docker-compose
Just FYI