Docker. Don't forget Docker. I pretty much checked out the day using it became a requirement just to talk to a GPU and didn't do much with machine learning until llama.cpp came along.