Containerize a GPU ML pipeline with Docker. How do you match CUDA to drivers and keep it reproducible?

The screen is whether you know the container ships the CUDA toolkit but NOT the driver: the host's NVIDIA driver is exposed through the container runtime, and the only hard constraint is driver >= the toolkit's minimum. Candidates who think you 'install the driver in the Dockerfile' or who can't explain why the same image runs on two driver versions reveal they've never actually shipped GPU containers. The strong answer pins toolkit and Python deps for reproducibility while leaving the driver as the host's job, and uses multi-stage builds to keep the runtime image lean.