Illustration of the distinction between higher-level components of the AI stack—such as the CUDA drivers, which are part of the execution container and homogenous across a training job—and the lower-level components involved in the training job, which are upgraded using maintenance trains in a rolling fashion, ensuring compatibility in a rollout between previous and new versions.