This section is internal implementation - we can't guarantee it will not be changed. As for unused GPU - in general if you run a task with the agent having the --gpu switch a GPU will be allocated for as long as the task is running. I think the main concern is trying to make sure your task makes the most out of the GPU...?