What I still don't get, is how you would create different queues, targeting different nodes with different GPUs, and having them using the appropriate Cuda image.
Looking at the template, I don't understand how that's possible.