Transport tuning should be part of setting up your cluster (Sparks or anything else).
It means choosing how data moves between GPUs and machines, including the communication backend and network path to use. Save the transport choice with that topology so it doesn't become something to revisit every time you run inference.