Regarding the server's lifecycle

If I want to perform inference with a trained model, is using trainer.dev the most convenient approach? Also, I’ve noticed that the server seems to be released after trainer.fit or trainer.dev finishes—how can I extend its lifetime?