How much of a performance overhead does this runtime add, anyway? Each request to a model would eat so much GPU for actual text generation, the cost to process

How much of a performance overhead does this runtime add, anyway? Each request to a model would eat so much GPU for actual text generation, the cost to process the request and response strings even in a slow, garbage-collected seems negligible both in latency.