A team wants to reduce the cost of generative AI inference while maintaining stable latency at scale as user demand grows. Which OpenShift AI component is specifically designed to manage these inference costs by using GPU resources efficiently?