Multi-tenant QoS in inference scenarios refers to the system providing differentiated service-level guarantees for multiple tenants sharing the same GPU pool, usually requiring three layers of mechanisms: admission control and quota management, priority-based request scheduling, and resource isolation.