The disaggregation of LLM serving has transformed GPU clusters into a complex mix of heterogeneous interconnects, such as PCIe, multi-rail RDMA, and NVLink. However, existing data transfer frameworks lack the ability to intelligently coordinate these diverse links, resulting in performance bottlenecks, reduced resilience, and increased operational overhead. This paper introduces Mooncake Transfer Engine NT (TENT), a novel data movement engine that shifts from the traditional Imperative Path Selection to a Declarative Slice Spraying approach. Instead of statically binding to a single transport, TENT automatically discovers all available interconnects and topologies, unifying them into a dynamic resource pool. At run-time, TENT’s adaptive scheduler intelligently “sprays” data slices across the healthiest and best-performing links based on real-time telemetry. Furthermore, its multi-layered resilience mechanism provides seamless failover within and across transports, enabling self-healing from link failures. Our evaluation shows that TENT simplifies operations in heterogeneous environments, increases elephant flow transfer throughput by over 33.7% compared to its predecessor. TENT is also effective in updating model weights for LLM reinforcement learning, reducing the update time by up to 50.5%.