vLLM's bidirectional KV transfer requires no additional hardware, and can be implemented with a KV-cache-aware router and EFA-based point-to-point tensor transmission.