Mooncake supported communication protocols (Transfer Engine docs)

The transport-path table, and the direct answer to 'what is the KV transfer path actually'. Protocols with hardware requirement: tcp (default, works everywhere), rdma (RDMA-capable NIC), efa (AWS EFA, libfabric SRD), nvmeof, nvlink (NVIDIA MNNVL, inter-node GPU-to-GPU), nvlink_intra (intra-node NVLink), hip (AMD ROCm), barex, cxl, shm (POSIX shared memory, same host, explicitly 'same-host DRAM copies without NIC loopback'), musa (Moore Threads MTLink), ascend (Huawei NPU), hylink (Hygon DCU), tpu (experimental TENT), mpcomm and flagcx. Python API support is marked primary for tcp/rdma/efa and advanced for the rest. This is what determines achieved bandwidth per deployment, and shm/nvlink_intra are why a single-node PD deployment does not pay NIC bandwidth.