nvidia-nixl

# NVIDIA Inference Xfer Library (NIXL) Target audience: operators wiring NIXL into Dynamo/vLLM/SGLang clusters, plugin authors writing new backends, developers using the agent API directly from Python (`references/python-api.md`). C++/Rust developers: consult `src/api/cpp/` headers and `examples/{cpp,rust}/` upstream directly — this skill does not carry a C++/Rust API reference. Assumes datacenter-class GPUs (H100/H200/B200/B300) with NVIDIA driver, CUDA 12.8+, RDMA NIC (Mellanox/EFA) for cross-node, and Linux (Ubuntu 22.04/24.04 or Fedora). macOS and Windows are not supported. ## What NIXL is — one paragraph NIXL is the *transport*, not a cache. If the goal is a KV cache that outlives a single vLLM process, the thing being configured is **`lmcache-mp`** (same `inference-cache` plugin) — a standalone LMCache server that can use NIXL as one of its backends. Come here for the agent API, plugins, and wire-level behaviour; go there for the cache server's deployment, sizing, and ZMQ wiring. NIXL is a thin abstraction over heterogeneous transport backends. A `nixlAgent` registers memory regions (DRAM, VRAM, FILE, BLOCK, OBJ), exchanges metadata with peer agents via either ETCD or socket side-channel, then issues asynchronous one-sided `READ`/`WRITE` transfers between local and remote registered memory. The agent picks the best backend (UCX for network, GDS for storage, etc.) based on memory types and what both sides have loaded. Same-process loopback, intra-node GPU-to-GPU, and...