# Runtime lock

Use the Unsloth Kimi-K3 full-size llama.cpp branch at immutable commit
`efc8bc38f0a9950cbb10ccef2cf48b951c39d3b2`.

Build configuration:

```text
GGML_CUDA=ON
GGML_RPC=ON
GGML_RPC_RDMA=ON
GGML_NATIVE=OFF
CMAKE_BUILD_TYPE=Release
CMAKE_CUDA_ARCHITECTURES=121a-real
```

The reference setup uses the local CUDA device plus three RPC CUDA devices
over a private RDMA/RoCE fabric. It does not expose worker RPC ports publicly.