On-device AI inference node for the ZeroGPU network (ONNX Runtime + vendored llama.cpp).