Loads model weights into a neural network. This is currently a no-op placeholder to keep the module compiling in environments where GGUF loading is not available. It preserves API compatibility.