Gemma certainly was trained for tool calling, but the implementation in llama.cpp has been troubled because Gemma uses a different chat template format. The processor from the transformers library works fine though.