Comment by verdverm

There is a bug in llama-cpp for qwen/gemma models, use vLLM instead