Considering the server version of llama.cpp for a commercial use case over bulkier options like vllm. But wondering if it's been battle-tested in production environments.