Building LLM inference libraries in pure Java and run them with LangChain4j locally on GPUs (No CUDA, No C++)