Java vLLM-like framework claim 90% perfomance of llamacpp inference on Nvidia HW