Hugging Face launches native-speed vLLM backend

The new backend integrates transformers for high-performance inference. It is designed to run at native speed directly within vLLM.