High-performance Inference and Deployment Toolkit for LLMs and VLMs based on PaddlePaddle