PaddlePaddle/PaddleNLP ? reverse-engineered prompt

Build me a Python NLP toolkit on top of PaddlePaddle that makes it easy to train, fine tune, and run language models from one place.

I want it to feel like a practical model library with a good set of pretrained models, simple training and inference helpers, and clear examples for common tasks like text classification, embeddings, question answering, information extraction, and large language model use. It should work well on different hardware, support distributed training, and include a fast path for inference and lightweight compression so bigger models are still usable. Please make the project clean, documented, and testable, with a README that shows how to install it, use it, and get started quickly. If you need to check current Paddle docs or model details online, do that.