huggingface/transformers ? reverse-engineered prompt

I want you to build me a comprehensive Python library that makes it easy to use all the state-of-the-art machine learning models. The main goal is to let me download and run pre-trained models from the Hugging Face Hub for lots of different tasks. It needs to handle text, audio, images, and even multimodal models.

The key feature should be a really simple "pipeline" API that abstracts away the complexity. A user should just be able to specify a task, like "text-generation," and a model name, and the library will handle downloading the model, preprocessing the input, and running the inference. While the main focus is making inference easy, it should also contain the actual model definitions so that advanced users could use this library to train or fine-tune models with popular frameworks like PyTorch.