microsoft/LoRA ? reverse-engineered prompt

Build me a small Python library for LoRA that works with PyTorch models, especially transformer models like GPT 2, RoBERTa, and DeBERTa.

I want it to let me freeze the original model weights and add low rank trainable layers so I can fine tune large language models with way fewer parameters, while keeping inference basically the same speed. Please include a clean installable package, plus a few working examples that show how to plug it into common Hugging Face style models for text classification and text generation.

It would be great if the examples cover the kind of tasks people would actually use this for, like GLUE style classification and GPT 2 generation, and if there are simple scripts or notebooks that make it easy to reproduce the results. Keep it focused on PyTorch only for now, and make the code easy to understand and modify. If you need to check the latest docs for anything, feel free to look them up online.