salesforce/BLIP ? reverse-engineered prompt

Build me a working version of this BLIP project so I can run image captioning, visual question answering, and image text retrieval from one codebase.

I want a simple setup where I can install the requirements, open the demo notebook, and try the model on my own images without fighting the environment. Please make sure the training and evaluation scripts for captioning, retrieval, VQA, and NLVR2 all run cleanly with the config files, and that the pretraining script is wired up too.

If anything is outdated or missing, clean it up so the repo is easy to use, and keep the README style instructions clear enough that I can follow them. If you need to check current docs or make small fixes to dependencies or notebook usage, do that. I mainly want something that works end to end for testing pretrained checkpoints, running the demo, and fine tuning on my own datasets.