Implement MiniGPT-v2

[abrichr]

Feature request

We would like to implement https://minigpt-v2.github.io/ as a CompletionProvider.

See also https://www.reddit.com/r/MachineLearning/comments/17b7iyy/r_minigptv2_large_language_model_as_a_unified/

Motivation

After our three-stage training, the experimental results show that MiniGPT-v2 achieves strong performance on many visual question answering and visual grounding benchmarks compared to other vision-language generalist models.

v2 is a developed version that can ground objects that it speaks about as it generates. text. From this aspect, it ground objects better than GPT4-V; you may see Jun Chen (my student)'s post https://twitter.com/garvinchen2/status/1714113425561784559