Results 3 comments of Elvis Nava

I was using CUDA 8.0, I now tried using a Docker container with CUDA 7.5 and can confirm it works. Apparently CUDA 8.0 silently breaks whatever functionality the instance_norm models...

I'm also having the same problem. It is somehow related to this step: > 3. Prepare the pretrained MiniGPT-4 checkpoint > To play with our pretrained model, download the pretrained...

Ok, I fixed this. So, apparently, you have to use an absolute path to your the vicuna checkpoint (starting with "/"), otherwise it gets interpreted as a hosted huggingface model.