Add multi-gpu support, model cache, fix README and deprecated containers

[catid]

This adds several improvements:
[x] Support for consumer GPUs for hosting
[x] Local model cache for docker container to avoid re-downloading the models on restart
[x] Fixes for deprecated container images and README corrections

[bluestyle97]

Hi, thanks for you contribution! Could you please add a new app_2gpu.py instead of editing the original app.py directly and then make a new pull request?

[catid]

No I think it makes more sense in app.py for use with gradio containers. I want to be able to just deploy this not manually run a different app

[bluestyle97]

I think using separate scripts for different environments can better empowers users to choose between using a single GPU or multiple GPUs.

[catid]

Don't see your point.. either way you'd want to use CUDA_VISIBLE_DEVICES=0 to do what you're saying. So might as well just have one script

[catid]

And for the docker container you can enforce single GPU usage like this:

docker run -it -p 43839:43839 --platform=linux/amd64 --gpus '"device=0"' -v $HOME/models/:/workspace/instantmesh/models instantmesh

[catid]

So there's no reason to have a ton of ugly code duplication for this feature.. Even if you wanted to select it via script features it's better to have an argparse --argument instead

[bluestyle97]

Ok, I get your point and agree with you. Before merging the commits, I notice that the CUDA version in the docker file is upgraded from 12.1 to 12.4, while pytorch 2.1.0 is compiled with CUDA 12.1 by default. Will the change of CUDA version potentially lead to incompatible problem? How about keeping the original CUDA version unchanged?

[catid]

Sometimes it can lead to problems but I tested it and it's working great. Also I regularly run 12.4.1 on all my Ubuntu GPU servers so I have some trust in this version of CUDA. I have had issues with prior versions so I know somewhat what to look out for (usually wheel build errors from nvcc).