Support for containerization on cloud machines

[twilwa]

So, installing this onto the nvidia NGC pytorch container that i'm running on a p3 instance, my goal is to provide a hosted version of pinokio on an isolated cloud VM. I've got most of the troubleshooting done; the torch(cuda.visible.devices) is true inside the container, nvcc is configured, and the runtime appears to be forwarded correctly. But, from /workspace in the container, running:

git clone pinokiod
cd pinokiod
npm install
npm start

will successfully launch the browser, but the console output reads as follows:

I can't quite tell from looking at the codebase, but I assume it's to do with isntalling the application to /workspace is causing the fs reads to look at the wrong part of the container filesystem to find gpu information. Any advice on how to modify?

Once I can get it figured out, I'll package it all up and push the dockerfiles/ngc instructions/etc so this can be deployed onto cloud infra with minimal setup.

[chymian]

hi @twilwa ,
did you have success with the dockerization?

[twilwa]

i ended up installing a torch container and was able to get the application to run -- it's a huge security risk though, so i'm working on a way to solve the catch-22 pinokio sort of puts you in:
'pinokio needs a priveleged container to operate + the frontend is a browser for installing applications that need to do things like install cuda drivers, etc, so if a user has access to the frontend it gives them admin access to the priveleged container, if they don't have access to the frontend/don't have admin access/the container isn't priveleged, pinokio doesn't work'

@cocktailpeanut if you have any insight on how to solve this i've got a network ready to host a ton of pinokio instances if that's something you're interested in seeing happen. current thought is to use https://katacontainers.io/ but could use a hand from the guy that wrote the application to check my logic :3