Should I just be trying to figure out how to use the docker images from NGC directly, launch them directly, figure out the jupyter notebook tunneling from them etc the hard way?
What am I getting by carrying on trying to use this thing? The git integration is not that helpful. Ultimately I would like to be able to run pytorch/cuda stuff from a container that works on the Spark or two Sparks, and then also works when I launch it in a cloud in a proper 8xH200 or whatever.
I would like it to be helpful, it’s kinda comforting, but I’m concerned that it’s just a toy to onboard people and I’m wasting my time trying to do things with it.
Thank you for asking this. I have been confused by the AI Workbench (and the spark in general).
I thought I was getting a push button platform where I could follow a pretty simple instruction set to install a project via the AI Workbench (lets say Llama factory), spin it up, and within a small set of time and simple instructions have a way to ‘do the thing’. Refine a model, have an AI chatbot I could use… etc.
Instead I have mostly had to keep dropping back to terminal and do all of this via the command line.
Can I get it to work? Yeah. But what is the point of the spark if there is ‘ease of use’ improvement.
I have spent 5-6 hours over the last week working on creating a couple projects to drop in install some of these apps but then I keep running into they dont recognize the chip and need all this special work to get it going. Can I get it working? yeah eventually. But then why bother to bend over backwards to make something work if you thought the machine was supposed to be a bit push button easy.
I don’t think it was ever going to be easy, but I am not sure how much easier AI Workbench really makes it. Especially since the AI Workbench images seem to be slightly different from the regular ngc docker images, and the AI Workbench ones have not been keeping up, I’m rapidly running out of runway with this tool and going to have to do it all on the CLI etc anyway.