Update: I was wrong, building the image is easy, but when you run it, you encounter numerous issues. And spent like 5 hours on this.

Here’s a summary:

static.ramith.io/gb10_af3_rep...

maybe ignoring the initial commit 😅

This would count any boilerplate code you copied from someone, though (say you start with some template/starter code) 🤔

wonder if there's any way to filter that

Now supports clusters behind VPNs... (openconnect to the rescue) -- idea is there's a button that lets you copy the vpn connection command (then you use duo etc)..

you can use this to check available gpus:

might need to adapt the script a bit because different clusters have different conventions for the information in `scontrol show node `
gist.github.com/ramithuh/132...

it's a common cluster.. didn't need any special priviledges

Essentially this executes a command equivalent to this
> ssh -i /path/to/ssh_key username@hostname "python3 ~/check_gpu_availability.py --json"

repo link: github.com/ramithuh/MLR...

1/2

this is fun to watch...

My biggest concern is using interactive jobs when prototyping... (connection interruptions, feel guilty if you are not actively using the compute node)

now, I do all the prototyping locally (on a GPU), and when it's ready, I choose a cluster that has idle GPUs and then submit the job 🤔

claude made a nice api spec..