For Reasons ™ I need an instance that supports:
- sm_86 or below (the H100 instances only support sm_90)
- 40 GB RAM per GPU or more (so A10 instances don’t work)
I could do with only a single GPU, but two or four would be nicer.
Unfortunately, everythings all full. I clicked the “reserve instance” button, and it said it had noted the reservation …
and nothing else happened.
What does it mean when I push this button?
Is there a dashboard where I can see reservation requests?
Will this create a paid-for reservation without me noticing, starting billing in the background?
Btw, “Reasons™” is “fine tuning mpt-7b-instruct” which uses a model that requires Torch 1.13.1 which doesn’t work with CUDA12/sm90, and it breaks on Torch 2 for other reasons. I’m sure you’re familiar with this level of snake wrangling …