Have you seen the new optimizer?

the way this usually works is that I spend so much time tweaking individual words in the second sentence that by the time I get to the third I’ve already forgotten the precise phrasing I used in the first

I’ve been meaning to actually host some stuff on it for a while to for friends (that was the original use case) but the slow entropic decay that comes with running one random project after another on the same machine has not really encouraged that

that’s evolved a bit over time but the most-reduced answer I have for that, realistically, is “make it easy spin up a bunch of different environments for random projects”

every homelab guide is directed towards somebody who just heard of Linux 5 minutes ago and what I need is a guide on how to build my own hassle-free cloud service provider from scratch

fucked up a distro upgrade and at this point I'm like. you know. this kind of stuff was fun to fix when I was 22 but now I'm getting pretty sick of it

this looks super cool, addresses the biggest pain points I’ve had using these kinds of frameworks (mostly lack of idempotence/statelessness), and also I am probably going to have a dozen questions about it with which to pester you

oh hell yeah

(er, on a CUDA device that is)

tbqh despite all of this and how old it is I think this repo is objectively easier to build and run than like 80% of other grad student machine learning code lmao, uv has made this quite nice

kinda incredible that you still can't natively do a matmul on int32s or int64s in PyTorch to this day

difficult to explain how shaky the library this was built off of was, what a ritual it was to extend to distributed training, how much I bled trying to write this fuckingisadng JIT-compiled PyTorch extension just so that I could get a proper GeMM for integer tensors