Download ML thing.
make new venv.
pip install -r requirements.txt.
pip can't find the right versions.
pip install --update pip.
pip still can't find the right versions.
install conda.
conda breaks for some reason.
fix conda.
install with conda.
pytorch won't compile with CUDA support.
install 2,000,000GB of nvidia crap from conda.
pytorch still won't compile.
install older version of gcc with conda.
pytorch still won't compile.
reinstall the entire operating system with debian 11.
apt can't find shitlib-1.
install shitlib-2.
it's not compatible with shitlib-1.
compile it from source.
automake breaks.
install debian 10.
It actually works.
"Join our discord to get the model".
give up.

[–] [email protected] 18 points 2 years ago

It feels like you stood behind me yesterday, taking notes.

[–] [email protected] 15 points 2 years ago

This comment gives me ptsd

[–] HarryEffingPotter 4 points 2 years ago* (last edited 2 years ago)

thats when you do

/usr/bin/python3.11 -m pip install

mother. fucking. hardcoded paths. 1 step forward, 10 steps backward.

[–] kabat 1 points 2 years ago

I recommend distrobox for adhoc distrohopping. Though for Nvidia stuff it links to drivers and cuda that you have installed on your host, so... I recently needed cuda 11.8 and that was hella fun to get going.