Quite the statement that Gentoo has survived for so long compiling from source but, even with ever advancing processor speeds, they’ve finally gone "Nah… Takes to long. ".
I mean, I don’t blame them. Yesterday I left my machine building a PyTorch package for 4 hours on a 12 core processor.
Errrm, could they please leave some memory to other processes? KDE already takes about 1.5GB of VRAM on my RX7600 8GB just running a desktop (dual head 4k + 1440p displays). Yes, things can get swapped out to main memory, but that becomes choppy. I’d rather run single buffered, get the odd screen tear, and have the VRAM back for real work.
My beef tends to be with software out of FANGs. Big teams and huge codebase to match. Completely inpetetrable for the rest of us and, I suspect, far more code then there should be.
Not really. I’ve had to do quite a bit of experimentation.
My setup that I’ve settled on:
Rocm system libraries from Arch Linux
PyTorch nightly for Rocm pip installed into a venv (see instructions on pytorch homepage)
Set HSA_OVERRIDE_GFX_VERSION to 11.0.0. This is just for the RX7600 and it tells it to use the RX7900 code as the pytorch version hasn’t been compiled with 7600 support.