Skip to content

Instantly share code, notes, and snippets.

View valentinsavenko's full-sized avatar
📈

Valentin valentinsavenko

📈
View GitHub Profile
@valentinsavenko
valentinsavenko / amd_gfx1030_llama.md
Last active August 1, 2025 17:15
Run llama.cpp on AMD GPU, verifiably. RX6700S (gfx1032) - Ubuntu 24

On my setup it was suprisingly hard to verify that a model is fully loaded and running on the dedicated GPU. The GPU can have some kind of activity, or the models can be just partially loaded even when the CPU is doing most of the work and on Ubuntu you'll need a handful of tools to verify rocm performance.

This is mostly a reminder for myself, but might prove useful if you have the same setup:

$ lscpu | grep -i model
Model name:                           AMD Ryzen 9 6900HS with Radeon Graphics

$ nvtop