Spent months thinking my Ryzen #5900x was broken due to reboots and #mce error logs. Turns out it was my #gpu driver.

Needlessly increased the voltage curve incrementally for nothing. Spent multiple days debugging.

Good to know #mce errors can't necessarily be trusted without further investigation and analysis.

I can now reliably trigger the #mce errors on startup by running some broken #hip workload on my #5700xt

After some trial and error, I now have AMD GPU accelerated LLMs running locally on my Ollama instance. I have an external 5700XT (gfx1010), which is not officially supported by ROCm. However, RDNA2 gfx1030 is supported.

So if you want to have a 5700XT working you need to:
- Install: sudo dnf rocblas
- Go to: /usr/lib64/rocblas/library
- Execute: sudo ln -s TensileLibrary_lazy_gfx1030.dat TensileLibrary_lazy_gfx1010.dat
- Edit systemd Ollama.service to let it know you have a 10.1.0 (that's what the gfx1010 stands for) via an environment variable by editing the file sudo nano /etc/systemd/system/ollama.service

It should look like this:
[Service]
Environment="HSA_OVERRIDE_GFX_VERSION=10.1.0"

What is cool is that the iGPU is also loaded up and it's CU cores and VRAM are available too!

#ollama #5700XT #amdgpu #llm #opensource