cross-posted from: https://lemmy.world/post/27088416
This is an update to a previous post found at https://lemmy.world/post/27013201
Ollama uses the AMD ROCm library which works well with many AMD GPUs not listed as compatible by forcing an LLVM target.
The original Ollama documentation is wrong as the following can not be set for individual GPUs, only all or none, as shown at github.com/ollama/ollama/issues/8473
AMD GPU issue fix
- Check your GPU is not already listed as compatibility at github.com/ollama/ollama/blob/main/docs/gpu.md#linux-support
- Edit the Ollama service file. This uses the text editor set in the
$SYSTEMD_EDITORenvironment variable.sudo systemctl edit ollama.service
- Add the following, save and exit. You can try different versions as shown at github.com/ollama/ollama/blob/main/docs/gpu.md#overrides-on-linux
[Service] Environment="HSA_OVERRIDE_GFX_VERSION=10.3.0"
- Restart the Ollama service.
sudo systemctl restart ollama
I would run it in a Podman container with the GPU passed though
Why not throw that into a VM with VFIO passthrough, plug the GPU in via an external dock and if we are already at abstracting shit away for unnecessary complexity and non-compatibility do all that on windows?
Because that is way more complicated?
It is really easy to run ollama in a container.
Really easy to start running it
Then everything goes wrong, from configuration over logs to cuda. And the worst fucking debugging ever.
On Linux you can download Alpaca. I think it is CPU only but it is simpler.
Ollama is simple too, I meant that containers make everything a nightmare to maintain.
If containers are the hard part you are doing it wrong. Containers should make it much easier.
Nested VMs stay performant about three levels deep, so do that as well.