https://github.com/1ordo/strix-halo-unsloth?utm_source=chatgpt.com
Using a local model practically seems to be around 30b?
So I brought in a 14b model for fine-tuning and was preparing for it, but AMD is causing problems like this. The driver for the processor that has been released for over a year hasn't come out, so I didn't know there was a problem.
There's no problem using the model as it is in lmstudio, but for fine-tuning, there's nothing officially released yet, and you have to use a workaround to make fine-tuning possible.
So I googled and found that there's a way like this on GitHub, and ended up using Fedora 43 version. The other computers are all Ubuntu...
Updates were also blocked, and only a specific version of the OS, PyTorch, and a specific kernel were confirmed to work. When I first installed it with Ubuntu, no matter how hard I tried, the AI agent wouldn't work. After being told about this method, I applied it and fine-tuning was successful.
If you are using this processor and thinking about fine-tuning, just keep in mind that there is a way like this. If you don't do fine-tuning, you can just use it as it is.
AMD is rumored to be announcing something at the end of June or early July, so I expect the driver for this to be released around that time.
