Selfhosted @lemmy.world TheBigBrother @lemmy.world 4 days ago

Is it possible to run a LLM on a mini-pc like the GMKtec K8 and K9?

I have experience in running servers, but I would like to know if it's possible to do it, I just need a GPT 3.5 like private LLM running.

33 comments

It's doable. Stick to the 7b models and it should work for the most part, but don't expect anything remotely approaching what might be called reasonable performance. It's going to be slow. But it can work.

To get a somewhat usable experience you kinda need an Nvidia graphics card or an AI accelerator.
- Intel Arc also works surprisingly fine and consistently for ML if you use llama.cpp for LLMs or Automatic for stable diffusion, it's definitely much closer to Nvidia in terms of usability than it is to AMD
  
  Would you suggest the K9 instead of the K8?
- I need it to make academic works pass the anti-AI systems, what do you recommend for that work? It's for business so I need a reasonable good performance but nothing extravagant..
  
  I believe commercial LLMs have some kind of watermark when you apply AI for grammar and fixing in general, so I just need an AI to make these works undetectable with a private LLM.
  
  Maybe just write the academic works yourself, then they should pass.
  
  I believe commercial LLMs have some kind of watermark when you apply AI for grammar and fixing in general, so I just need an AI to make these works undetectable with a private LLM.
  
  That's not how it works, sorry.
  
  Something with a GPU that's good for LLMs would be best.
They're Ryzen processors with "AI" accelerators, so an LLM can definitely run on hardware on one of those. Other options are available, like lower powered ARM chipsets (RK3588-based boards) with accelerators that might have half the performance but are far cheaper to run, should be enough for a basic LLM.
- I don't know of any project that already supports that AI processor. You'd still be using the CPU and GPU at the moment.
- The K8 it's Ryzen, the K9 Intel, money isn't a problem and it's not a spending it's a investment I need it for business, which of these two models would you recommend for a reasonable good LLM and Stable Diffusion?
  
  I'm looking for the most cost-effective solution.
Look into ollama. It shouldn't be an issue if you stick to 7b parameter models
- Yeah, I did see something related to what you mentioned and I was quite interested. What about quantized models?
  
  Quantized with more parameters is generally better than floating point with fewer parameters. If you can squeeze a 14b parameter model down to a 4-bit int quantization it'll still generally outperform a 16-bit Floating Point 7b parameter equivalent.
  
  I don't have any experience with them honestly so I can't help you there

You've viewed 33 comments.