Submitted 5 months ago by plsnotracking@lemmy.world to selfhosted@lemmy.world
Hi,
I wanted to run some Large Language Models locally. Something like Private GPT or Medium Article on my local Apple Silicon to enhance my privacy but also get some additional help.
Does anyone have recommendations or guides I could follow?
Thank you very much.
New Lemmy Post: Recommendations on running GPTs on Asahi - M1 Ultra? (https://lemmy.world/post/10881869)
Tagging: #SelfHosted
(Replying in the OP of this thread (NOT THIS BOT!) will appear as a comment in the lemmy discussion.)
I am a FOSS bot. Check my README: https://github.com/db0/lemmy-tagginator/blob/main/README.md
Guenther_Amanita@feddit.de 5 months ago
I don’t know what’s your intention.
I’m no expert or highly qualified in any way, so please correct me, but I don’t know if that’s the right way.
LLMs usually need lots of computing power, optimally in form of a GPU.
I use GPT4All, and when I send a prompt, I notice the temps of my GPU turning up instantly. If it’s a longer one, my PC sounds like a helicopter 😁
In terms of hosting a server, you want something barely good enough for your service, e.g. running your cloud. This results in way less power draw, which is what you want, since it runs 24/7. Something powerful enough to run LLMs comfortably would likely draw lots of power, even an Apple Silicon.
I think, you’re better off just using GPT4All on your gaming PC if you need it.
I hope I’m wrong, and M1s draw barely any power.
And even if I am, they (almost) can only run MacOS, which wouldn’t be a good server OS.
moonpiedumplings@programming.dev 5 months ago
The tldr as I understand it is that Mac M1/M2 devices are unique in that the vram (gpu ram) is the same as the normal ram. This sharing allows LLM models to run on the gpu of those chips, and in their “vram” as well.
Llama.cpp was the software that users do this. I can’t find the original guide/article I looked at, but here is a github gist, where the commenters have done benchmarks:
gist.github.com/…/e8d4cb0c4b1df6cc47ce8b18457ebde…
Guenther_Amanita@feddit.de 5 months ago
Alright, interesting… As I said, I’m no expert or anything and this was just my noob optinion.
Thank you for the correction and further resources!