this post was submitted on 21 Apr 2024
188 points (94.3% liked)
Selfhosted
60366 readers
937 users here now
A place to share alternatives to popular online services that can be self-hosted without giving up privacy or locking you into a service you don't control.
Rules:
-
Be civil.
-
No spam.
-
Posts are to be related to self-hosting.
-
Don't duplicate the full text of your blog or readme if you're providing a link.
-
Submission headline should match the article title.
-
No trolling.
-
Promotion posts require active participation, with an account that is at least 30 days old. F/LOSS without a paywall has exceptions, with requirements. See the rules link for details.
Resources:
- selfh.st Newsletter and index of selfhosted software and apps
- awesome-selfhosted software
- awesome-sysadmin resources
- Self-Hosted Podcast from Jupiter Broadcasting
Any issues on the community? Report it using the report flag.
Questions? DM the mods!
founded 3 years ago
MODERATORS
you are viewing a single comment's thread
view the rest of the comments
view the rest of the comments
I tried doing that on my home server, but running it on the CPU is super slow, and the model won't fit on the GPU. Not sure what I'm doing wrong
Sadly, can't really help you much. I have a potato pc and the biggest model I ran on it was Microsoft phi-2 using the candle framework. I used to tinker with Llama.cpp on colab, but it seems they don't handle llama3 yet. ollama says it does , but I've never tried it before. For the speed, It's kinda expected for a 70b model to be really slow on the CPU. How much slow is too slow ? I don't really know...
You can always try the 8b model. People says it's really great and even replaced the 70b models they've been using.
Show as in I waited a few minutes and finally killed it when it didn't seem like it was going anywhere. And this was with the 7b model...
It shouldn't happen for a 8b model. Even on CPU, it's supposed to be decently fast. There's definitely something wrong here.
Hm... Alright, I'll have to take another look at it. I kinda gave up, figuring my old server just didn't have the specs for it
Specs? Try mistral with llama.ccp.
It has a Intel Xeon E3-1225 V2, 20gb of ram, and a Strix GTX 970 with 4gb of VRAM. I've actually tried Mistral 7b and Decapoda Llama 7b, running them in Python with Huggingface's Transformers library (from local models)
Yeah, it's not a potato but not that powerful eaither. Nonetheless, it should run a 7b/8b/9b and maybe 13b models easily.
That's your problem right here. Python is great for making llms but is horrible at running them. With a computer as weak as yours, every bit of performance counts.
Just try ollama or llama.ccp . Their github is also a goldmine for other projects you could try.
Llama.ccp can partially run the model on the gpu for way faster inference.
Piper is a pretty decent very lightweight tts engine that can be directly run on your cpu if you want to add tts capabilities to your setup.
Good luck and happy tinkering!
Ah, that's good to know! I'll give those other options a shot. Thank you so much for taking the time to help me with that! I'm very new to the whole LLM things, and sorta figuring it out as I go
Completely forgot to tell you to only use quantized models. Your pc can run 4bit quantized versions of the models I mentioned. That's the key for running llms on at consumer level hardware. You can later read further about the different quantizations and toy with other ones like Q5_K_M and such.
Just read phi-3 got released and apparently it's a 4B that reach gpt 3.5 level. Follow the news and wait for it to be add to ollama/llama.ccp
I became fascinated with llms after the first AI booms but all this knowledge is basically useless where I live, so might as well make it useful by teaching people what i know.