Obviously there’s not a lot of love for OpenAI and other corporate API generative AI here, but how does the community feel about self hosted models? Especially stuff like the Linux Foundation’s Open Model Initiative?
I feel like a lot of people just don’t know there are Apache/CC-BY-NC licensed “AI” they can run on sane desktops, right now, that are incredible. I’m thinking of the most recent Command-R, specifically. I can run it on one GPU, and it blows expensive API models away, and it’s mine to use.
And there are efforts to kill the power cost of inference and training with stuff like matrix-multiplication free models, open source and legally licensed datasets, cheap training… and OpenAI and such want to shut down all of this because it breaks their monopoly, where they can just outspend everyone scaling , stealiing data and destroying the planet. And it’s actually a threat to them.
Again, I feel like corporate social media vs fediverse is a good anology, where one is kinda destroying the planet and the other, while still niche, problematic and a WIP, kills a lot of the downsides.
Very much pro Open Source AI. Especially as a concept digital public good. With https://petals.dev/ being the most promising option that regard (imagine something like RAG for the arch wiki with very large models supported by the community!).
It feel very enthusiasts right now. Where I feel like I’m just on the cusp of having usable set up.
I personally really want a full Dev that just takes gitlab issues and runs codes against tests until it passes, and then cycles between attempting to explain what it doing and refactoring until that explanation is reasonably simple, then submit PR.
At the moment I am trying to use it as a copilot (ollama lama3, continue, and devonAI vscode plugins) all on my MacBook (my Linux machine were too small gpu wise, at least first time I attempted). That said it ok for questions no real luck on a decent experience for actually making anything.
The next step to me for it to move from enthusiast to hobbiest would be:
I feel like there has to be a better way for this. Maybe its just selinux rules for data tags for locking down my local system and some routing config file at the root of my projects. Idk tbh
Honestly I am not sold on petals, it leaves so many technical innovations behind and its just not really taking off like it needs to.
IMO a much cooler project is the AI Horde: A swarm of hosts, but no splitting. Already with a boatload of actual users.
And (no offense) but there are much better models to use than ollama llama 8b, and which ones completely depends on how much RAM your Mac has. They get better and better the more you have, all the way out to 192GB. (Where you can squeeze in the very amazing Deepseek Code V2)
None taken! I’ll check out AI Horde!
Is there any objective measured ways or at least subject reviews based metrics for a model on g8ve problem set? I know the white papers tend to include it and sometimes the git repos, but I don’t see that info when searching through ollama for example.
I saw you other post about ollama alts and the concurrency mention in one of the projects README sounds promising.
Honestly I would get away from ollama. I don’t like it for a number of reasons, including:
Suboptimal quants
suboptimal settings
limited model selection (as opposed to just browsing huggingface)
Sometimes suboptimal performance compared to kobold.cpp, especially if you are quantizing cache, double especially if you are not on a Mac
Frankly a lot of attention squatting/riding off llama.cpp’'s development without contributing a ton back.
Rumblings of a closed source project.
I could go on and on, inclding some behavior I just didn’t like from the devs, but I think I’ll stop, as its really not that bad.
Oh, and as for benchmarks, check the huggingface open llm leaderbard. The new one.
But take it with a LARGE grain of salt. Some models game their scores in different ways.
There are more niche benchmarks floating around, such as RULER for long context performance. Amazon ran a good array of models to test their mistral finetune: https://huggingface.co/aws-prototyping/MegaBeam-Mistral-7B-512k
The splitting is 80% of the cool factor for me. Rather than bog down the one node that can handle those cooler models, and have more contribution opportunities.
I wonder honestly if a petals network could be a target host on horde lol
The problem is that splitting models up over a network, even over LAN, is not super efficient. The entire weights need to be run through for every half word.
And the other problem is that petals just can’t keep up with the crazy dev pace of the LLM community. Honestly they should dump it and fork or contribute to llama.cpp or exllama, as TBH no one wants to split up LLAMA 2 (or even llama 3) 70B, and be a generation or two behind for a base instruct model instead of a finetune.
Even the horde has very few hosts relative to users, even though hosting a small model on a 6GB GPU would get you lots of karma.
The diffusion community is very different, as the output is one image and even the largest open models are much smaller. Lora usage is also standardized there, while it is not on LLM land.
I guess to me be able to serve the 408b model even though I’m on a laptop is just awesome to me.
Also I saw Lora was an option for Petals but I haven’t messed with it at all.