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.
Really into local hosting and open LLM’s I’ve largely stepped back due to ‘fatigue’. I’ve downloaded tweaked and reshuffle models and programs then a couple months will pass and it’s lept forward again. Which is good but I figured I’d wait until it slowed a bit.
I will say the fact I can run a decent 7b and even 10b models and get decent responses and times with a 3070 is impressive. AnythingLLM has been a really handy program for me. Still in development but it’s been neat working with RAG. I also moved from textgen to LMStudio and am really liking it. I like textgen but I felt it got a bit side tracked. A lot of good suggestions in here so cheers OP.
You can probably run Nemo 12B pretty quickly, though llama 3.1/gemma 9b finetunes may be better tbh. Deepseek lite v2 code with offloading would still be fast, even though its a 16B, since its such a heavy MoE.
Hardware is such a limiting factor now. Once quad-channel APUs and such start coming out, I feel like it will open up the space, so people don’t have to hunt down used 3090s and built desktops around them.
Last I tried was a fimbul merge for 10.4b with rope for creative writing which was great but yeah 3.1 is where I’ve landed lately. I’ll have to check out nemo! Like you mentioned I was sitting on money to grab a 3090 but I think I’ll wait for rtx50xx to drive down prices or just for dedicated hardware. I’ll be sure to keep an eye the AI subs though, clearly there’s a community for it here that’s interested in discussion.
Oh and I forgot to mention, instead of a 5090, buy AMD Strix Halo if its any good.
I cannot emphasize how awesome 128GB on a fast APU would be. That opens up (admittedly slow, but usable) inference of “huge” models like Mistral Large, and very fast inference of large MoE models like 8x22B.
Good tips, thanks!! I’ll definitely check it out.
Don’t,Nvidia is going to price gouge the snot out of it. Honestly, if you want to buy new, just get a 7900 XTX. Screw Nvidia’s pricing on new cards, lol.
Speaking as someone who’s done a lot of merging, the “upscaling” merges are not great. Rope scaling the context is not either. You are better off finding models that were trained at the parameter count and context length you want in the first place, and there is a lot more choice these days.
Oh fuck buying Nvidia new, I was going to see if it depressed 40xx prices or even further for 3090 but I’m not sure it would.
Neat didn’t know that about rope, as you can guess largely due to having fuck all memory to work with. Is AMD viable with LLMs now? Honestly if I can make it work with an AMD GPU I just may because I agree screw Nvidia.
For inference? AMD is more finicky to setup but totally fine once you do. 7900 XTX prices can be very good.
I feel like 3090s have bottomed out, as they are just getting more rare now, and 4090s are so freaking expensive to start with I’m not sure how much they’ll come down.
Another feature you might not be aware of, that people use now, is quantized KV cache. With it, I can run a 19GB 35B model and still fit 131K context into vram, with basically no quality loss.
How are you people running cuda kernels?
rocm
exllama, llama.cpp, vllm/aphrodite, (I think) sglang, they all support it now.