It’s a small model by comparison. If you want something that’s offline and actually closer to comparing to ChatGPT 3.5, you’ll want the Mixtral 8x7B model instead (running on a beefy machine):
I always find it interesting how text is so much slower than image generation. I can do a 1024x1024 in probably 20s, but I get like 1 word a second with text.
I’d love to see some consumer level AI stuff, sadly it all seems to be designed for server farms and by the time it ages out into consumer prices it’s so obsolete there’s no point in getting it.
Nice! Thats a cool project, ill have to give it a try. I love the idea of self hosting local LLMs. Ive been playing around with: https://lmstudio.ai/ and it directly downloads from hugging face.
I tried llamafile for text gen too but I couldn’t get ROCm to properly work with it to run it through my GPU without having to build it myself, which I’m really not into. And CPU text gen is waaaaaay too slow for anything. Mixtral response was like ~250 seconds or so for ~1k context tokens, I think Mistral was about 52 seconds or something around that number.
https://github.com/Mozilla-Ocho/llamafile
Mixtral is definitely beefy, Mistral is quite a bit faster and there’s a few even smaller prebuilt ones. But the smaller you go the less complex the responses will be.
I think llamafile is a good step in the right direction though, but it’s still not a good out of the box experience yet. At least I got farther with it than with oobabooga, which is the recommendation for SillyTavern, which would just crash whenever it generated anything without even giving me an error.
Have you missed the first part where I explained that I couldn’t get it to run through my GPU? I would only have a 6650 XT anyway but even that would be significantly faster than my CPU. How far I can’t say exactly without experiencing it though, but I suspect with longer chats and consequently larger context sizes it would still be too slow to be really usable. Unless you’re okay waiting for ages for a response.
Direct link to the GitHub repo:
https://github.com/nickbild/local_llm_assistant?tab=readme-ov-file
It’s a small model by comparison. If you want something that’s offline and actually closer to comparing to ChatGPT 3.5, you’ll want the Mixtral 8x7B model instead (running on a beefy machine):
https://mistral.ai/news/mixtral-of-experts/
Sick, I only need 90gb of VRAM!
I’ve got it running with a 3090 and 32GB of RAM.
There are some models that let you run with hybrid system RAM and VRAM (it will just be slower than running it exclusively with VRAM).
Yeah but damn does it get slow.
I always find it interesting how text is so much slower than image generation. I can do a 1024x1024 in probably 20s, but I get like 1 word a second with text.
Languages are complex and, more importantly, much less forgiving to error
Hopefully we see more specific hardware for this. Like extension cards with pretty much just tensor cores and their own ram.
Graphic cards without video connection exists since a while.
I’d love to see some consumer level AI stuff, sadly it all seems to be designed for server farms and by the time it ages out into consumer prices it’s so obsolete there’s no point in getting it.
It’s not quite consumer level I’d say but Coral.ai has some small Google Edge based TPUs.
Do they want consumer ai cards to exist though?
Think about the data!
Card makers? They only want money, if theres enough consumer level demand they will make them.
I guess your right.
Nice! Thats a cool project, ill have to give it a try. I love the idea of self hosting local LLMs. Ive been playing around with: https://lmstudio.ai/ and it directly downloads from hugging face.
There’s also ollama which seems to be similar. Not sure if LMStudio is open source but ollama is.
I tried llamafile for text gen too but I couldn’t get ROCm to properly work with it to run it through my GPU without having to build it myself, which I’m really not into. And CPU text gen is waaaaaay too slow for anything. Mixtral response was like ~250 seconds or so for ~1k context tokens, I think Mistral was about 52 seconds or something around that number.
https://github.com/Mozilla-Ocho/llamafile Mixtral is definitely beefy, Mistral is quite a bit faster and there’s a few even smaller prebuilt ones. But the smaller you go the less complex the responses will be. I think llamafile is a good step in the right direction though, but it’s still not a good out of the box experience yet. At least I got farther with it than with oobabooga, which is the recommendation for SillyTavern, which would just crash whenever it generated anything without even giving me an error.
How fast are they with a good GPU?
Have you missed the first part where I explained that I couldn’t get it to run through my GPU? I would only have a 6650 XT anyway but even that would be significantly faster than my CPU. How far I can’t say exactly without experiencing it though, but I suspect with longer chats and consequently larger context sizes it would still be too slow to be really usable. Unless you’re okay waiting for ages for a response.
Sorry, I’m just curious in general how fast these local LLMs are. Maybe someone else can give some rough info.