Never really occurred to me before how huge a 10x savings would be in terms of parameters on consumer hardware.
Like, obviously 10x is a lot, but with the way things are going, it wouldn’t surprise me to see that kind of leap in the next year or two tbh.
Finally. Wrong answers to questions using my phone.
That would actually be insane. Right now, I still need my GPU and about 8-10 gigs of VRAM to run a 7B model tho, so idk how that’s supposed to work on a phone. Still, being able to run a model that’s as good as a 70B model but with the speed and memory usage of a 7B model would be huge.
I only need ~4 GB of RAM/VRAM for a 7B model, my GPU only has 6GB VRAM anyway. 7B models are smaller than you think, or you have a very inefficient setup.
That’s weird, maybe I actually am doing something wrong. Is it because I’m using GGUF models maybe?
llama2 gguf with 2bit quantisation only needs ~5gb vram. 8bits need >9gb. Anything inbetween is possible. There are even 1.5bit and even 1bit options (not gguf AFAIK). Generally fewer bits means worse results though.
Yeah, I usually take the 6bit quants, didn’t know the difference is that big. That’s probably why tho. Unfortunately, almost all Llama3 models are either 8B or 70B, so there isn’t really anything in between but I find Llama3 models to be noticeably better than Llama2 models, otherwise I would have tried bigger models with lower quants.
I’m even more excited for running 8B models at the speed of 1B! Laughably fast ok-quality generations in JSON format would be crazy useful.
Also yeah, that 7B on mobile was not the best example. Again, probably 1B to 3B is the sweetspot for mobile (I’m running Qwen2.5 0.5B on my phone and it works tel real for simple JSON)
EDIT: And imagine the context lengths we would be ablentonrun on our GPUs at home! What a time to be alive.
Being able to run 7B quality models on your phone would be wild. It would also make it possible to run those models on my server (which is just a mini pc), so I could connect it to my Home Assistant voice assistant, which would be really cool.
Something similar to this already kinda exists on HF with the 1.58 bit quantisation which seem to get very similar performance to the original Llama 3 8B model. That’s essentially a two bit quanitsation with reasonable performance!
That’s really interesting, gonna try out how well it runs
Slowly, is how
I have never worked on machine learning, what does the B stand for? Billion? Bytes?
I think it’s how many billion parameters the model has
Thanks!
