News for hackers. Stuff that matters.
Re:Large language models are having their Stable Diffusion m... by enduser on Saturday March 11, 2023 @10:25PM (#35116216)
I’m getting better results with 30B unquantized (f16) than with 65B at 4 bits, FWIW. On a Mac Studio with 128GB RAM.
Re:Large language models are having their Stable Diffusion m... by wjessup on Saturday March 11, 2023 @10:33PM (#35116276)
post example prompts and results please?
Re:Large language models are having their Stable Diffusion m... by koheripbal on Sunday March 12, 2023 @10:36AM (#35120513)
That's to be expected. 4 bits is too small. But the 8 bit 65b should outperform significantly.

Have you tried it?

Re:Large language models are having their Stable Diffusion m... by sebzim4500 on Sunday March 12, 2023 @01:54PM (#35123568)
Is 4 bits too small? Or is the quantization method just not sophisticated enough.

See https://arxiv.org/pdf/2210.17323.pdf

Re:Large language models are having their Stable Diffusion m... by enduser on Sunday March 12, 2023 @06:47PM (#35127428)
I’ll give it a try this evening.
Re:Large language models are having their Stable Diffusion m... by sebzim4500 on Sunday March 12, 2023 @01:53PM (#35123559)
From what I can tell, the quantization being used there is extremely naive.

If they were using techniques like GPTQ, presumably it would work much better.

https://arxiv.org/pdf/2210.17323.pdf