What is better: higher quantiation or higher parameter count?

For example, does a 13B parameter model at 2_K quantiation perform worse than a 7B parameter model at 8bit or 16bit?

19 points · 17 comments · view on lemmy.world

17 Comments

rufus@discuss.tchncs.de · 9 pts · 3y (2 replies)

https://github.com/ggerganov/llama.cpp#quantization

https://github.com/ggerganov/llama.cpp/pull/1684

Regarding your question: 13B 2_K seems to be on par with 7B 16bit and 8bit. Not much of a difference between all those. (Look at the perplexity values. Lower is better.) The second link has a nice graph.

Most people don't go as low as 2bit though. Look at the graph, below 4bit things start to deteriorate.

Wander@yiffit.net · 5 pts · 3y

That graph is great. Very easy to understand. Thank you!

noneabove1182@sh.itjust.works · 2 pts · 3y

These are good sources, to add one more, the GPTQ paper talks a lot about perplexity at several quantization and model sizes:

https://arxiv.org/abs/2210.17323

noneabove1182@sh.itjust.works · 2 pts · 3y (3 replies)

Anyone else see 11 comments on the post count but only 2 comments..?

alexthelion335@sh.itjust.works · 2 pts · 3y (2 replies)

Yes

rufus@discuss.tchncs.de · 5 pts · 3y (1 reply)

Well, a few of those extra numbers are my fault. I edited my answer a few times. And lemmy reportedly counts every edit as an additional comment. (When user and community are on different instances.) I hope they fix that soon.

noneabove1182@sh.itjust.works · 2 pts · 3y

ahh makes sense, i just made a post and deleted the comment i made on it but it glitched and deleted twice so now my post has -1 comments lmao