Hacker News Reader

Comment by scrlk | Hacker News Reader

IME, unquantised -> FP8 is pretty much lossless. What matters more is having an unquantized KV cache - using an FP8 KV cache can result in a significant drop in quality.

View on HN

>unquantised -> FP8 is pretty much lossless

Claude Shannon is rolling in his grave.

I don't know, sounds quite similar to his rate distortion theorem (analyzing minimum number of bits/symbol you need to stay under some fixed amount of distortion). I.e. lossy compression with a maximum amount of loss. I.e. "pretty much lossless" compression.

https://en.wikipedia.org/wiki/Rate%E2%80%93distortion_theory

"Pretty much" doing a lot of work. But it's kinda analogous to 99% JPEG compression: yes you can detect loss, but you get meaningful compression ratios out of it and the subjective appearance is nigh-on perfect.

Shannon would be pointing out that if you can throw away half the model without apparent degradation, we're nowhere near packing in all the information we could in training. There must be a better arrangement than we've currently got.

Do infra providers reveal that level of implementation detail?

I've seen a few articles from providers talking about KV cache quantisation, but it's not something they explicitly point out like they do with weights.

So you could end up paying more for unquantised weights, only to get silently hit with a quantised KV cache...

The official API is FP8, which should imply that it's lossless.

Reply on news.ycombinator.com