KVarN: Native vLLM backend for KV-cache quantization by Huawei

Posted by theanonymousone 5 days ago

14316Original

Comments

Comment by throwa356262 5 days ago

Better performance than TQ and better quality than FP16?

Am I reading this right??

Comment by qeternity 5 days ago

It's not better quality: 59.3% vs 59.4% fp16 on AIME 25

Comment by sheepscreek 5 days ago

0.1% is within margin of error. Depending on the performance boost, it might be worthwhile taking a minuscule quality hit.

Comment by qeternity 3 days ago

I think it very much is worth it!

But the point was that quality didn't magically increase.

Comment by electroglyph 5 days ago

any divergence (even if the benchmark is better) from full precision is error

Comment by 7e 5 days ago

Just pretend that it is the next step update when training. You didn’t train your model to step=inf, I hope?

Comment by thefox96 5 days ago

Faster than Fp16, not better quality i guess

Comment by pbich 5 days ago

[dead]

Comment by v3ss0n 5 days ago

Why this is not a PR for vLLM ?

Comment by esafak 5 days ago

It's the output of a research paper; the authors are not trying to build up vLLM, and they probably have no incentive to do so. You can submit a PR, though! It's easier now while the divergence is low, so don't wait. Since there are six authors, I bet you could get help with the inevitable review chores if you just take the step of creating the PR.

edit: It might not be clear that it is based on vLLM 0.22, which is the current version: https://github.com/huawei-csl/KVarN/commit/d6290e99098d7426d.... All you have to do is create a diff off it; it's fairly straightforward.

Comment by jmalicki 5 days ago

And with the help of AI, pointing at AI at this paper and saying "making a vLLM PR from this paper" tends to work surprisingly well, even if you need to nudge it a little bit along the way.

Comment by woadwarrior01 5 days ago

Last I heard, vLLM was backed by a company that has raised $150m in seed funding. I'm sure they've got the resources to port it.

Comment by electronsoup 5 days ago

Why this is not a PR for llama.cpp

Comment by thefox96 5 days ago

it should be easy to do btw

Comment by lukasc-ch 4 days ago

... and it's on llama.cpp that to this guy! https://www.reddit.com/r/LocalLLaMA/comments/1txlhxu/i_imple...

Comment by lukasc-ch 4 days ago

This is awesome! Let's give them some stars: - https://github.com/huawei-csl/KVarN (original repo, vLLM implementation) - https://github.com/Anbeeld/beellama.cpp (llama.cpp implementation + awesome evals)