Rendered at 10:14:42 GMT+0000 (Coordinated Universal Time) with Cloudflare Workers.
throwa356262 2 days ago [-]
Better performance than TQ and better quality than FP16?
Am I reading this right??
qeternity 2 days ago [-]
It's not better quality: 59.3% vs 59.4% fp16 on AIME 25
sheepscreek 1 days ago [-]
0.1% is within margin of error. Depending on the performance boost, it might be worthwhile taking a minuscule quality hit.
electroglyph 2 days ago [-]
any divergence (even if the benchmark is better) from full precision is error
7e 1 days ago [-]
Just pretend that it is the next step update when training. You didn’t train your model to step=inf, I hope?
thefox96 2 days ago [-]
Faster than Fp16, not better quality i guess
pbich 2 days ago [-]
[dead]
v3ss0n 2 days ago [-]
Why this is not a PR for vLLM ?
woadwarrior01 2 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.
electronsoup 1 days ago [-]
Why this is not a PR for llama.cpp
esafak 2 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.
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.
Am I reading this right??
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.