Live data from Hacker News

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

github.com

11–20 of 21 posts

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

#15

Better performance than TQ and better quality than FP16? Am I reading this right??

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

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

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

#16

Better performance than TQ and better quality than FP16? Am I reading this right??

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

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

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

#20

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

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)
Post reply on HN