Next.js Hacker News
KVarN: Native vLLM backend for KV-cache quantization by Huawei
67 points by theanonymousone 3 hours ago | 7 comments
  • throwa356262
    Better performance than TQ and better quality than FP16?

    Am I reading this right??

    • qeternity
      It's not better quality: 59.3% vs 59.4% fp16 on AIME 25
    • thefox96
      Faster than Fp16, not better quality i guess
    • pbich
      [dead]
  • v3ss0n
    Why this is not a PR for vLLM ?
    • esafak
      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.

      • jmalicki
        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.
    • thefox96
      it should be easy to do btw