4 ms·

Similar article for vLLM: https://vllm-website-pdzeaspbm-inferact-inc.vercel.app/blog/deepseek-v4 https://vllm-website-pdzeaspbm-inferact-inc.vercel.app/blog/..

by Palmik 6mo ago

Similar article for vLLM: https://vllm-website-pdzeaspbm-inferact-inc.vercel.app/blog/deepseek-v4 https://vllm-website-pdzeaspbm-inferact-inc.vercel.app/blog/... Bechmarks from InferenceX (they do not have apples-to-apples setups to compare the different engines for whatever reason): https://inferencex.semianalysis.com/inference?i_hc=1&g_model=DeepSeek-V4-Pro&g_rundate=2026-04-25&g_runid=24943464864&i_prec=fp4%2Cfp8 https://inferencex.semianalysis.com/inference?i_hc=1&g_model... I find it odd that sglang, vLLM, TRTLLM don't seem to want to publish benchmarks comparing each other. They used to, but now there seems to be some unspoken rule against it. At least we get comparison against "other OSS engine" this time, but that could be HF's Transformers as well :)

imjonse 6mo ago
They're OSS projects in a friendly competition, both working towards the goal of having alternatives to big closed source players. No need for jabs.
- Palmik 6mo ago
  I don't think "friendly" and "publishing benchmarks" are at odds with each other. Model makers (both open and closed weight) typically publish benchmarks against other models and when they do not, people rightfully call them out. Including comparison against "other OSS engine" is just not helpful (what if it's a sandbagged baseline like HF Transformers?)
  - rfoo 6mo ago
    The problem here is both aimed for Day 0 support, both got embargoed preliminary model weights and arch, and I don't think they have access to the other sides embargoed code.
mirekrusin 6mo ago
Too early, they simply didn't yet have access?
embedding-shape 6mo ago
> I find it odd that sglang, vLLM, TRTLLM don't seem to want to publish benchmarks comparing each other. They used to, but now there seems to be some unspoken rule against it. Someone always cries foul with "you're using the wrong version/patch" or similar, and they get hopelessly outdated so fast too, especially since the labs tend to release when everyone else is releasing too, hassle to keep them up to date, and why would you when sometimes the alternatives manages to get ahead of you? :)

by Palmik 6mo ago

imjonse 6mo ago

They're OSS projects in a friendly competition, both working towards the goal of having alternatives to big closed source players. No need for jabs.

Palmik 6mo ago
I don't think "friendly" and "publishing benchmarks" are at odds with each other. Model makers (both open and closed weight) typically publish benchmarks against other models and when they do not, people rightfully call them out. Including comparison against "other OSS engine" is just not helpful (what if it's a sandbagged baseline like HF Transformers?)
- rfoo 6mo ago
  The problem here is both aimed for Day 0 support, both got embargoed preliminary model weights and arch, and I don't think they have access to the other sides embargoed code.

mirekrusin 6mo ago

Too early, they simply didn't yet have access?

embedding-shape 6mo ago

> I find it odd that sglang, vLLM, TRTLLM don't seem to want to publish benchmarks comparing each other. They used to, but now there seems to be some unspoken rule against it. Someone always cries foul with "you're using the wrong version/patch" or similar, and they get hopelessly outdated so fast too, especially since the labs tend to release when everyone else is releasing too, hassle to keep them up to date, and why would you when sometimes the alternatives manages to get ahead of you? :)