Primary sourcearXiv: OpenAI
View sourceEvent ·Official first-hand
研究分析 5398 个仓库:开发者实际只用到一小部分 LLM 推理配置项
一项大规模实证研究分析了 5398 个使用 Hugging Face Transformers、vLLM 或 OpenAI API 的 GitHub 开源仓库,共识别出 345117 次显式配置的选项。结果显示,虽然 98.8% 的后端文档选项至少出现一次,但 78.9% 的选项只在不超过 10% 的相关仓库中被配置。配置规模很小,Transformers 仓库中位数为 2 个选项,vLLM 和 OpenAI 为 4 个,但选项组合差异极大,56.8%-61.6% 的多选项组合只出现在一个仓库中。不同应用场景之间的选项采用率差异最高达 30.7 个百分点。
Current read
三个后端、五千多个仓库被扒了一遍配置,想知道大家平时真的调哪些推理参数、哪些基本没人碰,看这篇就够了。
11 sources27° AI heatLast updated 2026/10/20 09:52:31
Evidence
11 sourcesRelated source 1IT之家
View sourceRelated source 2Tech in Asia
View sourceRelated source 3Decoder
View sourceRelated source 4TestingCatalog
View sourceRelated source 5The Information
View sourceRelated source 6Teortaxes
View sourceRelated source 7a16z
View sourceRelated source 8Artificial Analysis
View sourceRelated source 9techcrunch
View sourceRelated source 10OpenAI Blog
View sourceConflict check
Dedup data only shows that these sources discuss the same event, not that they agree. There is no structured supporting or refuting evidence yet, so no conflict is inferred automatically.