Primary sourcevLLM
View sourceEvent ·Multi-source
vLLM v0.33.0 发布:新增 4 个模型支持与 GB300 解码优化
vLLM v0.33.0 共包含 927 次提交,由 389 位贡献者完成,其中 122 位是首次参与。新版本加入 Reflection Beam-1.0、Step-5 Preview、Nemotron Parse 3.5、MiniCPM-OCR 1.0 四个模型的支持。推理引擎升级到 PyTorch 2.14,提供 CUDA 13.0 专用 wheel 包。投机解码方面新增 XPress refiner 与 FLy lossy verify(1.04–1.31 倍加速),DSpark 在 GB300 上从 1373 提升到 3390 tok/s。NVIDIA 解码路径实现 1.5–2 倍延迟降低,TTFT 下降 39–48%,同时 ROCm 10.0 成为默认配置。
Current read
vLLM 又发新版了,这版支持 Step-5 Preview 等 4 个新模型,GB300 上解码快了最多 2 倍,自建推理服务的可以升级看看。
11 sources71° AI heatLast updated 2026/11/3 23:53:16
Evidence
11 sourcesRelated source 1PyTorch
View sourceRelated source 2marktechpost
View sourceRelated source 3Artificial Analysis
View sourceRelated source 4NVIDIA AI
View sourceRelated source 5LlamaIndex
View sourceRelated source 6IT之家
View sourceRelated source 7Fireworks AI
View sourceRelated source 8arXiv cs.AI
View sourceRelated source 9Tech in Asia
View sourceRelated source 10Gorden Sun
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.