vLLM 0.28.0 Upgrade Guide: Rust Frontend & Quantization
vllm_project · x · 2026-08-27
vLLM v0.28.0 is released with key updates in serving and frontend:
- E/P/D: Removed duplicate image preprocessing and moved it to the GPU.
- Rust Frontend: Added a standalone renderer and image inference over gRPC.
- Quantization: Online MXFP4 quantization, plus batch-invariant NVFP4 MoE via CUTLASS.
- Priority: Request priority can now come from an HTTP header.
- Security: Fixed a DoS vulnerability via audio sample-rate forgery.
Before you upgrade:
- Default Changes: maxnumbatchedtokens increased to 16384; prefix caching enabled by default for Mamba models.
- Plugins: bitsandbytes is now an out-of-tree plugin; calculatekvscales and overrideattentiondtype are removed.
- Dependencies: Upgraded to Transformers 5.15.0; runtime image moved to Ubuntu 24.04.
- Metrics Renamed: KV offload tiering metrics renamed from ...block... to ...chunk....
Related event: vLLM v0.28.0 Ships with 584 Commits from 270 Contributors(3 posts)→
More from Infra
- Actual Computer serves 8,000+ tokens/s from a single NVIDIA DGX Station — cephaloform · 2026-10-16
- Meta open-sources Muse Spark Open: 816B-parameter model that runs locally on two Mac Studios — gabriberton · 2026-10-16
- Jensen Huang joins BG2 podcast: does AI revenue justify capex, '27 bottlenecks — elonmusk · 2026-10-16
- Mistral's on-device infra rack revealed: DGX Spark, Mac Studio and more — gregmushen · 2026-10-16
- NVIDIA and Sweden's Wallenberg launch Sferical One, a shared AI supercomputer — nvidia · 2026-10-16
- Next infra frontier: self-evolving systems that adapt and self-verify — _shreya_s · 2026-10-16