vLLM 0.28.0 · Runner V2, tiered KV offload, bitsandbytes out-of-tree

Major engine cut (26 Aug): Runner V2, tiered KV with disk offload, Rust/gRPC multimodal path, new models (Muse Glimmer, Ling 3.0 Flash, Dots3 NOTE).¹ Batch token default doubles to 16384. bitsandbytes is a plugin; pin Transformers ≥5.15.0.¹

Relevant to: anyone running vLLM servers or pinning CUDA/ROCm serving images