vLLM v0.29.0 makes Model Runner V2 the default for every model and deletes ten model architectures

GitHub·high signalXBlueskyLinkedInCopy link

Released 2026-09-09T08:54Z with 594 commits from 277 contributors (91 new), v0.29.0 completes the MRV2 rollout that started with pooling models, adding CUDA graph memory profiling for KV cache auto-sizing and batch-sharded sampling that cuts per-step logits memory by 1/TP. MRV1 survives only for a few ROCm models. The breaking set is large: ten deprecated architectures removed, FlexOlmo/Olmo3/Hunyuan moved to the Transformers backend, the PyAV video decoder gone, and `python -m vllm.entrypoints.openai.api_server` deprecated in favor of `vllm serve`. FlashInfer all-reduce is now on by default for TP CUDA groups, opt out with VLLM_ALLREDUCE_USE_FLASHINFER=0.