vLLM v0.29.0 upgrade guide: new defaults, ten architectures removed, DoS fix
vllm_project · x · 2026-09-11
Key upgrade notes for vLLM v0.29.0:
- New defaults: MRV2 for every model, FlashInfer all-reduce for TP CUDA groups, prefixcacheretentioninterval 0 on SWA/SSM models
- MRV1 stays for a few ROCm models and not-yet-covered features
- Prefix-cache NONEHASH is now deterministic, so distributed KV cache users can drop PYTHONHASHSEED
- Ten deprecated architectures removed; FlexOlmo, Olmo3 and Hunyuan V1/VL move to the Transformers backend
- PyAV video decoder removed — use OpenCV or Torchcodec
- python -m vllm.entrypoints.openai.apiserver deprecated in favor of vllm serve
Serving/frontend highlights: Mooncake Store can offload decode KV; queue admission control via --max-num-queued-reqs/--max-num-queued-tokens; a /v1/messages/render endpoint for the Anthropic Messages API; Rust frontend adds audio/video over gRPC; bounded cachesalt length closes a scheduler CPU exhaustion DoS.
Related event: vLLM v0.29.0 Cuts Blackwell Latency 33.6%, Model Runner V2 Default(4 posts)→
More from Infra
- REVA Mines LLM Attention into Reusable Evidence Views, Cutting RAG Compression Overhead up to 15.6x — _reachsumit · 2026-09-11
- Edge0-35B-A3B preview MoE model for edge inference trends on Hugging Face — Edge0 · 2026-09-11
- Reflect Orbital wants to sell sunlight via volleyball-court mirrors on satellites — kyliebytes · 2026-09-11
- Data centers are for startups, not frontier labs: more compute is the anti-monopoly move — arthurcolle · 2026-09-11
- DeepSeek cut KV cache per token 54x in 9 months, called the third frontier lab — max_paperclips · 2026-09-11
- Anatomy of Jensen Huang's 'AGI Is Here' Tweet: Zero-Cost Signaling and Abilene's 357-Job Data Center Deal — 创业邦 · 2026-09-11