vllm v0.30.0 released — new models, Fast Start IPC cache, HiSparse
vllm v0.30.0 (762 commits from 315 contributors (104 new)) adds DeepSeek-V4.1-Flash, GLM-5.3-Flash, Fast Start per-GPU IPC weight-cache, Gumbel-max watermarking, HiSparse, Model Runner V2, plus new wheels and docker images.
What changed
Adds DeepSeek-V4.1-Flash and GLM-5.3-Flash; Fast Start per-GPU IPC weight-cache with --load-format ipc_cache; Gumbel-max watermarking; HiSparse host KV tier; Model Runner V2 optimizations and new wheels/docker images.
Why it matters
Developers can map post-quantized, TP-sharded weights over CUDA IPC with --load-format ipc_cache instead of reloading from disk; use new models and wheels for immediate evaluation.
Who should care
DevelopersResearchersEnterprises
Worth trying
Yes — Install via pip ('pip install vllm') or docker ('docker pull vllm/vllm-openai:v0.30.0').