vLLM v0.30.0 Release — Expanded Model Support and Model Runner V2 Enhancements

New Features and Improvements

• Expanded Model Support: Added support for DeepSeek-V4.1-Flash, DeepSeek-V4-Flash-Vision-Exp, GLM-5.3-Flash, K2-Horizon, Cohere Compass, Bailing V3 VL, and Nanbeige4.2 via the Transformers backend.
• Model Runner V2 Enhancements: Introduced dual-batch overlap in eager mode, full CUDA graphs for microbatched steps, and adaptive verification for speculative decoding.
• Fast Start: Added a persistent per-GPU weight-cache daemon using CUDA IPC to speed up engine restarts, now supporting FP4 checkpoints and multi-node TP.
• Watermarking: Implemented Gumbel-max watermarked generation and detection with a keyed PRF and per-request opt-out.
• HiSparse: Introduced a host-resident tier for sparse-MLA decode to manage GPU memory pressure by spilling KV pages to pinned host memory.
• Quantization: Added support for targeted online quantization and various new formats including NVFP4 and W4A16.
• Performance Optimizations: Includes hardware-specific improvements for NVIDIA (Qwen3.8-Flash-Next, Kimi K3), AMD ROCm, Intel XPU, and CPU (DeepSeek-V4).

Bug Fixes

• Fixed an approximately 2x RL step-time regression related to GPU-compacted sampling masks.
• Resolved various issues including invalid structured-output requests no longer stopping the engine, and handled absolute bounds for validation-error response bodies to prevent response amplification.

Breaking Changes and Deprecations