vLLM 0.31.0 adds preload CLI and gates multimodal kwargs

The vLLM project released version 0.31.0 on October 5, 2026. It carries 717 commits from 307 contributors and includes a new vllm preload command, draft-model speculative decoding in Model Runner V2 and a security-motivated breaking change. If you serve models with vLLM, read the breaking changes before you upgrade.

What changed

• The new vllm preload CLI launches a weight-cache daemon that keeps post-quantized weights in GPU memory across engine restarts.
• Model Runner V2 gains draft-model speculative decoding and custom logits processors.
• Per-request mm_processor_kwargs and media_io_kwargs now return an error unless the server starts with --trust-request-mm-kwargs. Server-level flags and offline LLM use are unchanged.
• tokenizer_mode="slow" is removed, and --enable-mamba-fine-grained-prefix-cache is renamed to --enable-mamba-shared-prefix-checkpoint.
• The release notes list a fused KimiViT QK RoPE kernel at up to 29x faster and a halved DeepGEMM FP8 workspace.

Why it matters

Restart time is a real cost when you run agents against a self-hosted model: every redeploy means reloading weights. A preload daemon attacks that directly. The multimodal change is the one to watch. Clients that pass processor kwargs per request will start failing, and we think rejecting them by default is the right call for any server exposed beyond your own team.