vLLM v0.31.0: 717 commits, `vllm preload` fast restarts, DeepSeek-V4.1-Flash FlashMLA default, security gates and breaking changes

vLLM v0.31.0 (717 commits, 307 contributors) targets restart cost and multi-tenant safety. The new vllm preload daemon keeps post-quantized weights resident in GPU memory across engine restarts (with DP, MTP drafts, a /health endpoint and readiness wait), and experimental vllm snapshot create/restore uses CRIU to restore an initialized TP1 engine. DeepSeek-V4.1-Flash now defaults to FlashMLA mega attention with NVFP4 compressed KV on SM100; Model Runner V2 gains draft-model speculative decoding and custom logits processors; --max-num-active-seqs caps running admission. Security: per-request multimodal kwargs are rejected unless --trust-request-mm-kwargs is set, and LoRA names/paths can no longer collide with cache_salt in prefix-cache keys. No end-to-end throughput headline was published; component numbers include 1.6–4.8x faster GLM-5.3-Flash metadata ops, up to 29x fused KimiViT RoPE, and ~25 s faster Qwen3.8-Flash-Next loading on DGX Spark. Breaking changes: tokenizer_mode="slow" removed, online FP8 via fp8_per_tensor, renamed Mamba prefix-cache flag, --enforce-eager also disables JIT warmup. Install with pip install vllm or vllm/vllm-openai:v0.31.0.