vLLM v0.31.0 adds DeepSeek V4.1 Flash support and faster restarts

Context

vLLM is a widely used open source engine for serving language models. The project says version 0.31.0 includes 717 commits from 307 contributors, 96 of them contributing for the first time. DeepSeek V4.1 Flash is also the model CrowdStrike says the attacker in the South Korean bank breaches used as the main backend.

Deep dive

The vLLM team grouped the release into six areas in its announcement.

For DeepSeek V4.1 Flash, it adds FlashMLA mega attention with NVFP4 KV caching, DeepGEMM sparse MQA, Mega Gate and shared Engram host tables.

A new fast restart feature, vllm preload, keeps model weights in GPU memory across restarts, with support for data parallelism and MTP drafts, plus experimental engine snapshots. Restarting a large model server usually means reloading weights, so this targets downtime.

Model Runner V2 gains speculative decoding with a draft model, LiLiCorr, async DFlash and DSpark adaptive verification for Gemma4. For large scale serving, the release lists MoonEP balanced all to all communication, prefill context parallelism combined with data parallelism, DeepEPv2 and weight transfer for reinforcement learning.

Scheduling changes include independent limits on active sequences, adaptive chunking of long prefills and priority for requests already holding KV cache blocks. HiSparse gets fixes for MTP verification, chunked prefill livelocks and prefix caching across host and GPU. The release graphic says the stack covers NVIDIA, AMD ROCm, Intel XPU and CPU.

What to expect next

The full release notes on GitHub list every change, and teams upgrading should check them because the release touches scheduling, caching and speculative decoding at once.

Watch for user benchmarks of DeepSeek V4.1 Flash on vLLM against other serving engines, and for reports on how much vllm preload cuts restart downtime in production. Patch releases may follow as users find issues in the new features.