vLLM v0.30.0 Ships a Persistent GPU Weight-Cache Daemon That Skips the Disk on Engine Restart

vLLM (GitHub releases)·high signalXBlueskyLinkedInCopy link

vLLM v0.30.0 was published 2026-09-22T05:20Z with 762 commits from 315 contributors (104 new). Fast Start adds a persistent per-GPU daemon holding post-quantized, TP-sharded weights in GPU memory, so restarting engines map them over CUDA IPC with --load-format ipc_cache instead of reloading checkpoints, now covering FP4 and multi-node TP. The release also adds Gumbel-max watermarked generation and keyed-PRF detection (with a dual-key variant compatible with speculative decoding), HiSparse host-resident KV spill for sparse-MLA decode, and DeepSeek-V4.1-Flash support; freezing GC during CUDA graph capture cuts capture from 12s to 2s and engine init from 28.9s to 8.2s on an H200.