vLLM 0.30.0 deletes four things you may be running and shortens two model context windows without asking.
vLLM shipped v0.30.0 at 05:20 UTC, 762 commits from 315 contributors, and it is a stable release rather than a candidate.1 The upgrade notes carry removals, and two of them change behaviour silently rather than erroring.
The quiet one is YaRN. vLLM aligned its aliases with Transformers and stopped multiplying max_position_embeddings by the scaling factor twice, so derived max_model_len drops for some models: the notes name TeleChat3-36B-Thinking going from 131072 to 32768, and sarvam-105b from 5242880 to 131072.1 A serving stack that relied on the old arithmetic will start refusing long requests it accepted yesterday, with no deprecation warning in between.
Read the upgrade notes before this reaches a pinned deployment. The removals are recoverable; the YaRN context-length change can surface as user-visible request failures instead of a startup error.
The full removal list, and what 0.30.0 gives you in exchange
GPTQ group and dynamic activation ordering is gone: g_idx is now ignored, and the Marlin, GPTQ, CPU and RDNA3 kernels behind it were deleted. The scale-out endpoints /render, /derender and /inference/v1/generate no longer register on a plain vllm serve without --enable-scale-out, and the VLLM_ENABLE_SCALE_OUT_ENDPOINTS variable that used to turn them on was removed. Four items deprecated for 0.29 are now deleted outright, including the VLLM_PREFIX_CACHE_RETENTION_INTERVAL and VLLM_MM_HASHER_ALGORITHM environment variables and the ROCm CUDA_VISIBLE_DEVICES fallback.1
Fast Start adds a persistent per-GPU weight-cache daemon that holds post-quantized, TP-sharded weights in GPU memory, so a restarting engine maps them over CUDA IPC with --load-format ipc_cache instead of reloading from disk. It now covers FP4 checkpoints and multi-node tensor parallelism.
Freezing the garbage collector during CUDA-graph capture cuts capture from 12s to 2s and engine init from 28.9s to 8.2s on an H200, per the release notes. Both figures are vLLM's own, measured on their hardware, not ours.
Also new: Gumbel-max watermarked generation with a keyed PRF and per-request opt-out, and HiSparse, a host-resident KV tier that spills sparse-MLA decode pages to pinned host memory under GPU pressure.
Sources
- vLLM v0.30.0 release notes · read