What breaks when upgrading to vLLM v0.30.0?

vLLM v0.30.0 makes scale-out endpoints opt-in on plain `vllm serve` behind `--enable-scale-out`, replacing the old environment variable. GPTQ activation ordering (`g_idx`) is removed, the `all` Mamba cache mode is deprecated, and `python -m vllm.entrypoints.grpc_server` gives way to `vllm serve --grpc`. YaRN now follows Transformers, so vendor aliases stop re-scaling context length.

Which new models does vLLM v0.30.0 serve?

New in vLLM v0.30.0 are DeepSeek-V4.1-Flash, which stores its whole KV cache in MXFP8, plus DeepSeek-V4-Flash-Vision-Exp with ROCm and LoRA support, GLM-5.3-Flash, K2-Horizon with its reasoning and tool parsers, Cohere Compass, Bailing V3 VL, and Nanbeige4.2 through the Transformers backend. A DeepSeek-V4 CPU backend with AVX512 and AMX kernels also lands.

How does watermarking in vLLM v0.30.0 work for users?

Watermarking in vLLM v0.30.0 uses Gumbel-max sampling with a keyed pseudorandom function, and the release ships an example endpoint that detects the mark later. Callers can opt out per request, and a dual-key variant keeps the watermark compatible with speculative decoding. The Rust frontend forwards the per-request controls.

Can vLLM v0.30.0 run without NVLink between GPUs?

Yes — vLLM v0.30.0 adds an opt-in FlashInfer PCIe IPC all-reduce aimed at boxes without NVLink. The same large-scale serving work brings prefill and decode context parallelism to sparse-MLA models, Elastic EP that reuses CUDA graphs across reconfiguration, and encoder-cache sharing over NIXL and Mooncake.

vLLM · 2026-09-22 · major

vLLM v0.30.0 — engine restarts skip the disk with a GPU weight cache

vLLM v0.30.0 adds Fast Start, a per-GPU daemon that holds quantized weights in GPU memory so a restarting engine maps them over CUDA IPC instead of reloading from disk. The release lands 762 commits from 315 contributors.

GitHub card for the vLLM inference and serving engine repository

Fast Start keeps post-quantized weights resident in GPU memory, so a restarting vLLM engine maps them instead of reloading them.

Quick facts

Maker	vLLM project
Version	v0.30.0
Released	September 22, 2026
License	Apache-2.0
Contributors	315 across 762 commits
Headline feature	Fast Start weight-cache daemon
Breaking change	Scale-out endpoints need --enable-scale-out

What is it?

Fast Start is the headline addition in vLLM v0.30.0: a persistent per-GPU daemon that holds post-quantized, tensor-parallel-sharded weights in GPU memory. Engines started with `--load-format ipc_cache` map those weights over CUDA IPC rather than reading a checkpoint off disk. The cache covers FP4 checkpoints and multi-node tensor parallelism. A second new feature, Gumbel-max watermarking, can mark and later detect vLLM-generated text.

How does it work?

Startup time in v0.30.0 also drops because garbage collection is frozen during CUDA graph capture: capture falls from 12s to 2s and engine init from 28.9s to 8.2s on an H200. A new tier called HiSparse spills KV pages to pinned host memory when a sparse-MLA model runs out of GPU room, then serves top-k misses from a per-request GPU hot buffer. Kimi K3 decoding gains 5.2–7.7% end-to-end throughput after a gather/scatter step was removed from its KDA path.

Why does it matter?

Restart cost is what makes autoscaling a serving fleet expensive, and mapping cached weights removes most of it — useful for anyone who redeploys often or runs reinforcement-learning loops that tear engines down between steps. The upgrade is not free, though: scale-out endpoints are now opt-in, GPTQ activation ordering is gone, and YaRN now matches Transformers, so vendor YaRN aliases no longer re-scale `max_model_len`.

Who is it for?

inference and platform engineers

Frequently asked questions

What breaks when upgrading to vLLM v0.30.0?: vLLM v0.30.0 makes scale-out endpoints opt-in on plain `vllm serve` behind `--enable-scale-out`, replacing the old environment variable. GPTQ activation ordering (`g_idx`) is removed, the `all` Mamba cache mode is deprecated, and `python -m vllm.entrypoints.grpc_server` gives way to `vllm serve --grpc`. YaRN now follows Transformers, so vendor aliases stop re-scaling context length.
Which new models does vLLM v0.30.0 serve?: New in vLLM v0.30.0 are DeepSeek-V4.1-Flash, which stores its whole KV cache in MXFP8, plus DeepSeek-V4-Flash-Vision-Exp with ROCm and LoRA support, GLM-5.3-Flash, K2-Horizon with its reasoning and tool parsers, Cohere Compass, Bailing V3 VL, and Nanbeige4.2 through the Transformers backend. A DeepSeek-V4 CPU backend with AVX512 and AMX kernels also lands.
How does watermarking in vLLM v0.30.0 work for users?: Watermarking in vLLM v0.30.0 uses Gumbel-max sampling with a keyed pseudorandom function, and the release ships an example endpoint that detects the mark later. Callers can opt out per request, and a dual-key variant keeps the watermark compatible with speculative decoding. The Rust frontend forwards the per-request controls.
Can vLLM v0.30.0 run without NVLink between GPUs?: Yes — vLLM v0.30.0 adds an opt-in FlashInfer PCIe IPC all-reduce aimed at boxes without NVLink. The same large-scale serving work brings prefill and decode context parallelism to sparse-MLA models, Elastic EP that reuses CUDA graphs across reconfiguration, and encoder-cache sharing over NIXL and Mooncake.

Try it

pip install -U vllm && vllm serve <model> --load-format ipc_cache