vLLM 0.30.0 ships 762 commits from 315 contributors. A persistent per-GPU weight cache lets restarted engines map already sharded and quantized weights over CUDA IPC instead of loading them from disk. HiSparse adds a pinned-host-memory tier for sparse-MLA KV pages, while Model Runner V2 expands pipeline-parallel speculative decoding and cuts one H200 startup path from 28.9 seconds to 8.2 seconds. The release also closes response-amplification, sparse-embedding and media-input denial-of-service paths.
vLLM 0.33 moves stable builds to CUDA 13 and adds EFA serving
2026-09-22Verified announcementADOPT
vLLM 0.33 moves its stable CUDA artifacts to CUDA 13, adds a dedicated AWS EFA image for multi-node KV transfer, and makes ROCm 10 the default while retaining ROCm 7.2 builds. It expands speculative decoding, low-precision KV paths, watchdog diagnostics and large-scale serving. The release also removes CUDA 12.9 artifacts and several legacy executors, so deployment compatibility must be tested before rollout.
Imported from the existing research archive. Migration preserves the recorded judgment and is not a fresh verification.
What this means for you
For self-hosted inference teams, 0.33 broadens deployable hardware and serving controls but raises the migration floor. Project-reported kernel gains are workload-specific, and the release notes still contain an artifact-verification TODO, so verify images, wheels and real traces rather than treating headline throughput as portable.
Evidence & related material
vLLM 0.31 extends fast restart to DP and MTP and adds experimental initialized-engine snapshots; it also rejects request-level multimodal processor options by default and hardens prefix-cache and LoRA hash boundaries.
vLLM 0.32.0 stable tightens serving-security defaults, expands low-precision inference paths and removes Python 3.10.
vLLM 0.33 stable moves to CUDA 13, adds an AWS EFA image, and expands speculative decoding, low-precision KV and large-scale serving while removing several legacy artifacts and executors.
Subsequent changes
vLLM 0.31 extends fast restart and tightens multimodal request boundaries by default.
vLLM 0.32 tightens security defaults and expands low-precision inference
vLLM 0.33 moves to CUDA 13 and adds EFA serving and new inference paths
Technical details & scores
| Release Scale | 762 commits, 315 contributors, 104 new contributors |
|---|---|
| Fast Start | persistent per-GPU post-quantization and TP-sharded weight cache; CUDA IPC; FP4 and multi-node TP support |
| Hisparse | pinned-host-memory tier for sparse-MLA KV pages with per-request GPU hot buffer and Prometheus counters |
| Model Runner V2 | pipeline-parallel MTP/EAGLE3/DFlash/DSpark; adaptive verification; H200 engine init 28.9s to 8.2s in the reported path |
| Security | bounded validation-error bodies and sparse embeddings; capped request-controlled video sampling; validated cache_salt; credential redaction |
| Breaking Changes | scale-out endpoints require --enable-scale-out; GPTQ g_idx removed; 0.29 deprecations removed; YaRN aligned with Transformers |
| Artifacts | CUDA 13.0/12.9, ROCm, CPU and XPU wheels and images |
Technical impact 4
Engineering value 5
Adoption signal 5
Maturity 4
Verification cost 3
Risk 3