vLLM: VLLM through 0.29.0 fails to properly clean up decode-side metadata for rejected inference…

Who is affected

vLLM is software that runs AI language models on a server so other programs can send it questions and get answers. This flaw affects vLLM versions through 0.29.0, but only in setups where the work of answering requests is split between two separate groups of workers (one that processes the incoming request and one that produces the answer). When certain requests are rejected, the product does not clean up after them. Someone over a network could send specially shaped requests that keep using up memory on the answer-producing workers until those workers restart. This flaw is rated High.

What to do

1. If your organization runs vLLM itself, the person who installed or runs it should act on this advisory. If an IT provider or vendor runs it for you, send them this advisory and ask when the update will be done.
2. Find out which version you run and whether your setup splits request handling between separate workers as described above. Versions through 0.29.0 in that kind of setup are affected.
3. No fixed version is named for this flaw yet, so check the vendor advisory linked on this page for a fixed release.
4. Install the fixed release as soon as it is available.

Analysis written from the source record as of September 28, 2026 and reviewed before publishing.