Disclosed weakness · no CVE recorded

Embedding archives could exhaust vLLM memory before its size limit ran

vLLM project · vLLM · client-supplied prompt and media embeddings

First documented disclosure · Review checkpoint

Summary

An inference server accepting client-provided embeddings could allocate far more memory than its configured decode limit.

Severity and exploitation evidence

No CVSS score recorded in this review.

Exploitation not established: The reporter demonstrated excessive allocation in the loader. The source reports no malicious use or production outage.

Exploitation evidence source

Technical severity is not a measure of observed attacks or of risk to your deployment. Attempts, demonstrations, and successful compromise retain the qualifications in the source.

How the weakness works

Validation ran after the archive loader had allocated backing storage. A small tensor view could keep that storage alive.

Conditions required

Prompt or multimodal embedding input is enabled and a client can supply an embedding archive. The project documents these inputs for trusted clients.

Potential impact

The server could run out of memory or delay other embedding loads.

Affected versions and configurations

Reproduced on the pre-fix main branch. Affected packaged releases have not been established.

Fix and mitigation status

PR 67316 merged into main November 16. A packaged fix was not verified.

Practical next steps

Deploy a build containing PR 67316 and keep the decode limit enabled. Limit embedding input to trusted clients.

Operating lesson

Enforce the memory budget before decoding or allocating archive storage.

Evidence and limits

The public November 12 report describes a component-level reproduction. The maintainer merged the fix November 16 and supplies no CVSS score.

Sources