Disclosed weakness · no CVE recorded

FunAudioChat requests could exceed vLLM’s profiled audio budget

vLLM project · vLLM · FunAudioChat preprocessing

First documented disclosure · Review checkpoint

Summary

Long audio and unnecessarily padded short clips could consume more preprocessing resources than the server had budgeted.

Severity and exploitation evidence

No CVSS score recorded in this review.

Exploitation not established: The upstream sources document the correction and identify no malicious use.

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

Runtime audio length was not bounded to the profiled per-item duration, while short inputs were padded to a full-length feature tensor.

Conditions required

A vLLM deployment serves FunAudioChat and accepts client-supplied audio.

Potential impact

Audio requests could overrun the profiled token budget or allocate excessive memory before generation.

Affected versions and configurations

The public report provides no complete release range. The reviewed 0.33.0 source lacks the runtime length guard.

Fix and mitigation status

0.34.0.

Practical next steps

Upgrade to 0.34.0 or later. Apply request and audio budgets for clients using multimodal inference.

Operating lesson

Enforce the same resource budget during preprocessing that capacity planning assumes.

Evidence and limits

PR 57411 was public September 17 and merged October 28. The November 16 release lists the correction, and its tagged FunAudioChat processor contains the runtime sample-count guard. This is separate from the previously published artifact-revision pinning and generic audio-decoding issues.

Sources