CVE-2026-115913
CVE-2026-115913 updated by NVD
Summary
vLLM is an inference and serving engine for large language models. Prior to 0.32.0, the Phi-4 multimodal audio encoder in phi4mm_audio.py and phi4mm_utils.py uses eager attention by default, allowing a remote client to submit one legal long audio clip that causes MultiHeadedAttention.forward_attention to materialize and retain full attention matrices in self.attn. The scaled dot-product attention branch does not materialize these matrices. The resulting quadratic GPU-memory use can cause an out-of-memory condition and terminate EngineCore for all users. This issue is fixed in version 0.32.0.
Affected packages
| Ecosystem | Package | Affected versions | Fixed versions |
|---|---|---|---|
| PyPI | vllm | — | 0.32.0 |
Remediation: Upgrade to 0.32.0 or later.
References
Includes data from the GitHub Advisory Database, licensed under CC-BY 4.0.
Includes data from the National Vulnerability Database (NIST). NVD data is in the public domain; this page is not endorsed by NIST.
CVE® is a registered trademark of The MITRE Corporation. CVE content reproduced under the CVE Terms of Use; copyright designation © MITRE.
EPSS scores provided by the FIRST.org Exploit Prediction Scoring System.