CVE-2026-115913: vLLM: Phi-4 multimodal audio encoder materializes and retains full attention matrices: a single max-length audio OOM-crashes the vLLM engine
vLLM is an inference and serving engine for large language models. Prior to 0.32.0, the Phi-4 multimodal audio encoder in phi4mmaudio.py and phi4mmutils.py uses eager attention by default, allowing a remote client to submit one legal long audio clip that causes MultiHeadedAttention.forwardattention 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 Software
Remediation
Recommended actions to resolve this vulnerability, in priority order.
- Upgrade
Upgrade
vLLMto a version that resolves this vulnerability.Fixed in 0.32.0
Event History
Frequently Asked Questions
Which deployments are exposed to this denial-of-service condition?
vLLM deployments using the Phi-4 multimodal audio encoder are exposed when running a version earlier than 0.32.0. A remote client can trigger the issue by submitting a legal maximum-length audio clip.
Does exploitation require authentication or user interaction?
The supplied CVSS vector indicates low privileges are required and no user interaction is required. Exploitation is network-reachable and has low attack complexity.
What is the operational impact if the issue is exploited?
Quadratic GPU-memory consumption can cause an out-of-memory condition that terminates EngineCore, disrupting service for all users of that engine.
What should be done if the affected encoder is in use?
Upgrade vLLM to version 0.32.0, which fixes the issue. The data provided does not specify an alternative mitigation for environments that cannot yet upgrade.