vLLM: Phi-4 multimodal audio encoder materializes and retains full attention matrices: a single max-length audio OOM-crashes the vLLM engine

[ Description ]Official description

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.

[ Fixes ]Fixes and patches

• vllm 0.32.0Fixed version · GitHub Advisories · pip package vllm: upgrade to 0.32.0 or later (affected < 0.32.0)
• Patch on github.comVendor patch or bulletin · NVD · Reference tagged 'Patch' by NVD
• Patch on github.comVendor patch or bulletin · NVD · Reference tagged 'Patch' by NVD
• Patch on github.comVendor patch or bulletin · NVD · Reference tagged 'Patch' by NVD

[ Advisories ]Government CERT advisories

• BSI CERT-Bund19 Oct 2026WID-SEC-2026-4051

  vllm: Multiple vulnerabilities Translated from German

  CVE-2026-115918CVE-2026-115911CVE-2026-115912+6 more CVEs

[ Packages ]Affected open-source packages

• vllm pipAffected < 0.32.0 · fixed in 0.32.0 · GHSA-8397-8qf6-f2gm

[ References ]References

• github.com/vllm-project/vllm/commit/3d52578e13454de0bb8352cac988347cdcc89928Patch
• github.com/vllm-project/vllm/pull/60614Patch
• github.com/vllm-project/vllm/releases/tag/v0.32.0Release Notes
• github.com/vllm-project/vllm/security/advisories/GHSA-8397-8qf6-f2gmExploitPatchVendor Advisory