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

影響: pip:vllm < 0.32.0 → 0.32.0

The Phi-4 multimodal audio encoder (phi4mm_audio.py / phi4mm_utils.py) runs its conformer attention in eager mode by default: the official HF config of microsoft/Phi-4-multimodal-instruct does not set use_pt_scaled_dot_product_attention, so the default False applies (phi4mm_audio.py:166, :876; phi4mm_utils.py:1669). The eager path fully materializes the attention probability matrix of shape [n_chunks, 16, 500, 500] per layer — and then stores it on the module: self.attn = attn (phi4mm_utils.py:…

Read on GitHub Advisory Database

On 10/29/2026, 06:15 AM, GitHub Advisory Database published "vLLM: Phi-4 multimodal audio encoder materializes and retains full attention matrices: a single max-length audio OOM-crashes the vLLM engine" (type: Advisory). Severity is "Medium" (medium). Related technologies: Python. CVE: CVE-2026-115913.