vLLM: Phi-4 multimodal audio encoder materializes and retains full attention matrices: a single max-length audio OOM-crashes the vLLM engineGHSA-8397-8qf6-f2gm

Summary

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:1802, a torch.jit.Attribute declared at :1693), keeping a live reference for every one of the 24 encoder layers.

The only gates on audio length are the server-side limits VLLM_MAX_AUDIO_CLIP_FILESIZE_MB=25 and VLLM_MAX_AUDIO_DECODE_DURATION_S=600 (envs.py:83-84) — i.e. a 10-minute clip is a legal input by default. A 600 s clip becomes ~60,000 mel frames, which the encoder unfolds into 500-frame chunks (phi4mm_audio.py:996, max_seq_len = 500), giving n_chunks ≈ 120. Each layer therefore materializes a 120×16×500×500 matrix (~1–2 GiB depending on dtype), and the module attribute keeps all completed layers' matrices alive, so GPU memory climbs monotonically within a single request until an allocation fails.

Verified end-to-end on the official vllm/vllm-openai:v0.31.0 image: one 19.2 MB silent 600 s WAV → worker torch.OutOfMemoryError: CUDA out of memory. Tried to allocate 2.29 GiB ... 501.75 MiB free ... 22.09 GiB allocated by PyTorch inside get_audio_features (phi4mm_audio.py:1279 → :1332 → phi4mm_utils.py:1386) → EngineCore encountered a fatal error → API EngineDeadError → whole server down; /health refuses connections. Total time from request to dead service: ~6 seconds.

A secondary effect: because self.attn persists on the module after the request ends, the matrices of the last audio request stay resident indefinitely (observed +50 MiB pinned after a single 5 s control request); they are only overwritten by the next audio request, so the resident footprint equals the largest audio processed since startup.

Details

1. Root cause. MultiHeadedAttention.forward_attention in eager mode computes attn = masked_softmax(scores, mask) and unconditionally executes self.attn = attn (phi4mm_utils.py:1795-1802). The attribute is a leftover inference/debug hook (used by torch.jit.Attribute(None, Tensor | None) at :1693); in serving it serves no purpose but pins the full [batch·chunks, heads, T, T] tensor per layer past both the layer's forward and the request's lifetime. The SDP attention branch (use_pt_scaled_dot_product_attention=True) never materializes this matrix — but the HF config does not enable it and the code default is False.

2. Why one request suffices. The conformer chunks long audio to max_seq_len = 500 frames (phi4mm_audio.py:996). Duration is bounded only by VLLM_MAX_AUDIO_DECODE_DURATION_S=600, so n_chunks reaches ~120 for a maximal clip. With 16 heads and fp32 scores, one layer's attention tensor is ~1.8 GiB (the observed failing allocation was 2.29 GiB, consistent with the scores+bias materialization). The encoder has num_blocks=24, attention_heads=16, attention_dim=1024, time_reduction=8 (HF config). Weights plus a handful of retained per-layer matrices exhaust the ~2.3 GiB headroom that --gpu-memory-utilization 0.9 leaves on a 24 GB card — mid-forward, on the first request.

4. Deployment reality. --trust-remote-code is the documented, standard way to serve this model (the HF model card instructs it), so the finding applies to default deployments. The crash is deterministic, not load-dependent: scheduler stats at death show a single running request and near-zero KV cache usage.

Impact

Any unauthenticated client that can reach the chat/completions API of a default deployment serving Phi-4-multimodal-instruct (or any model reusing this audio tower) can kill the entire engine with one request carrying a perfectly legal 10-minute audio clip (silent noise suffices — no crafted content needed). Full denial of service, requiring a process restart. On GPUs with more headroom (e.g. 80 GB at default utilization) a single request may survive, but the eager path still materializes GiB-scale per-layer matrices and retains the last request's tensors indefinitely, so a few sequential max-length requests — or normal traffic following one — achieve the same result.

Relation to prior art: the only public Phi-4-mm advisory we found, GHSA-vc6m-hm49-g9qg, concerns a CPU-side quadratic-cost path — different subsystem, different mechanism. GitHub search for "phi4mm audio OOM/memory" (31 results, 2026-10-08) shows no report of this attention-retention issue; main (2026-10-08) is unchanged.

Suggested remediation (any one breaks the chain; 1+2 recommended):

1. Remove self.attn = attn from the serving path (or gate it behind an explicit debug/output_attentions flag) so attention matrices are freed when the layer returns.
2. Default use_pt_scaled_dot_product_attention to True (flash/SDPA kernels never materialize the [B, H, T, T] matrix); keep eager only as an opt-in.
3. Additionally bound the encoder input (frames or duration) for this model so the eager path cannot reach GiB-scale single-layer allocations even if enabled.

This advisory was fixed in MR60614.