Allocation of Resources Without Limits or Throttling Affecting vllm package, versions [,0.28.0)

vllm is an A high-throughput and memory-efficient inference and serving engine for LLMs

Affected versions of this package are vulnerable to Allocation of Resources Without Limits or Throttling via the VideoMediaIO.merge_kwargs and DeepStream video decoding paths in vllm/multimodal/media/video.py and vllm/multimodal/video.py. An attacker can cause partial denial of service by sending a video request that selects backend="deepstream" at request time and forces the service to initialize and use the process-wide DeepStream decode pool. Because DeepStream was not treated as a GPU backend and its decode path did not enforce the video pixel limit, unauthenticated requests could bypass the startup-time backend controls and submit oversized video workloads that other backends reject. This degrades availability for concurrent requests, causing timeouts and reducing service capacity.

Workarounds

• Disable request-time backend="deepstream" selection for video inputs, and allow DeepStream only when it is statically configured at startup; this prevents unauthenticated clients from activating the process-wide DeepStream decode pool on demand.
• Enforce VLLM_MAX_IMAGE_PIXELS checks for DeepStream video decoding, or reject oversized video dimensions before decoding; this prevents oversized video requests from bypassing the pixel-limit resource control.

• GitHub Commit
• Maintainer's Advisory