The vLLM inference engine for large language models has a resource consumption vulnerability (CWE-400) in versions before 0.28.0. Specifically, request bodies for Chat Completions and Responses can specify media_io_kwargs.video.video_backend as pynvvideocodec, which causes MediaConnector.fetch_video to use the VideoMediaIO backend regardless of the startup configuration. The engine's memory budgeting logic does not account for decoder memory allocated dynamically by this backend, allowing creation of CUDA contexts and decoder surfaces that exhaust shared GPU memory. This leads to request failures, worker crashes, or denial of service. The vulnerability is addressed in vLLM version 0.28.0.

An attacker with the ability to submit video requests to a GPU deployment running vLLM with PyNvVideoCodec installed can exhaust shared GPU memory. This results in denial of service conditions such as request failures and worker crashes. There is no impact on confidentiality or integrity reported, only availability is affected.

Upgrade vLLM to version 0.28.0 or later, which contains the fix for this uncontrolled resource consumption vulnerability. No other mitigation is indicated or required.