SGLang versions up to 0.5.21 allocate memory unbounded during video preprocessing in the qwen_vl module. The implementation decodes frames at their full source resolution before any pixel‑count limits are applied, creating a flaw that falls under CWE‑770. When high‑resolution videos are submitted to the Qwen‑VL chat completion endpoint, the server can consume all available worker memory, causing the application to become unresponsive and effectively denying service to legitimate users.

The affected product is the sgl‑lang framework distributed by sgl‑project, specifically versions up through 0.5.21. Hosts running Qwen‑VL deployments that expose the /v1/chat/completions API and use the SGLang 0.5.21 code base are vulnerable.

The CVSS score of 8.7 classifies this flaw as High severity. The EPSS score is not available, but the exploit requires a target that accepts arbitrary‑resolution video uploads and a vulnerability in the preprocessing pipeline; thus the attack surface is relatively narrow. Because the flaw leads to denial of service rather than data compromise, the immediate impact is availability degradation rather than confidentiality or integrity loss. The lack of a KEV listing suggests no known large‑scale exploitation, yet the inherent resource exhaustion risk remains for any exposed Qwen‑VL instance.