CVE-2026-73557

Name of the Vulnerable Software and Affected Versions vLLM versions 0.20.2rc0 through 0.25.x

Description A race condition exists in the safe load prompt embeds function within vllm/renderers/embed utils.py when enable prompt embeds is enabled. The issue occurs because torch.sparse.check sparse tensor invariants uses a process-global state for saving, enabling, and restoring that can be interfered with by concurrent prompt embeds parts. These parts are submitted to the 'POST /v1/chat/completions' endpoint via AsyncMultiModalItemTracker.resolve items, asyncio.gather, and the default executor. This allows an invalid sparse tensor to reach the tensor.to dense operation.

Recommendations Update to version 0.26.0.