vllm-project vllm
| Product | Affected | Fixed in |
|---|---|---|
| vllm PyPI |
|
|
vLLM is an inference and serving engine for large language models (LLMs). In versions from 0.6.4 to before 0.12.0, users can crash the vLLM engine serving multimodal models that use the Idefics3 vision model implementation by sending a specially crafted 1x1 pixel image. This causes a tensor dimension mismatch that results in an unhandled runtime error, leading to complete server termination. This issue has been patched in version 0.12.0.
From the CVE record · GitHub_M
vLLM versions 0.6.4 through versions before 0.12.0 are vulnerable when serving multimodal models that use the Idefics3 vision model implementation. A user can send a specially crafted 1x1 pixel image that triggers an unhandled tensor dimension mismatch and terminates the server. The input does not specify whether authentication is required.
Operators running PyPI vllm versions >=0.6.4 and <0.12.0 should check their deployments.