ThinkFacility Sign in
  1. Home
  2. Security
  3. Python package (vllm)
HighPython package (vllm)

vLLM 0.7.3 to before 0.31.0: one image request can knock your model server offline

Anyone serving a multimodal model with vLLM's API server, especially one other people or the internet can reach.

How to fix it

You haveUpdate to
0.7.3 and later, before 0.31.0→0.31.0

Check your version

Which version you have
pip show vllm

Update

pip
pip install --upgrade "vllm>=0.31.0"
uv
uv add "vllm>=0.31.0"

Any API client could send its own mm_processor_kwargs, setting image size or frame count for models like Isaac, InternVL and H2OVL, and run the server out of memory for everyone. vLLM's API key is optional, so a server with no login in front of it is open to anyone who reaches it. The fix refuses those per-request settings by default.

Sources