vLLM: Assert-based security check bypassed in optimized mode allows arbitrary code execution

An assert-based security check in vLLM's activation function loading is stripped when Python runs in optimized mode, allowing arbitrary code execution via a malicious HuggingFace model.

Users loading cross-encoder models (e.g., BERT, RoBERTa) with vLLM running under `python -O` or `PYTHONOPTIMIZE=1`.

Update vLLM to the latest commit that replaces the assert with an explicit conditional raise, or avoid running vLLM with Python optimization enabled.