llm.cpp OOB in SamplingParams (GHSA-w27m-xm79-ngpc) fix

github.com 2026-09-19 查看中文 →

Security Advisory GHSA-w27m-xm79-ngpc High vLLM

Affected:

vllm-project/vllm using allowed_token_ids with multimodal models like LLaVA where tokenizer vocab size exceeds model output vocab size

Referenced CVEs: CVE-2026-93840 · 3.7

文章内图片已隐藏以节省流量 · Upgrade to Pro to view images & offline archive

This content was auto-fetched from github.com, cleaned by our LLM pipeline, and translated to English. View original.