Entity · DiggingBeagle record
vLLM
Open-source LLM inference engine with scale-out, cache and multimodal input paths.
Record description
vLLM is an open-source model inference and serving framework that supports multimodal decoding, distributed workers and shared-memory feature caches. Distinct advisories document a media-type confusion path that can reach vulnerable Ghostscript through Pillow and a forged cache-handle path that can reach pickle deserialization. These are separate trust boundaries, each with specific prerequisites and an identified repair in vLLM 0.32.0. Neither advisory should be generalized to all vLLM deployments or to identical conditions for remote execution. Decoder case and cache case.