An issue exists where the software forwards a request-controlled `mm processor kwargs.max patches` value to the Hugging Face processor for Cohere2 Vision models without a server-side upper bound. This leads to quadratic CPU work during multimodal preprocessing. Remote clients can exploit this by sending small image chat-completion requests with large, varying `max patches` values to bypass the LRU (Least Recently Used) cache, which is a caching mechanism that discards the least recently used items first, exhaust preprocessing workers, and cause a denial of service for other clients.