b8952

Llama.cpp, the popular open-source C++ implementation for running LLMs locally, has released version b8952 with significant router enhancements. The headline feature enables the router to forward form-data to model servers, specifically fixing issue #22044 to allow /v1/audio/transcriptions in router mode. This opens up audio transcription capabilities for users running llama.cpp in a distributed or load-balanced setup.

Under the hood, the update brings several technical improvements. It uses a non-throwing json::parse overload for better error handling, extends file representation to include filename and content-type metadata, and makes the RNG thread_local for safer concurrent access. The multipart body builder switched from std::string to std::ostringstream for efficiency, and a sanitize_field lambda was added for secure key, filename, and content-type handling. These changes enhance stability and security when processing form-data requests through the router.