llama.cpp Release b11139 Fixes Server Crash During Sleep State
llama.cpp released version b11139, which resolves a crash in the server token counting API when the server is in a sleep state. The patch ensures that sleeping server instances wake up correctly upon receiving token counting requests. This bug fix enhances the reliability of llama.cpp HTTP server deployments, especially for applications querying token counts while idle. It ensures uninterrupted service uptime when integrating local LLM backends into external developer workflows. The fix addresses PR #29309 by properly handling server wakeups during token counting and cleaning up local aliases. Pre-compiled binaries were published across numerous targets including macOS, Linux, Windows, and Android, supporting CPU, CUDA, Vulkan, ROCm, OpenVINO, SYCL, and Snapdragon backends.
## BACKGROUND
llama.cpp is a popular open-source C/C++ framework for running LLM inference efficiently on local hardware ranging from smartphones to multi-GPU servers. Its included HTTP server allows developers to host OpenAI-compatible endpoints locally, incorporating power-saving mechanisms such as putting server threads to sleep when idle.