llama.cpp Release b11114 Fixes Server Router Eviction Race Conditions
llama.cpp build b11114 addresses server router race conditions by ensuring all model loading requests pass through the central queue. It also prevents incoming requests from being routed to model instances that are currently stopping or unloading. This release enhances server stability and reliability for multi-model inference deployments, preventing dropped requests and premature model eviction. Developers running llama.cpp in production server environments will experience cleaner request lifecycle management and fewer concurrency errors. Implemented in PR #29217, fast-path model loads are now routed through the main queue to protect loaded models until waiting requests complete. Additionally, requests targeting a stopping model now join the queue to be served by the next model instance, with status flags synchronized under a single lock.
## BACKGROUND
llama.cpp is a widely used open-source C/C++ LLM inference engine that allows high-performance model execution across CPU and GPU hardware. Its included HTTP server feature supports dynamically loading and evicting models from RAM or VRAM based on client request traffic.