llama-server's router could evict a model mid-load or proxy requests into a dying child
GitHub·medium signal
llama.cpp #29217 (merged 2026-09-22) fixes two races reproduced from issue #28698. A model loaded through the fast path had no queue entry, so tick() evicted it before its own request was proxied. A request for a model being stopped was also proxied into the dying child process. Every load now goes through the scheduler queue, and stopping requests wait for the next instance. The PR carries llama.cpp's 'AI usage disclosure: YES' line.