This release fixes a critical bug where opening the llama.cpp Manager could break the installed llama-server binary, causing “is not executable” errors when starting the server from the main screen.
What’s Fixed:
- Manager Cleanup No Longer Breaks Installations: The installed
llama-serveris now an independent real file in~/.local/bin/, so cleaning the CMake build cache (~/llama.cpp/build) can no longer break it. - Atomic Real-File Installation: Source builds (Option A) now install
llama-serverby copying to a temporary file, validating it (executable + valid ELF), and atomically replacing the target — no symlinks are used. - Legacy Symlink Migration: Existing symlink-based installations are automatically upgraded to real files when the manager opens or before any build cache cleanup; external symlinks are left untouched.
- Broken Symlink Recovery: A leftover broken symlink is replaced with a proper real binary during the next build/install.
- Post-Install Verification: After every install, the binary is verified (exists, not a symlink, executable, valid ELF via
file, andllama-server --helpsucceeds) and the result is logged. - Robust Server Lookup: The main screen re-validates its cached
llama-serverpath before use, so a stale or broken path is never reused; the strict executability check is kept unchanged.