Create a project venv, then install requirements.txt.
.venv/Scripts/python.exe -m unittest discover -s tests -v(Windows).venv/bin/python -m unittest discover -s tests -v(Linux)node --check manager/web/app.jsnode --check manager/web/monitor.jsnode --check manager/preset_bridge.cjsnpm ci && npm test: monitor arithmetic and browser interactions.
CI runs on Windows and Linux. Windows release builds bundle portable Python, Git, and Node, freeze the manager, and compile the Inno Setup executable. Native two-machine LAN discovery and GPU inference require physical hosts.
Opt-in real integration checks (Windows, with downloads):
python scripts/prepare_runtimes.py.venv/Scripts/python scripts/smoke_upstream.py: two unmodified upstream copies, independent runtime files, clean Git status, CPU backend installation, a tiny test model, preset compilation, inference counters, and independent restart.- After building:
.venv/Scripts/python scripts/smoke_packaged.py: two packaged managers, auto-start, UDP discovery, direct federation, and remote lifecycle.
Integration artifacts stay under .test-data; test processes stop on exit.
These scripts do not configure Windows login startup or firewall rules.
.venv/Scripts/python scripts/smoke_installer.py: silent clean setup, installed executable startup, upgrade, and uninstall in an isolated workspace directory. Refuses to overwrite an existing user installation. Runs in release CI.
- 17 backend tests, 3 monitor arithmetic tests, and 4 Playwright browser tests passed.
- Windows and Linux CI passed.
- Two independent upstream checkouts installed successfully, remained clean, and
retained
https://github.com/thomas9120/LLama-GUI.gitas their origin. - Two CPU llama-server processes ran
stories260K.gguf, launched through saved presets using the original flag compiler. Inference, metrics, slots, and independent restart passed with upstream CPU build b12282. - Two packaged manager processes passed automatic preset startup, local UDP discovery, direct federation, remote stop/start/restart, and preset identity recovery after restarting the manager while inference continued.
- The Windows setup installed, upgraded, started again, and uninstalled while preserving test settings and model files.
The LAN tests used two processes on one Windows host; separate physical hosts and other GPU/driver combinations have not been exercised in this release run.