Runtime

Choose a llama.cpp build

The runtime path starts unset. Noullama searches managed installs, nearby builds and PATH. In Runtime → llama.cpp, select a discovered build under Installed builds, or enter a llama-server under Custom path.

Click Check to list devices. For Liquid d1-2B, the runtime must contain llama.cpp PR #31275, such as b11954 or a compatible newer build.

Download an official build

1. Under Download a build, choose a backend.
2. Click Check official releases to preview matching OS/architecture assets from ggml-org/llama.cpp. This fetches metadata only.
3. Choose Download nightly + fallback to approve the displayed downloads: the latest nightly and the binary build linked by the latest official release. The fallback uses CPU, or Metal on macOS.
4. Click Use beside the installed runtime, then check or load it normally.

Archives are checked against GitHub SHA-256 digests and installed in the app-data runtimes/ directory. Existing custom paths are preserved. Linux uses official Ubuntu-built archives; use a custom build if your platform is unsupported.

CPU, GPU and libraries

Set GPU layers under Runtime → Model. Use 0 for CPU inference. Load the model to apply your chosen settings.

If a GPU build cannot find its libraries, set Shared-library folders. On Linux, Noullama also checks common CUDA 12 and 13 locations and passes these paths to the runtime process.

For image input, enable the multimodal projector before loading. See Images & vision.

Read the runtime card

The sidebar runtime card shows whether a model is ready and which llama.cpp build is selected. While a model is running, it shows the build in use and marks a different selection as not applied.

Hover or focus the card for the loaded device, projector, context size, GPU layers, API address and graphics-memory use. Runtime-specific use is available on NVIDIA; total card use is available on NVIDIA and AMD. These values are read only while the card is open.

Closing the window and quitting

With an active system tray, closing the window keeps the model and API running. Choose Open Noullama in the tray menu to restore it.

Quit Noullama in the tray or sidebar asks for confirmation. Confirming stops the API, owned runtime and active downloads, removes partial files from cancelled downloads, and exits. Keep running or Escape cancels.

Linux tray mode needs an AppIndicator library and tray host. GNOME may need an AppIndicator extension. Without tray support, closing the window asks for confirmation before quitting.

On Linux, relaunching the same executable with the same app-data directory shuts down its older instance before starting the API. Startup cleans up orphaned llama-server processes marked as belonging to that Noullama data directory, leaving other applications’ processes alone. Full exit releases Noullama’s own ports.