ComfyUI-llamacpp-Unloader

ComfyUI-llamacpp-Unloader

Unload llama.cpp models from VRAM to free GPU memory in ComfyUI.

What it does

Queries a running llama.cpp inference server for currently loaded models and unloads them, freeing their VRAM so other ComfyUI workflows can run without OOM errors.

Installation

Via ComfyUI Manager (recommended)

1. Open ComfyUI Manager
2. Search for ComfyUI-llamacpp-Unloader
3. Click Install
4. Restart ComfyUI

Manual installation

cd ComfyUI/custom_nodes
git clone https://github.com/schroldgames/ComfyUI-llamacpp-Unloader.git
cd ComfyUI-llamacpp-Unloader
pip install -r requirements.txt

Usage

Wire the node into your workflow. The llama_api input defaults to http://host.docker.internal:8081, which works for Docker-based llama.cpp deployments on macOS/Windows. Adjust if your server runs at a different address.

Excluding models

The optional Exclude Models text area lets you specify model IDs that should not be unloaded (one per line or comma-separated). Any loaded model whose ID matches an excluded entry will be skipped — all other loaded models are still unloaded.

Gotcha: execution order

ComfyUI does not guarantee execution order between parallel branches. To ensure the unload fires after your generation workflow completes, insert a Reroute node to enforce ordering — wire the last node of your generation chain into the Reroute, then from the Reroute into this node's inputs. This forces ComfyUI to execute the unload only after generation finishes.

Outputs

• success (Boolean) — true if the operation completed successfully (including when no models were loaded), false on error
• result (String) — human-readable status message with per-model results, also shown as inline preview text in the node UI

Failure mode

If the llama API is unreachable (no llama.cpp server at the given address), the node returns success: false with an error message rather than crashing your workflow.