Unload idle Ollama models and log GPU usage with HTTP and file writes

Quick Overview

This workflow runs manually or every 10 minutes to check which Ollama models are loaded, log state changes to a file, and optionally unload models that have been idle too long to free GPU memory.

How it works

1. Runs on a manual trigger or on a 10-minute schedule.
2. Loads configuration values (Ollama URL, idle thresholds, dry-run mode, and log file path).
3. Calls Ollama’s /api/ps endpoint to list currently loaded models and their expires_at timestamps.
4. Calculates which models are idle (and which are “pinned” based on a far-future expires_at) and marks each model as keep, pinned-skip, would-unload, or unload while tracking the previous state.
5. Appends a log line to the configured file only when a model’s state changes since the last run.
6. When dry-run is disabled, sends a POST request to Ollama’s /api/generate with keep_alive: 0 for each model marked for unload.

Setup

1. Ensure your n8n instance can reach your Ollama server and set the correct base URL in the workflow settings (for example http://host.docker.internal:11434 when using Docker).
2. Create the log directory and update logPath to a writable location for the n8n runtime.
3. Review and adjust idleMinutes, assumedKeepAliveMinutes (to match your Ollama keep-alive behavior), and whether pinned models can be unloaded (unloadPinned).
4. Run the workflow once with dryRun set to true and verify the log output, then set dryRun to false to enable live unloading.