A script to immediately free up VRAM used by ollama services.

ollama-unload

#!/bin/bash

# The default ollama server address
OLLAMA_URL="http://localhost:11434"

# Help function
show_help() {
  echo "Usage: $0 [OPTIONS] [URL]"
  echo "Unload all loaded models in ollama."
  echo ""
  echo "Options:"
  echo "  -h, --help    Show this help message and exit."
  echo ""
  echo "Arguments:"
  echo "  URL           Optional. The URL of the ollama server. Defaults to 'http://localhost:11434'."
}

# Parse arguments
while [[ $# -gt 0 ]]; do
  case "$1" in
    -h|--help)
      show_help
      exit 0
      ;;
    *)
      OLLAMA_URL=$1
      shift
      ;;
  esac
done

# Identify the models being currently loaded
loaded_models=$(curl -s "${OLLAMA_URL}/api/ps" | jq -r '.models[].name')

# Request unloading for each of the models
for model in $loaded_models; do
  curl -s -X POST "${OLLAMA_URL}/api/generate" -d "{\"model\": \"${model}\", \"keep_alive\": 0}" > /dev/null 2>&1
  echo "Unloading model: ${model}" >&2
done

# reference: https://www.reddit.com/r/ollama/comments/1cnxnrv/how_to_set_keepalive_1_on_ollama_linux/