inference is not usually slow for classical ML (save lightgbm)