According to the Gemma team, the optimal config for inference is temperature = 1.0, top_k = 64, top_p = 0.95, min_p = 0.0 (optional 0.01)