Accelerating Gemma 4: faster inference with multi-token prediction drafters