Speculative Decoding

What Is Gemma 4 assistant-MTP: How Multi-Token Prediction Draft Models Speed Up Inference