Faster Gemma 4 on MLX with multi-token prediction · HackerLangs