Dual-Flow Transformer activates additional computation only from the end of the text during response generation

This approach decouples the extra computation used in large language model response generation from the prefill stage, reducing total inference cost on hardware and providing finer control over prefilling and continuation costs, especially in MoE models.