Curious about GPU utilization

[spidercatfly]

Excellent work! Macro-o1 really supports long-term self-thinking!

I found a magical phenomenon that when Macro-o1 is inferencing, it will continue to occupy the memory of the last GPU until it is about to OOM, and then all the calculation tensors will be removed, leaving only the initial model. Could you tell me how this is done?

Thanks in advance.