deepseek-ai/DualPipe ? reverse-engineered prompt

Build me a small Python project that demonstrates the DualPipe idea from DeepSeek’s training report.

I want two working examples, one for DualPipe and one for the simpler DualPipeV version, so I can run them and see the pipeline schedule in action. It should show how forward and backward passes can overlap, and make it easy to understand where the pipeline bubbles are reduced. Please keep it practical, with clear example scripts, simple defaults, and enough comments that I can follow what is happening without already knowing the algorithm.

Use PyTorch 2.0 or newer, and if you need to check the current docs or best practices while implementing it, go ahead. Also make sure the code is set up so someone could later adapt it for their own model by replacing the example overlapped forward and backward logic. If useful, include a simple visual or printed schedule so I can compare the two approaches quickly.