deepseek-ai/LPLB ? reverse-engineered prompt

Build me a Python project for balancing Mixture of Experts workloads with linear programming.

I want a planner that takes the current expert assignments, a topology map for redundant experts, and a live token count, then returns better physical expert assignments for the batch. It should support the idea of reordering experts first, then using replicated experts to smooth out overloads across an expert parallel group. If possible, include a way to initialize from an existing DeepEP buffer, but keep it usable with simple inputs too.

Please make it work as a research prototype with tests, a small example, and a clean install path for CUDA enabled machines. The core should be fast enough to run inside training, and it should expose a simple interface that feels easy to call from PyTorch code. If you need current CUDA or library details, look them up online first.