mlc-ai/pith-train ? reverse-engineered prompt

Build me a compact Python training system for mixture of experts language models that feels easy for an AI coding agent to work with.

I want it to handle pretraining first, with the ability to run on NVIDIA Hopper or Blackwell GPUs, support single node and multi node training, resume from checkpoints automatically, and export and import checkpoints in Hugging Face format. It should include example configs for models like Qwen3 and DeepSeek style models, plus scripts for tokenizing a corpus and launching training. Make it practical for real use, with clear project structure, good docs, tests, and a simple way to change batch size, learning rate, and parallelism settings without digging through a lot of hidden code.

If you need current details for any external libraries or best practices, look them up online.