agentscope-ai/TuFT — reverse-engineered prompt

Build me a local multi tenant fine tuning service for LLMs that several people can use at once on shared machines, with one simple Tinker compatible API. I want to be able to start a server, connect a client to it, see which base models are available, create a LoRA training session for a chosen model, send tokenized training data through forward and backward steps, run optimizer updates, and get back basic training metrics.

It should also let me save a checkpoint so training can resume later, save sampler weights for inference, and then load those saved weights into a sampling client to generate tokens from a prompt. Please keep user sessions separated cleanly so each tenant can inspect their own runs and artifacts.

Make it easy to install and run locally, ideally with a one command setup plus a Docker option if that already fits the project. Include a small working example that shows the full flow from training data to sampling. Look up current docs online if you need to.