ibm-granite/granite-tsfm — reverse-engineered prompt

Build me a Python project for working with time series foundation models, focused on making forecasting easy to try and demo. I want it to help someone load Granite time series models from Hugging Face, run simple forecasts on sample data, and show clear examples for getting started, transfer learning, fine tuning, and benchmarking.

Include friendly Jupyter notebooks for the main workflows, especially TinyTimeMixer, PatchTSMixer, PatchTST, FlowState, and PatchTST FM examples. Also include reusable utility code so the notebooks are not messy, plus tests for the important pieces.

I’d also like a simple inference service so a user can send time series data and get predictions back, with basic setup instructions. Make the README clear for someone who knows Python but isn’t an expert, including install steps, supported Python versions, how to run notebooks, and where to find the models. Look up the current Hugging Face and transformers docs online if you need to.