With hierarchical forecasting multiple timeseries are used together to train (local or global).
However during inference it might be that not at all times inference is requested for all timeseries instances on which was trained.
This part does not seem to be supported.
Now it requires to train a separate model for the subset.

Would it not be nice to add the possibility during inference to allow sending a subset of the hierarchy to get forecasts for.
I imagine cross-learning covers this partially for global forecasting. But the case is not covered for local and global (panel) forecasting.