Using PyTorch models for training and hyperparameter tuning with Dask

I was speaking with a group today who is using Dask for I/O and preprocessing of a larger-than-memory dataset and then wants to train a single pytorch model on that large dataset. My initial thought was to look at using Dask-ML's `Incremental` meta-estimator, however after reading the docs more closely I realized that `Incremental` only supports models with a `partial_fit` method.

I'm wondering what are some best practices for training a pytorch model (or running a hyperparameter optimization) on a large Dask collection today? Should users use [skorch](https://skorch.readthedocs.io) to wrap their pytorch models for a scikit-learn compatible API? Is there some other approach users should take?

cc'ing @stsievert as you may have thoughts on this topic