Hello, I added some comments.

Thank you! They should all have been addressed now.

Although it might be necessary I think it's a little unfortunate that you had to reimagine the implementation a bit. When I added it was pretty much a straight copy/paste from numpy. It made it very easy to review and see through possible errors (although not the cupy errors apparently...). But the numpy implementation has already changed since I copy/pasted it so keeping up with upstream is not easy.

Yes, supporting multiple backends can come with this sort of adapting when some functionality isn't supported. With that said, I'm not really worried about Dask having a diverging implementation to that of NumPy, due to its different requirements and different project paces it's generally impractical to keep everything in sync. Regardless of that, the new code is indeed much simpler than it used to be, thanks for working on that!

Thanks John for merging!