A while back I was wondering if I could have J-style notation for handling arrays on GPU. This lets me do that, and the Python/Numpy interface is a wonderful bonus. Awesome implementation! A few exploratory questions:

1. I've written a bit of J for Advent of Code/Project Euler, and I love J's tacit/point-free style. Does KlongPy have hooks and forks? What is the general difference between J and K/Klong/KlongPy?

2. Have you considered adding PyTorch as a backend, in addition to numpy/cupy? Array programming on GPU with autograd would be a magic combination.

3. Would it benefit to rewrite some of the verbs (or add new/optimized verbs) in C? I know numpy has an ok C API, and I've tried the C FFI in J with interesting results.

Again, beautiful work -- starred the repo, bookmarked, I hope to check this out sometime this summer. Thank you so much for sharing! (also thanks to Nils for your work!)