Consider upgrading pybind11 to nanobind

[steven-johnson]

Per https://github.com/wjakob/nanobind:

TLDR: nanobind bindings compile ~2-3× faster, producing ~3× smaller binaries, with up to ~8× lower overheads on runtime performance (when comparing to pybind11 with -Os size optimizations).

[steven-johnson]

Update: reading further, there are at least two issues with the current (Mar 2022) version of nanobind:

• Eigen and NumPy integration have been removed.
• Buffer protocol functionality was removed.

We rely (heavily) on numpy and buffer protocol; the notes currently say it's "not clear" whether they will be reintroduced.

[rootjalex]

Looks like they’ve been reintroduced! https://github.com/wjakob/nanobind/blob/master/docs/tensor.md

[alexreinking]

This does seem worth evaluating now that the buffer protocol is back.

[steven-johnson]

Agreed, but availability could be an issue -- pybind11 is widely available on many systems already, and (more importantly to me personally), nanobind isn't available inside Google yet, so we'd have to get it imported and updated on a regular basis if we adopted it.)

[steven-johnson]

FYI, I took a stab at doing a basic adaptation, but it looks like there are a number of API changes that aren't well-documented yet (eg buffer->tensor) and other things that we use that are (still) missing in nanobind -- see wjakob/nanobind#77 for a handful of things I noticed. I pushed my (very very very rough) work-in-progress to the srj/nanobind branch if anyone is interested, but right now it looks like nanobind probably needs more work to be ready for us (and we'd probably have to contribute that work).

[steven-johnson]

(One part of nanobind that is better than pybind11 is that it appears to have support for dlpack built in, which is something we really want/need, not sure if it will be backported to pybind11)

[alexreinking]

It's been almost two years. Is nanobind ready? It now provides a nb::ndarray argument type that is supposedly even more powerful than the buffer protocol binding was. They claim support with "NumPy, PyTorch, TensorFlow, JAX, etc."

I am particularly interested in nanobind for its support for the Python 3.12+ Stable ABI, which will dramatically reduce the number of wheels we'll need to publish.

https://nanobind.readthedocs.io/en/latest/porting.html#removed

https://nanobind.readthedocs.io/en/latest/ndarray.html

[steven-johnson]

I talked with a Xoogler who until recently maintained PyBind for Google; he strongly discouraged us from moving to nanobind; he considered it a 'hobby' project, and PyBind 3.x (on the way) will contain most of the improvements we want.