Migrate to nanobind, remove pybind11 by BenWibking · Pull Request #595 · AMReX-Codes/pyamrex

BenWibking · 2026-07-18T23:57:49Z

This is a vibe-coded experiment to see if it is feasible to replace pybind11 with nanobind. Apparently, it is!

"~4× faster compile time"
"~5× smaller binaries"
" ~10× lower runtime overheads compared to pybind11"
Hidden bonus feature: DLPack batteries included via nb::ndarray<...>

Generated using ~1 million tokens from GPT-5.6-Sol.

ax3l · 2026-07-19T01:13:23Z

Cool! @BenWibking van you try to use nanobind's stub generator to replace our pybind11-stubgen? We need stubs for docs and ipython completion.

BenWibking · 2026-07-19T03:34:30Z

+                ${pyAMReX_SOURCE_DIR}/src/amrex/space${D}d/amrex_${D}d_pybind/__init__.pyi
+            PYTHON_PATH $<TARGET_FILE_DIR:pyAMReX_${D}d>
+            DEPENDS pyAMReX_${D}d
+        )


@ax3l nanobind stubs are generated here using nanobind's CMake interface: https://nanobind.readthedocs.io/en/latest/typing.html#cmake-interface

BenWibking · 2026-07-19T03:38:34Z

Amazingly, all tests pass on all compilers. I'm marking it ready for review.



-void init_TagBox (py::module& m)
+void init_TagBox (nb::module_& m)


 {
    template<int dim>
-    void init_IntVectND(py::module &m)
+    void init_IntVectND(nb::module_ &m)




-void init_RealBox(py::module &m) {
+void init_RealBox(nb::module_ &m) {




-void init_RealVect(py::module &m) {
+void init_RealVect(nb::module_ &m) {




-void init_PhysBCFunct(py::module& m)
+void init_PhysBCFunct(nb::module_& m)


BenWibking · 2026-07-19T12:25:11Z

 {
    template<typename T>
-    void make_FabArray_T(py::module &m, std::string const &name)
+    void make_FabArray_T(nb::module_ &m, std::string const &name)


I don't understand this comment. @ax3l do you?

Huh, I think that is a false positive.

            ArrayOfStructs<T_ParticleType, amrex::PinnedArenaAllocator> h_data;
            h_data.resize(aos.size());
-            //py::array_t<T_ParticleType> h_data(aos.size());
+            //nb::array_t<T_ParticleType> h_data(aos.size());


ax3l

This looks pretty great!

I added a few inline thoughts to research.

I will push a commit to undo the .pyi changes and might isolate the setup.py binary distr update once I understand it to a separate PR.

ax3l · 2026-07-29T17:14:58Z

+cmake_minimum_required(VERSION 3.24)
+project(pyAMReXDownstreamNanobind LANGUAGES C CXX)
+
+find_package(Python 3.8 COMPONENTS Interpreter Development.Module REQUIRED)


Oh, that's a neat downstream integration test.

The AI came up with that after it broke on an earlier change. I think it's useful.

ax3l · 2026-07-29T18:15:32Z

+            PYTHON_PATH $<TARGET_FILE_DIR:pyAMReX_${D}d>
+            DEPENDS pyAMReX_${D}d
+        )
+        add_dependencies(pyAMReX_stubs pyAMReX_${D}d_stub)


We might need to add a dependency pyAMReX_${D}d_stub -> pyAMReX_${D}d as well.

Maybe also env var helpers in case we do not want to rely on a pip_install (to check).

BenWibking · 2026-07-29T18:37:03Z

I was confused to whether the .pyi files should be checked in. What is the policy on that? Is it documented somewhere?

ax3l · 2026-07-29T18:44:50Z

Sorry for the confusion.

Yes, the .pyi files are checked in, because we need them to build Sphinx docs from.
(We have not enough CPU hours to rebuild the C++ bindings in RTD CI builders.)
Additionally, tools like IPython's Jedi auto-complete need them to give interactive users auto-completion hints.

But, they are only checked in via an auto-commit on development.
The reason for that is that if PRs change them, then nearly every PR will have a merge conflict with other open PRs. (I tried for a few days and realized that does not scale.)

Thus, our PR's CIs do build them, check they are valid, but they do not commit them.
Only the development branch does the last step, too: commit on changes.

yes, we should document this workflow / automation in the development or maintenance sections of the docs:
https://pyamrex.readthedocs.io

BenWibking · 2026-07-29T18:46:39Z

Sorry for the confusion.

Yes, the .pyi files are checked in, because we need them to build Sphinx docs from.

But, they are only checked in via an auto-commit on development. The reason for that is that if PRs change them, then nearly every PR will have a merge conflict with other open PRs. So that does not scale.

ah, okay :) That makes sense.

ax3l · 2026-07-29T18:49:36Z

Before we merge this in (maybe early in a release cycle), I would do a whole BLAST build test to see if mixing nanobind and pybind11 works well for the mid-term, then migrate the rest of BLAST too. (I expect no issues.)

BenWibking · 2026-07-29T19:10:11Z

Ok, I can extract the build fixes to a separate PR. I have some uncommitted fixes to make things easier when using uv to build/install.

for more information, see https://pre-commit.ci

We do not want to modify the generated stub files on this branch. Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>

Resolutions port the new features to nanobind: - dependencies.json: keep nanobind pins, take new AMReX commit - PlotFileUtil.cpp: port write_multi_level_plotfile to nb:: - ParticleHeader.cpp: port new bindings to nb:: (def_rw, nb::arg) - ParticleContainer.H, pyAMReX.cpp: nb:: for merged declarations Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>

BenWibking · 2026-07-29T20:03:48Z

Build fixes moved to #599.

plan

8929924

BenWibking mentioned this pull request Jul 19, 2026

[Experiment] Nanobind #127

Closed

BenWibking commented Jul 19, 2026

View reviewed changes

BenWibking marked this pull request as ready for review July 19, 2026 03:38

BenWibking changed the title ~~[Experiment] Nanobind 2026~~ Migrate to nanobind, remove pybind11 Jul 19, 2026

ax3l self-assigned this Jul 19, 2026

ax3l self-requested a review July 19, 2026 04:30

github-advanced-security AI found potential problems Jul 19, 2026

View reviewed changes

ax3l reviewed Jul 29, 2026

View reviewed changes

BenWibking and others added 11 commits July 29, 2026 14:53

[Experiment] Nanobind 2026

ae6eac6

[pre-commit.ci] auto fixes from pre-commit.com hooks

5d7b859

for more information, see https://pre-commit.ci

MPI fix

2f05341

Merge branch 'nanobind' of github.com:BenWibking/pyamrex into nanobind

1a79edd

fix pytest

c56a823

fix MPI test

5ebb62d

fix warnings

18fef93

Delete docs/nanobind_migration_inventory.md

0397152

Delete docs/nanobind_migration_plan.md

366cfa4

Revert .pyi stub changes vs development

7b690c4

We do not want to modify the generated stub files on this branch. Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>

BenWibking force-pushed the nanobind branch from e4e644d to 5817151 Compare July 29, 2026 20:01



		void init_TagBox (py::module& m)
		void init_TagBox (nb::module_& m)



		void init_RealBox(py::module &m) {
		void init_RealBox(nb::module_ &m) {



		void init_RealVect(py::module &m) {
		void init_RealVect(nb::module_ &m) {



		void init_PhysBCFunct(py::module& m)
		void init_PhysBCFunct(nb::module_& m)

Uh oh!

Conversation

BenWibking commented Jul 18, 2026 • edited Loading Uh oh! There was an error while loading. Please reload this page.

Uh oh!

Uh oh!

ax3l commented Jul 19, 2026

Uh oh!

BenWibking Jul 19, 2026 • edited Loading Uh oh! There was an error while loading. Please reload this page.

Uh oh!

Choose a reason for hiding this comment

Uh oh!

BenWibking commented Jul 19, 2026

Uh oh!

BenWibking Jul 19, 2026

Choose a reason for hiding this comment

Uh oh!

ax3l Jul 29, 2026

Choose a reason for hiding this comment

Uh oh!

ax3l left a comment

Choose a reason for hiding this comment

Uh oh!

ax3l Jul 29, 2026

Choose a reason for hiding this comment

Uh oh!

BenWibking Jul 29, 2026

Choose a reason for hiding this comment

Uh oh!

ax3l Jul 29, 2026

Choose a reason for hiding this comment

Uh oh!

Uh oh!

Uh oh!

BenWibking commented Jul 29, 2026

Uh oh!

ax3l commented Jul 29, 2026 • edited Loading Uh oh! There was an error while loading. Please reload this page.

Uh oh!

Uh oh!

BenWibking commented Jul 29, 2026

Uh oh!

ax3l commented Jul 29, 2026 • edited Loading Uh oh! There was an error while loading. Please reload this page.

Uh oh!

Uh oh!

BenWibking commented Jul 29, 2026

Uh oh!

BenWibking commented Jul 29, 2026

Uh oh!

Reviewers

Assignees

Labels

Projects

Milestone

Development

Uh oh!

3 participants

BenWibking commented Jul 18, 2026 •

edited

Loading

BenWibking Jul 19, 2026 •

edited

Loading

ax3l commented Jul 29, 2026 •

edited

Loading

ax3l commented Jul 29, 2026 •

edited

Loading