Port compiler and iree-dialects to nanobind #19580
Conversation
marbre
changed the title
|
Unfortunately, this does not solve the failures we're seeing in the macOS clang workflow (https://github.com/iree-org/iree/actions/runs/12583361482/job/35070746468). Feel free to review but we probably don't want to land this before the release in case we revert the nanobind patch in our llvm fork. |
There was a problem hiding this comment.
Nice! Pretty minimal changes needed and tests all still pass.
IMO let's still revert the changes that broke macOS and then include changes like this after the 3.1.0 release. I want to prioritize derisking releases first and foremost.
Ideas for the PR description:
- Link to the Discourse thread with context on nanobind vs pybind
- Include any metrics we have as a result of this change, like binary size changes or performance improvements (the upstream work was motivated by performance improvements for JAX)
- Mention that now we only have a single library for bindings across the compiler and runtime
| @@ -176,6 +176,7 @@ declare_mlir_python_sources(IREECompilerPythonExtensions) | |||
| declare_mlir_python_extension(IREECompilerPythonExtensions.Registration | |||
| MODULE_NAME _site_initialize_0 | |||
| ADD_TO_PARENT IREECompilerPythonExtensions | |||
| PYTHON_BINDINGS_LIBRARY nanobind | |||
There was a problem hiding this comment.
Tested on my Windows machine. Build completed with no errors.
|
Both works for me :) @makslevental pushed a fix to this branch thus it should build for macOS now. However, I don't have strong feelings and we can revert the relevant commits in our LLVM fork for next week's release and apply them again (as well as this patch) afterwards. Another option would be to just apply the linker fix (without porting pybind11 to nanobind). I assume that would fix things as well. Btw, thanks sharing your ideas about improving the PR description. Will take those into account. |
|
Landed with #19789. |
No description provided.