_PyBytes_Join was not declared in this scope

[i-l-s-h-a-t]

Dear

I have an issue during python wheel generation for my module on Linux, Python 3.13:

      ../numpy/_core/src/common/pythoncapi-compat/pythoncapi_compat.h: In function ‘PyObject* PyBytes_Join(PyObject*, PyObject*)’:
      ../numpy/_core/src/common/pythoncapi-compat/pythoncapi_compat.h:1607:12: error: ‘_PyBytes_Join’ was not declared in this scope; did you mean ‘PyBytes_Join’?

My code has numpy dependency and it uses pybind11 C++ library. I generate a wheel for it.

Could it be due to this renaming?
There is a check here but it checks only Python 3.14, and not 3.13:

I see there is a backward compatibility was assumed

As you can see it doesn't work in my case.

The whole log is here https://pastebin.com/Am07UcZg
I have this issue only for Python 3.13, only on Linux (docker image python:3.13). Just install numpy and that works there. Looks like for my wheel creation, it does something else which fails.
And there is no such an issue for other Python versions (3.12 or 3.14) and Windows (Python 3.10-3.14).

The error appears in gitlab CI/CD for my project. There are those images used:

I could not trigger numpy compilation using one of that images and reproduce the issue locally. I don't know how to run two layered docker image locally, although gitlab manages to do that inside the runner.
I tried to compile numpy using Python 3.13 from conda on a Linux machine (not that docker image) and it succeeded.

I will appreciate you help.

Best,
ilshat

[vstinner]

What are the steps to reproduce your issue?

[vstinner]

And there is no such an issue for other Python versions (3.12 or 3.14) and Windows (Python 3.10-3.14). The error appears in gitlab CI/CD for my project.

So maybe something is wrong with the docker image used by gitlab CI/CD. Try another image maybe?

[i-l-s-h-a-t]

Dear @vstinner
Sorry, I cannot provide more details.
I tried to reproduce the issue locally using:

1. dockcross/manylinux-2014-x64 image and install python, then I have SSL issue so I can't install build module for python there. I could not resolve it to proceed.
2. Installing conda (instead of Python): conda doesn't work on manylinux-2014 as the system in that image seems to be too old for conda build.
   If you can suggest I can add commands to CI/CD pipeline script to see what happens there.

[i-l-s-h-a-t]

And there is no such an issue for other Python versions (3.12 or 3.14) and Windows (Python 3.10-3.14). The error appears in gitlab CI/CD for my project.

So maybe something is wrong with the docker image used by gitlab CI/CD. Try another image maybe?

I tried another image python:3.13**.9** and the issue persists for it as well. Can you maybe suggest any other images to try?

[vstinner]

According to what you wrote, the issue is not in numpy or pythoncapi-compat, but in the docker image used to run your tests. Otherwise, you would be be able to reproduce the issue locally.

[i-l-s-h-a-t]

According to what you wrote, the issue is not in numpy or pythoncapi-compat, but in the docker image used to run your tests. Otherwise, you would be be able to reproduce the issue locally.

Dear @vstinner, thank you for working on this issue.

I just could not manage to run it locally in the same way as it run in CI/CD, since I have not much experience in installing Python in dockers.

What do you think on this behavior in my log - when compiling different c files, the compile shows warnings in two cases and an error in one case:

In your pipeline Ubuntu is used, could it be that -Wimplicit-function-declaration different than in my case?

Also, from my log I see this hash for numpy f170575ce40f4f5993e0d3ee5648d07a. However, there is no such a commit in numpy.

I hope these notes could help you to guess the cause.

Kind regards,
ilshat

[vstinner]

In Python 3.13, _PyBytes_Join() is declared in Include/cpython/bytesobject.h:

$ grep -A1 _PyBytes_Join Include/cpython/bytesobject.h
/* _PyBytes_Join(sep, x) is like sep.join(x).  sep must be PyBytesObject*,
   x must be an iterable object. */
PyAPI_FUNC(PyObject*) _PyBytes_Join(PyObject *sep, PyObject *x);