metaclass differs in nanobind.stubgen

[matwey]

Hello,

I am using nanobind.stubgen to generate pyi files for my module. When I am validating the result with mypy.stubtest, then I see the following error for every exported class:

error: pyweif.sf.Mono is inconsistent, metaclass differs
Stub: in file /home/matwey/venv/3.13/lib64/python3.13/site-packages/pyweif/sf.pyi:6
N/A
Runtime:
<class 'nanobind.nb_type_0'>

What do I do with it?

[wjakob]

That's a quite pedantic error. I guess you could patch the pyi files to include the nanobind metaclass.

[matwey]

I guess this error appears for every nanobind-based package, right?

[wjakob]

Yes. You are likely the first user who has used mypy.stubtest with nanobind-generated stubs. My own experience with MyPy is not so great, I personally prefer PyRight and derivatives.

[wjakob]

btw, one issue with adding packages from the nanobind module to make MyPy's stub checker happy is that these are created on the fly, potentiallyby any package using nanobind. There could be arbitrarily many different nb_type_... classes based on the size of the nb::supplement<T>. They are in the nanobind module but the official nanobind PyPI package actually doesn't include them (it only ships Python code, but these classes are written in C++ and dynamically created by libnanobind). Trying to expose this behavior in the stubs sounds like an incredible pain without any tangible benefit on type checking. The custom metaclass does not cause any different behavior and is an implementation detail of nanobind.

I suggest adding a # type:ignore directive or something of that sort.

[matwey]

Basically, type:ignore won't solve the issue. stubtest is very simple and cannot ignore separate errors, only the whole type. And if I ignore the type then the check is useless because all the types are affected.

What I initially want to achieve is to prove that my generated TypeStubs are consistent with C++ code (This must be very generic problem for almost every nanobind-based project, right?). It is possible do it in two ways:

1. Regenerate TypeStubs in CI in pre.commit
2. Check the consistency (manually or in CI)

Now, I am trying the latter way, but it seems that it doesn't work for me due to stubtest limitations and this metaclass issue. You mentioned that you prefer pyright but from what I can see, pyright isn't able to check consistency only the style, right?