Incorrect docstrings for PyTorch tensor class

[chrismile]

Hi,

I face the problem that docstrings generated by pybind11 for functions with PyTorch at::Tensor objects have at::Tensor in the Python interface instead of torch.Tensor, e.g.:

foo(X: at::Tensor, k: int) -> at::Tensor

I saw that the docs were mentioning this type of problem (https://pybind11.readthedocs.io/en/stable/advanced/misc.html#avoiding-c-types-in-docstrings).

I tried the following workaround, but to no avail (at::Tensor still remains in the docstring).

#include <torch/script.h>
#include <torch/types.h>
#include <torch/extension.h>

at::Tensor foo(at::Tensor X, int k) { return X; }

PYBIND11_MODULE(torch, m) {
    auto tensorClass = py::class_<at::Tensor>(m, "Tensor");
    tensorClass.def(py::init());
}
PYBIND11_MODULE(mymodule, m) {
    m.def("foo", foo, "Description here", py::arg("X"), py::arg("k"));
}

Am I using pybind11 incorrectly or is this actually a bug?

Here's the corresponding Python code:

import torch
import mymodule
help(mymodule.foo)

[chrismile]

Seems like this is a PyTorch bug, which was fixed last November by pytorch/pytorch#89115. I recommend to anyone facing the same issues as me to upgrade their PyTorch version.