Implicit conversion into nb::ndarray in C++ by custom type_cast

[Char-Aznable]

Hi, thanks for this great library. I'm new to this so any suggestion is welcome. I have a custom multi-dimensional array C++ class that I need to convert into pytorch tensor upon some function call. I have written the nb::detail::type_cast specialization for it by following the Eigen array conversion:

nanobind/include/nanobind/eigen/dense.h

Line 124 in 827c083

struct type_caster<T, enable_if_t<is_eigen_plain_v<T> &&

I then define a nb::class_ method that returns directly the custom multi-dimensional array C++ class and expect I that would get the desired pytorch tensor but instead I got the error when calling the function on python: TypeError: Unable to convert function return value to a Python type!

My question is: should I expect this implicit conversion to work? What should I look into about the conversion error?

The structure of the code I mentioned looks something like:

template <class MDArray>
struct type_caster<
  MDArray,
  std::enable_if_t<isMDArray<MDArray>>>> {
  
  // a lot of code mapping the type traits of MDArray to nb::ndarray

  using NDArrayCaster = make_caster<NDArray>;

  NB_TYPE_CASTER(MDArray, NDArrayCaster::Name);

  bool from_python(handle src, uint8_t flags, cleanup_list* cleanup) noexcept =   delete;

  static handle from_cpp(
    MDArray&& a, rv_policy policy, cleanup_list* cleanup) noexcept {
   // basically following Eigen array conversion: https://github.com/wjakob/nanobind/blob/827c0837ddb60b83ef6dc946f1be5b317926f1b9/include/nanobind/eigen/dense.h#L124
  }
};

NB_MODULE(MyModule, m) {
  nb::class_<MyOperator>(m, "MyOperator").def("getMDArray", [](...) { ... return MDArray(...); } );
}

[Char-Aznable]

BTW, I have tried not using type_cast at all but merge those conversion code into the getMDArray to return a nb::ndarray and it works in that case

[Char-Aznable]

It turns out type_caster's SFINAE is not working. The type cast works if I do the following without the SFINAE:

template <class ... T>
struct type_caster<MDArray<T...>> {
// the static_assert never fires, which means the OP's code should have worked
static_assert(isMDArray<MDArray<T...>>>, "should not fire");
};

Now I'm not sure what the expected signature of the SFINAE as in the Eigen array case...