std::bad_cast on returning ndarray of vector

[jeffective]

I have a std::bad_cast when I attempt to read this iomap property of my class by returning an ndarray

class

using BytesVector = std::vector<uint8_t>;
using BytesArray = nb::ndarray<nb::numpy, uint8_t, nb::shape<-1>, nb::c_contig>;


class SOEM_wrapper {
    public:
  
    BytesVector iomap;

    SOEM_wrapper(uint16_t maxslave_, uint8_t maxgroup_, size_t iomap_size_bytes){
        if (maxslave_ == 0) {
                throw std::invalid_argument("maxslave cannot be zero.");
            }
        if (maxgroup_ == 0) {
            throw std::invalid_argument("maxgroup cannot be zero.");
        }
        if (iomap_size_bytes == 0) {
            throw std::invalid_argument("iomap_size_bytes cannot be zero.");
        }

        iomap.resize(iomap_size_bytes, 0); 
    }
}

binding

.def_prop_ro("iomap", [](SOEM_wrapper &wrapper){
            size_t shape[1] = {wrapper.iomap.size()};
            return BytesArray(wrapper.iomap.data(), 1, shape, nb::handle());
        }, nb::rv_policy::reference_internal)

python code that produces the error

main_device = SOEM(maxslave=512, maxgroup=2, iomap_size_bytes=4096)
print(main_device.iomap)

error

terminate called after throwing an instance of 'std::bad_cast'
  what():  std::bad_cast
Aborted

[jeffective]

whats odd is this doesnt happen has part of a test

def test_iomap_mutability():
    context = SOEM(maxslave=512, maxgroup=2, iomap_size_bytes=4096)
    print(context.iomap)

[wjakob]

Can you see if #549 helps?

[jeffective]

I used these instructions https://stackoverflow.com/questions/27567846/how-can-i-check-out-a-github-pull-request-with-git

to put my nanobind submodule at commit b0f684d (which is correct for PR #549)

and recompiled etc.

and the error still occurs as part of a regular python script (but not as part of a pytest).

[jeffective]

the error also does not occur when running using the python debugger in vscode

[jeffective]

i'm not able to repeat the error after rebuilding my vscode dev container. perhaps the 100's of pip install . during development caught up to me. closing