Get raw DLPack capsule from nb::ndarray

[awni]

From the docs I thought the following would work (e.g. no framework annotation):

      const float data[] = { 1, 2, 3, 4, 5, 6, 7, 8 };
      m.def("ret_dlpack", [](const array& a) {         
        size_t shape[2] = { 2, 4 };
        return nb::ndarray<const float, nb::shape<2, -1>>(
            /* data = */ data,                       
            /* ndim = */ 2,                          
            /* shape pointer = */ shape,             
            /* owner = */ nb::handle());             
      })    

But it fails with the error:

RuntimeError: nanobind::detail::ndarray_wrap(): copy failed: AttributeError: 'PyCapsule' object has no attribute 'copy'

Is that expected? If so, is it possible to access and return the DLPack capsule in a bound function (e.g. __dlpack__).

[wjakob]

The error message is weird, I agree with that. I think the key problem is that you didn't specify an owner, which means that nanobind must make a copy. But it cannot make a copy of a raw DLPack capsule because the data might be wherever (e.g. on a CUDA or ROCm device, and nanobind doesn't link against these libraries. So we cannot memcpy). What happens when you set the owner argument to nb::none()?

[awni]

Indeed it runs without the error if I set the owner to nb::none(). Thanks!

In the actual use case the data pointer is coming from an MLX array. Presumably the nb::ndarray will not copy the data. If that's correct, then I should use nb::rv_policy::reference_internal as the return value policy? That way the capsule keeps the array alive. When the capsule is dereferenced, the array releases the data. Does that sound correct to you?