Writable tensor?

[qnzhou]

I am wondering if it possible to expose C++ buffers as a writable tensor in python. E.g. if I have the Foo class as

struct Foo {
    std::vector<int> tags = {1, 2, 3};
};

and the tags member variable is exposed in python using nb::tensor<> like the following.

using IntArray = nb::tensor<nb::numpy, int, nb::shape<nb::any>>;
nb::class_<Foo>(m, "Foo")
    .def_property("tags", 
    [](Foo& foo) {
        size_t shape[] = {foo.tags.size()};
        return IntArray(reinterpret_cast<void*>(foo.tags.data()), 1, shape);
    }, [](Foo& foo, IntArray tensor) {
        ...
    });

In python, the numpy array corresponding to tags appears to be not writable.

a = Foo()
print(a.tags.flags)
#  C_CONTIGUOUS : True
#  F_CONTIGUOUS : True
#  OWNDATA : False
#  WRITEABLE : False
#  ALIGNED : True
#  WRITEBACKIFCOPY : False

Is there a way of creating writable tensor such that in place update like a.tags[0] = 10 is allowed?

[brettc]

As @qnzhou said in #53, this looks like a limitation of using from_dlpack() and is by design. A related issue is the one mentioned in #42, where you cannot send a read-only numpy array to c++, though it looks like this might be worked around in nanobind.

So working through dlpack currently puts some constraints on passing tensors.

1. You can't pass writeable tensors from c++ -> python.
2. You can't pass read-only tensors from python -> c++.

I see that the reimplementation of the numpy interface is "open for discussion". @wjakob, is it worthwhile submitting a PR that ports the numpy code from pybind11? I could give this a go, though it might be a challenge given my limited understanding of the internals.

[wjakob]

I would be fine with having a nb::writable annotation to force the buffer protocol path in NumPy.

[qnzhou]

An interesting observation is that with older numpy (e.g. v1.21.5), the numpy array that wraps around a nanobind tensor is writeable out of the box. With newer version of numpy (e.g. v1.24.1), it is not, and the writeable flag cannot be set.

[wjakob]

That's likely the cutoff where nanobind switches over from talking to numpy using the buffer protocol vs DLPack.

[qnzhou]

FYI, this issue seems to be fixed in v1.0.0 release with ndarray replacing tensor. I still have a few questions:

• What exact change fixed this issue? (Just curious).
• Now, is there a way to bind a read-only buffer with ndarray with writable bit off? Ideally without const_cast.

[wjakob]

For #1, see the changelog. For #2, unfortunately not yet. nb::ndarray does not include a read-only bit yet, help on this front is appreciated.