[BUG]: Cannot resolve numpy array as tensor input if flag write=False is set

[chrisrichardson]

Problem description

If we have a function which takes a const reference, it should be possible to do something like this:

m.def("get_dim", [](const nb::tensor<const double>& x){ return x.ndim(); });

and send in an unwriteable numpy array:

    w = np.array([[2.0,2.0]])
    w.setflags(write=False)
    assert (module.get_dim(w) == 2)

But this doesn't seem to work at the moment:

E       TypeError: get_dim(): incompatible function arguments. The following argument types are supported:
E           1. get_dim(arg: tensor[dtype=float64], /) -> int
E       
E       Invoked with types: ndarray

Removing the write=False, it works fine, but sometimes we have an unwriteable array from somewhere else...

Reproducible example code

No response

[brettc]

I think I have the same problem. I'm converting code from pybind11, and I pass in a numpy array which is constructed from a buffer:

buf = fstream.read(RAW_CHUNK_SIZE)
arr = np.frombuffer(buf, dtype=np.int8)

Passing arr into a function with just nb::tensor<> as the parameter gives me the same error as above, I'm guessing because frombuffer makes it not writeable.

[brettc]

I spent a bit of time futzing around trying to debug this. If I change this line:

nanobind/src/tensor.cpp

Line 152 in 8555cf8

if (PyObject_GetBuffer(o, view.get(), PyBUF_RECORDS)) {

to this:

  if (PyObject_GetBuffer(o, view.get(), PyBUF_RECORDS_RO)) {

Then my code works. But I have no idea about the further implications of this. There is a lot to take in!

@chrisrichardson Does this work for you?

[brettc]

It appears the readonly problem is not just with nanobind.

import numpy as np
w = np.array([[2.0,2.0]])
w.setflags(write=False)
w.__dlpack__()

gives

TypeError: NumPy currently only supports dlpack for writeable arrays

Some discussion here: numpy/numpy#20742

So is the issue that dlpack doesn't support readonly buffers?

[whateverforever]

Relevant when working with mikedh/trimesh, as the vertex and normal arrays have write=False by default

[wjakob]

I will close this issue, the current master version of nanobind addresses it to my knowledge. If not, feel free to reopen.