[BUG]: nb::c_contig constraint is not enforced for non-contiguous PyTorch arrays

[bathal1]

Problem description

When binding a function processing tensors, one can specify contiguity flags, like nb::c_contig. Nanobind is expected to ensure the input tensors are contiguous in memory before calling the function.

However, when passing a non-contiguous PyTorch tensor to such a function will raise an error, instead of making it contiguous under the hood.

Modifying the Nanobind example project with the following function definition allows to reproduce this issue.

Function binding

m.def("add", [](nb::ndarray<float, nb::c_contig> a, nb::ndarray<float, nb::c_contig> b) { return 0; }, "a"_a, "b"_a);

Reproducer

from nanobind_example import add
import drjit as dr
import torch
a = torch.ones((3,3))
b = dr.ones(dr.llvm.TensorXf, (3,3))

# DrJit tensors, works as expected
print(b,b)
print(b,b[::-1])

# PyTorch tensors, fails
print(add(a,a))  # Contiguous tensors, returns 0
print(add(a,a.T)) # a.T is non-contiguous, this fails

This raises the following issue

Traceback (most recent call last):
  File "<stdin>", line 11, in <module>
TypeError: add(): incompatible function arguments. The following argument types are supported:
    1. add(a: ndarray[dtype=float32, order='C'], b: ndarray[dtype=float32, order='C']) -> int

Invoked with types: torch.Tensor, torch.Tensor

[wjakob]

Just double-checking: wasn't the error in the cholespy repository about a crash? (Rather than a TypeError mentioned here).

[bathal1]

The bug mentioned a crash, but the reproducer produced this error, without crashing.

[wjakob]

Does that mean that the code in the original issue didn't have the annotation? Otherwise it sounds like there are two separate issues.

[bathal1]

The bindings in cholespy already had the nb::c_contig annotations, prior to this issue.

[wjakob]

Fixed in 23ea320.

[wjakob]

@bathal1 : I released nanobind v1.5.2 -- you might want to re-release cholespy with those fixes.