python3.13-keras> =================================== FAILURES ===================================
python3.13-keras> __________________ NumpyTwoInputOpsCorrectnessTest.test_cross __________________
python3.13-keras> 
python3.13-keras> self = <keras.src.ops.numpy_test.NumpyTwoInputOpsCorrectnessTest testMethod=test_cross>
python3.13-keras> 
python3.13-keras>     def test_cross(self):
python3.13-keras>         x1 = np.ones([2, 1, 4, 3])
python3.13-keras>         x2 = np.ones([2, 1, 4, 2])
python3.13-keras>         y1 = np.ones([2, 1, 4, 3])
python3.13-keras>         y2 = np.ones([1, 5, 4, 3])
python3.13-keras>         y3 = np.ones([1, 5, 4, 2])
python3.13-keras>         self.assertAllClose(knp.cross(x1, y1), np.cross(x1, y1))
python3.13-keras>         self.assertAllClose(knp.cross(x1, y2), np.cross(x1, y2))
python3.13-keras>         if backend.backend() != "torch":
python3.13-keras>             # API divergence between `torch.cross` and `np.cross`
python3.13-keras>             # `torch.cross` only allows dim 3, `np.cross` allows dim 2 or 3
python3.13-keras> >           self.assertAllClose(knp.cross(x1, y3), np.cross(x1, y3))
python3.13-keras>                                                    ^^^^^^^^^^^^^^^^
python3.13-keras> 
python3.13-keras> keras/src/ops/numpy_test.py:3659: 
python3.13-keras> _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ 
python3.13-keras> 
python3.13-keras> a = array([[[[1., 1., 1.],
python3.13-keras>          [1., 1., 1.],
python3.13-keras>          [1., 1., 1.],
python3.13-keras>          [1., 1., 1.]]],
python3.13-keras> 
python3.13-keras> 
python3.13-keras>        [[[1., 1., 1.],
python3.13-keras>          [1., 1., 1.],
python3.13-keras>          [1., 1., 1.],
python3.13-keras>          [1., 1., 1.]]]])
python3.13-keras> b = array([[[[1., 1.],
python3.13-keras>          [1., 1.],
python3.13-keras>          [1., 1.],
python3.13-keras>          [1., 1.]],
python3.13-keras> 
python3.13-keras>         [[1., 1.],
python3.13-keras>          [1., 1.],
python3.13-keras>   ...
python3.13-keras>          [1., 1.],
python3.13-keras>          [1., 1.]],
python3.13-keras> 
python3.13-keras>         [[1., 1.],
python3.13-keras>          [1., 1.],
python3.13-keras>          [1., 1.],
python3.13-keras>          [1., 1.]]]])
python3.13-keras> axisa = 3, axisb = 3, axisc = -1, axis = None
python3.13-keras> 
python3.13-keras>     @array_function_dispatch(_cross_dispatcher)
python3.13-keras>     def cross(a, b, axisa=-1, axisb=-1, axisc=-1, axis=None):
python3.13-keras>         """
python3.13-keras>         Return the cross product of two (arrays of) vectors.
python3.13-keras>     
python3.13-keras>         The cross product of `a` and `b` in :math:`R^3` is a vector perpendicular
python3.13-keras>         to both `a` and `b`.  If `a` and `b` are arrays of vectors, the vectors
python3.13-keras>         are defined by the last axis of `a` and `b` by default, and these axes
python3.13-keras>         must have 3 dimensions.
python3.13-keras>     
python3.13-keras>         Parameters
python3.13-keras>         ----------
python3.13-keras>         a : array_like
python3.13-keras>             Components of the first vector(s).
python3.13-keras>         b : array_like
python3.13-keras>             Components of the second vector(s).
python3.13-keras>         axisa : int, optional
python3.13-keras>             Axis of `a` that defines the vector(s).  By default, the last axis.
python3.13-keras>         axisb : int, optional
python3.13-keras>             Axis of `b` that defines the vector(s).  By default, the last axis.
python3.13-keras>         axisc : int, optional
python3.13-keras>             Axis of `c` containing the cross product vector(s).  By default, the last axis.
python3.13-keras>         axis : int, optional
python3.13-keras>             If defined, the axis of `a`, `b` and `c` that defines the vector(s)
python3.13-keras>             and cross product(s).  Overrides `axisa`, `axisb` and `axisc`.
python3.13-keras>     
python3.13-keras>         Returns
python3.13-keras>         -------
python3.13-keras>         c : ndarray
python3.13-keras>             Vector cross product(s).
python3.13-keras>     
python3.13-keras>         Raises
python3.13-keras>         ------
python3.13-keras>         ValueError
python3.13-keras>             When the dimension of the vector(s) in `a` or `b` does not equal 3.
python3.13-keras>     
python3.13-keras>         See Also
python3.13-keras>         --------
python3.13-keras>         inner : Inner product
python3.13-keras>         outer : Outer product.
python3.13-keras>         linalg.cross : An Array API compatible variation of ``np.cross``.
python3.13-keras>         ix_ : Construct index arrays.
python3.13-keras>     
python3.13-keras>         Notes
python3.13-keras>         -----
python3.13-keras>         Supports full broadcasting of the inputs.
python3.13-keras>     
python3.13-keras>         Examples
python3.13-keras>         --------
python3.13-keras>         Vector cross-product.
python3.13-keras>     
python3.13-keras>         >>> import numpy as np
python3.13-keras>         >>> x = [1, 2, 3]
python3.13-keras>         >>> y = [4, 5, 6]
python3.13-keras>         >>> np.cross(x, y)
python3.13-keras>         array([-3,  6, -3])
python3.13-keras>     
python3.13-keras>         One vector with dimension 2.
python3.13-keras>     
python3.13-keras>         >>> x = [1, 2, 0]
python3.13-keras>         >>> y = [4, 5, 6]
python3.13-keras>         >>> np.cross(x, y)
python3.13-keras>         array([12, -6, -3])
python3.13-keras>     
python3.13-keras>         Both vectors with dimension 2.
python3.13-keras>     
python3.13-keras>         >>> x = [1, 2, 0]
python3.13-keras>         >>> y = [4, 5, 0]
python3.13-keras>         >>> np.cross(x, y)
python3.13-keras>         array([0, 0, -3])
python3.13-keras>     
python3.13-keras>         Multiple vector cross-products. Note that the direction of the cross
python3.13-keras>         product vector is defined by the *right-hand rule*.
python3.13-keras>     
python3.13-keras>         >>> x = np.array([[1,2,3], [4,5,6]])
python3.13-keras>         >>> y = np.array([[4,5,6], [1,2,3]])
python3.13-keras>         >>> np.cross(x, y)
python3.13-keras>         array([[-3,  6, -3],
python3.13-keras>                [ 3, -6,  3]])
python3.13-keras>     
python3.13-keras>         The orientation of `c` can be changed using the `axisc` keyword.
python3.13-keras>     
python3.13-keras>         >>> np.cross(x, y, axisc=0)
python3.13-keras>         array([[-3,  3],
python3.13-keras>                [ 6, -6],
python3.13-keras>                [-3,  3]])
python3.13-keras>     
python3.13-keras>         Change the vector definition of `x` and `y` using `axisa` and `axisb`.
python3.13-keras>     
python3.13-keras>         >>> x = np.array([[1,2,3], [4,5,6], [7, 8, 9]])
python3.13-keras>         >>> y = np.array([[7, 8, 9], [4,5,6], [1,2,3]])
python3.13-keras>         >>> np.cross(x, y)
python3.13-keras>         array([[ -6,  12,  -6],
python3.13-keras>                [  0,   0,   0],
python3.13-keras>                [  6, -12,   6]])
python3.13-keras>         >>> np.cross(x, y, axisa=0, axisb=0)
python3.13-keras>         array([[-24,  48, -24],
python3.13-keras>                [-30,  60, -30],
python3.13-keras>                [-36,  72, -36]])
python3.13-keras>     
python3.13-keras>         """
python3.13-keras>         if axis is not None:
python3.13-keras>             axisa, axisb, axisc = (axis,) * 3
python3.13-keras>         a = asarray(a)
python3.13-keras>         b = asarray(b)
python3.13-keras>     
python3.13-keras>         if (a.ndim < 1) or (b.ndim < 1):
python3.13-keras>             raise ValueError("At least one array has zero dimension")
python3.13-keras>     
python3.13-keras>         # Check axisa and axisb are within bounds
python3.13-keras>         axisa = normalize_axis_index(axisa, a.ndim, msg_prefix='axisa')
python3.13-keras>         axisb = normalize_axis_index(axisb, b.ndim, msg_prefix='axisb')
python3.13-keras>     
python3.13-keras>         # Move working axis to the end of the shape
python3.13-keras>         a = moveaxis(a, axisa, -1)
python3.13-keras>         b = moveaxis(b, axisb, -1)
python3.13-keras>         if a.shape[-1] != 3 or b.shape[-1] != 3:
python3.13-keras> >           raise ValueError(
python3.13-keras>                 f"Both input arrays must be (arrays of) 3-dimensional vectors, "
python3.13-keras>                 f"but they are {a.shape[-1]} and {b.shape[-1]} dimensional instead."
python3.13-keras>             )
python3.13-keras> E           ValueError: Both input arrays must be (arrays of) 3-dimensional vectors, but they are 3 and 2 dimensional instead.
python3.13-keras> 
python3.13-keras> /nix/store/i4ajph3w20gpspwsmhaqkbgj8mkjlr4n-python3.13-numpy-2.5.0/lib/python3.13/site-packages/numpy/_core/numeric.py:1679: ValueError
python3.13-keras> =========================== short test summary info ============================
python3.13-keras> FAILED keras/src/ops/numpy_test.py::NumpyTwoInputOpsCorrectnessTest::test_cross - ValueError: Both input arrays must be (arrays of) 3-dimensional vectors, but they are 3 and 2 dimensional instead.
python3.13-keras> = 1 failed, 12630 passed, 482 skipped, 33 deselected, 7 xfailed, 6 xpassed, 4 subtests passed in 1702.65s (0:28:22) =

Yeah, that’s separate. cctools will still be cctools, but darwin.binutils will just have LLVM bintools. There are probably still a handful of incompatibilities, but I just don’t care at this point. Things that really need that functionality can use the versions of those tools from cctools.

IIRC, unless things have changed in LLVM 22, lipo doesn’t support archives, which Meson needs for its tests; and install_name_tool doesn’t support reexports, which breaks the libiconv, but reexports are pretty rare.

There were also issues with LLVM’s version of libtool and ranlib, but meh at this point. Things can use cctools if they really need them. Using LLVM binutils and LLD should make Linux to Darwin cross mostly just work.

In reply to @reckenrode:matrix.org
It was using an unwrapped clang for CUDA stuff. I had a PR, but it sat for a while. I saw a ping today about an approach that does it the right way by passing the path to an unwrapped clang.