Exposed array flags

[stemann]

Array flags could be convenient to be able to check - in particular when using the C interface from a system which typically uses column-major ordering such as Julia.

[awni]

Array flags in MLX are kind of like strides. They are meaningless unless the array is evaluated. That's one reason we don't expose strides and flags in e.g. Python (because it's pretty easy to use them incorrectly).

What do you plan to use them for? Usually it's an error to use the flags directly.. but it would be helpful to understand if you have a reasonable use case in which case we can consider exposing them.

We could for example expose them but raise / throw if you try to access flags on an unevaluated array.

[stemann]

The use case here was to check/ensure that the array was properly represented/stored - because it's likewise pretty easy to get confused by row-major and column-major representations :-)

E.g. I'm currently reversing the dimension order, returned by mlx_array_ndim, when reporting the size of a (row-major order) MLX array when reporting the size to Julia (which usually expects a column-major ordering):

function Base.size(array::MLXArray)
    return Tuple(
        Int.(
            reverse(
                unsafe_wrap(
                    Vector{Cint},
                    Wrapper.mlx_array_shape(array.mlx_array),
                    Wrapper.mlx_array_ndim(array.mlx_array),
                ),
            )
        ),
    )
end

https://github.com/stemann/MLX.jl/blob/master/src/array.jl#L65-L77

Exposing the flags would also allow replication of the C++ tests that checks the flags on the Julia-side.

It would be fine to expose the flags and throw an exception for non-available arrays.