mlx-rs 0.33.0

Unofficial rust wrapper for Apple's mlx machine learning library.
1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
37
38
39
40
41
42
43
44
45
46
47
48
49
50
51
52
53
54
55
56
57
58
59
60
61
62
63
64
65
66
67
68
69
70
71
72
73
74
75
76
77
78
79
80
81
82
83
84
85
86
87
88
89
90
91
92
93
94
95
96
97
98
99
100
101
102
103
104
105
106
107
108
109
110
111
112
113
114
115
116
117
118
119
120
121
122
123
124
125
126
127
128
129
130
131
132
133
134
135
136
137
138
139
140
141
142
143
144
145
146
147
148
149
150
151
152
153
154
155
156
157
158
159
160
161
162
163

use crate::{
    error::{Exception, Result},
    utils::{guard::Guarded, Closure, IntoOption},
    Array,
};

use super::{value_and_gradient, ClosureValueAndGrad};

fn build_value_and_gradient_inner<'a>(
    closure: Closure<'a>,
    argnums: &'a [i32],
) -> impl FnMut(&[Array]) -> Result<(Vec<Array>, Vec<Array>)> + 'a {
    move |arrays: &[Array]| unsafe {
        let cvg = ClosureValueAndGrad::try_from_op(|res| {
            mlx_sys::mlx_value_and_grad(res, closure.as_ptr(), argnums.as_ptr(), argnums.len())
        })?;
        value_and_gradient(cvg.as_ptr(), arrays.iter())
    }
}

fn build_value_and_gradient<'a, F>(
    f: F,
    argnums: &'a [i32],
) -> impl FnMut(&[Array]) -> Result<(Vec<Array>, Vec<Array>)> + 'a
where
    F: FnMut(&[Array]) -> Vec<Array> + 'a,
{
    let closure = Closure::new(f);
    build_value_and_gradient_inner(closure, argnums)
}

fn build_fallible_value_and_gradient<'a, F>(
    f: F,
    argnums: &'a [i32],
) -> impl FnMut(&[Array]) -> Result<(Vec<Array>, Vec<Array>)> + 'a
where
    F: FnMut(&[Array]) -> Result<Vec<Array>> + 'a,
{
    let closure = Closure::new_fallible(f);
    build_value_and_gradient_inner(closure, argnums)
}

/// Trait for functions/closures that can be converted into a closure that computes the value and
/// gradient.
pub trait IntoValueAndGrad<'a, Err> {
    /// Convert the function/closure into a closure that computes the value and gradient.
    fn into_value_and_grad(
        self,
        argnums: impl IntoOption<&'a [i32]>,
    ) -> impl FnMut(&[Array]) -> Result<(Vec<Array>, Vec<Array>)> + 'a;
}

impl<'a, F> IntoValueAndGrad<'a, ()> for F
where
    F: FnMut(&[Array]) -> Vec<Array> + 'a,
{
    // refining_impl_trait is fine here because we have restricted the Args and Output types
    // in the generics.
    #[allow(refining_impl_trait)]
    fn into_value_and_grad(
        self,
        argnums: impl IntoOption<&'a [i32]>,
    ) -> impl FnMut(&[Array]) -> Result<(Vec<Array>, Vec<Array>)> + 'a {
        let argnums = argnums.into_option().unwrap_or(&[0]);
        build_value_and_gradient(self, argnums)
    }
}

impl<'a, F> IntoValueAndGrad<'a, Exception> for F
where
    F: FnMut(&[Array]) -> Result<Vec<Array>> + 'a,
{
    #[allow(refining_impl_trait)]
    fn into_value_and_grad(
        self,
        argnums: impl IntoOption<&'a [i32]>,
    ) -> impl FnMut(&[Array]) -> Result<(Vec<Array>, Vec<Array>)> + 'a {
        let argnums = argnums.into_option().unwrap_or(&[0]);
        build_fallible_value_and_gradient(self, argnums)
    }
}

/// Returns a function which computes the value and gradient of `f` with a
/// default argument number `&[0]`.
///
/// See also [`value_and_grad_with_arg_nums`] for a version that allows
/// specifying the argument numbers
pub fn value_and_grad<'a, F, Err>(
    f: F,
) -> impl FnMut(&[Array]) -> Result<(Vec<Array>, Vec<Array>)> + 'a
where
    F: IntoValueAndGrad<'a, Err> + 'a,
{
    f.into_value_and_grad(None)
}

/// Returns a function which computes the value and gradient of `f`.
///
/// See also [`value_and_grad`] for a version that uses the default argument
/// numbers `&[0]`.
pub fn value_and_grad_with_argnums<'a, F, Err>(
    f: F,
    argnums: impl IntoOption<&'a [i32]>,
) -> impl FnMut(&[Array]) -> Result<(Vec<Array>, Vec<Array>)> + 'a
where
    F: IntoValueAndGrad<'a, Err> + 'a,
{
    f.into_value_and_grad(argnums)
}

#[cfg(test)]
mod tests {

    use crate::{array, transforms::value_and_grad, Array};

    use super::*;

    // The unit tests below are adapted from the mlx c++ codebase
    #[test]
    fn test_value_and_grad() {
        let x = &[Array::from_f32(1.0)];
        let fun = |argin: &[Array]| -> Vec<Array> { vec![&argin[0] + 1.0] };
        let argnums = &[0];
        let (y, dfdx) = value_and_grad_with_argnums(fun, argnums)(x).unwrap();
        assert_eq!(y[0].item_exact::<f32>(), 2.0);
        assert_eq!(dfdx[0].item_exact::<f32>(), 1.0);

        let (y, dfdx) = value_and_grad(fun)(x).unwrap();
        assert_eq!(y[0].item_exact::<f32>(), 2.0);
        assert_eq!(dfdx[0].item_exact::<f32>(), 1.0);
    }

    #[test]
    fn test_value_and_grad_with_error() {
        let fun = |argin: &[Array]| -> Result<Vec<Array>> {
            argin[0].add(array!(1.0)).map(|res| vec![res])
        };

        // Success case
        let argnums = &[0];
        let x = array!(1.0f32);
        let y = array!(1.0f32);
        let args = &[x, y];
        let result = value_and_grad_with_argnums(fun, argnums)(args);
        assert!(result.is_ok());
        let result = value_and_grad(fun)(args);
        assert!(result.is_ok());

        // Error case
        // Use non-broadcastable shapes
        let a = array!([1.0, 2.0, 3.0]);
        let b = array!([4.0, 5.0]);
        let args = &[a, b];
        let result = value_and_grad_with_argnums(fun, argnums)(args);
        assert!(result.is_err());
        let result = value_and_grad(fun)(args);
        assert!(result.is_err());

        // Check that the error is not just "mlx_closure returned a non-zero value"
        let err = result.unwrap_err();
        assert!(!err.what().contains("non-zero value"))
    }
}