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
164
165
166
167
168
169
170
171
172
173
174
175
176
177
178
179
180
181
182
183
184
185
186
187
188
189
190
191
192
193
194
195
196
197
198
199
200
201
202
203
204
205
206
207
208
209
210
211
212
213
214
215
216
217
218
219
220
221
222
223
224
225
226
227
228
229
230
231
232
233
234
235
236
237
238
239
240
241
242
243
244
245
246
247
248
249
250
251
252
253
254
255
256
257
258

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

use super::{value_and_gradient, ClosureValueAndGrad};

#[inline]
fn build_gradient_inner<'a>(
    closure: Closure<'a>,
    argnums: &'a [i32],
) -> impl FnMut(&[Array]) -> Result<Vec<Array>> + 'a {
    move |arrays: &[Array]| -> Result<Vec<Array>> {
        let cvg = ClosureValueAndGrad::try_from_op(|res| unsafe {
            mlx_sys::mlx_value_and_grad(res, closure.as_ptr(), argnums.as_ptr(), argnums.len())
        })?;
        let result = value_and_gradient(cvg.as_ptr(), arrays.iter())?;
        Ok(result.1)
    }
}

fn build_gradient<'a, F>(
    f: F,
    argnums: &'a [i32],
) -> impl FnMut(&[Array]) -> Result<Vec<Array>> + 'a
where
    F: FnMut(&[Array]) -> Vec<Array> + 'a,
{
    let argnums = argnums.into_option().unwrap_or(&[0]);
    let closure = Closure::new(f);
    build_gradient_inner(closure, argnums)
}

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

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

impl<'a, F> IntoGrad<'a, &[Array], Vec<Array>, ()> 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_grad(
        self,
        argnums: impl IntoOption<&'a [i32]>,
    ) -> impl FnMut(&[Array]) -> Result<Vec<Array>> + 'a {
        let argnums = argnums.into_option().unwrap_or(&[0]);
        build_gradient(self, argnums)
    }
}

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

impl<'a, F> IntoGrad<'a, &Array, Array, ()> for F
where
    F: FnMut(&Array) -> Array + 'a,
{
    #[allow(refining_impl_trait)]
    fn into_grad(
        mut self,
        argnums: impl IntoOption<&'a [i32]>,
    ) -> impl FnMut(&Array) -> Result<Array> + 'a {
        let f = move |args: &[Array]| -> Vec<Array> { vec![self(&args[0])] };
        let argnums = argnums.into_option().unwrap_or(&[0]);
        let mut g = build_gradient(f, argnums);
        move |args: &Array| -> Result<Array> {
            let args_clone = std::slice::from_ref(args);
            let result = g(args_clone)?;
            Ok(result.into_iter().next().unwrap())
        }
    }
}

impl<'a, F> IntoGrad<'a, &Array, Array, Exception> for F
where
    F: FnMut(&Array) -> Result<Array> + 'a,
{
    #[allow(refining_impl_trait)]
    fn into_grad(
        mut self,
        argnums: impl IntoOption<&'a [i32]>,
    ) -> impl FnMut(&Array) -> Result<Array> + 'a {
        let f = move |args: &[Array]| -> Result<Vec<Array>> { self(&args[0]).map(|res| vec![res]) };
        let argnums = argnums.into_option().unwrap_or(&[0]);
        let mut g = build_fallible_gradient(f, argnums);
        move |args: &Array| -> Result<Array> {
            let args_clone = std::slice::from_ref(args);
            let result = g(args_clone)?;
            Ok(result.into_iter().next().unwrap())
        }
    }
}

impl<'a, F> IntoGrad<'a, &[Array], Array, ()> for F
where
    F: FnMut(&[Array]) -> Array + 'a,
{
    #[allow(refining_impl_trait)]
    fn into_grad(
        mut self,
        argnums: impl IntoOption<&'a [i32]>,
    ) -> impl FnMut(&[Array]) -> Result<Array> + 'a {
        let f = move |args: &[Array]| -> Vec<Array> { vec![self(args)] };
        let argnums = argnums.into_option().unwrap_or(&[0]);
        let mut g = build_gradient(f, argnums);
        move |args: &[Array]| -> Result<Array> {
            let result = g(args)?;
            Ok(result.into_iter().next().unwrap())
        }
    }
}

impl<'a, F> IntoGrad<'a, &[Array], Array, Exception> for F
where
    F: FnMut(&[Array]) -> Result<Array> + 'a,
{
    #[allow(refining_impl_trait)]
    fn into_grad(
        mut self,
        argnums: impl IntoOption<&'a [i32]>,
    ) -> impl FnMut(&[Array]) -> Result<Array> + 'a {
        let f = move |args: &[Array]| -> Result<Vec<Array>> { self(args).map(|res| vec![res]) };
        let argnums = argnums.into_option().unwrap_or(&[0]);
        let mut g = build_fallible_gradient(f, argnums);
        move |args: &[Array]| -> Result<Array> {
            let result = g(args)?;
            Ok(result.into_iter().next().unwrap())
        }
    }
}

impl<'a, F> IntoGrad<'a, &Array, Vec<Array>, ()> for F
where
    F: FnMut(&Array) -> Vec<Array> + 'a,
{
    #[allow(refining_impl_trait)]
    fn into_grad(
        mut self,
        argnums: impl IntoOption<&'a [i32]>,
    ) -> impl FnMut(&Array) -> Result<Vec<Array>> + 'a {
        let f = move |args: &[Array]| -> Vec<Array> { self(&args[0]) };
        let argnums = argnums.into_option().unwrap_or(&[0]);
        let mut g = build_gradient(f, argnums);
        move |args: &Array| -> Result<Vec<Array>> {
            let args_clone = std::slice::from_ref(args);
            let result = g(args_clone)?;
            Ok(result)
        }
    }
}

impl<'a, F> IntoGrad<'a, &Array, Vec<Array>, Exception> for F
where
    F: FnMut(&Array) -> Result<Vec<Array>> + 'a,
{
    #[allow(refining_impl_trait)]
    fn into_grad(
        mut self,
        argnums: impl IntoOption<&'a [i32]>,
    ) -> impl FnMut(&Array) -> Result<Vec<Array>> + 'a {
        let f = move |args: &[Array]| -> Result<Vec<Array>> { self(&args[0]) };
        let argnums = argnums.into_option().unwrap_or(&[0]);
        let mut g = build_fallible_gradient(f, argnums);
        move |args: &Array| -> Result<Vec<Array>> {
            let args_clone = std::slice::from_ref(args);
            let result = g(args_clone)?;
            Ok(result)
        }
    }
}

/// Returns a function which computes the gradient of `f` with the default
/// argument numbers `&[0]`.
///
/// See also [`grad_with_arg_nums`] for a version that allows specifying the
/// argument numbers
pub fn grad<'a, F, Args, Output, Err>(f: F) -> impl FnMut(Args) -> Result<Output> + 'a
where
    F: IntoGrad<'a, Args, Output, Err>,
{
    f.into_grad(None)
}

/// Returns a function which computes the gradient of `f`.
///
/// See also [`grad`] for a version that uses the default argument numbers
/// `&[0]`.
pub fn grad_with_argnums<'a, F, Args, Output, Err>(
    f: F,
    argnums: impl IntoOption<&'a [i32]>,
) -> impl FnMut(Args) -> Result<Output> + 'a
where
    F: IntoGrad<'a, Args, Output, Err>,
{
    f.into_grad(argnums)
}

#[cfg(test)]
mod tests {

    use crate::{
        transforms::{grad, grad_with_argnums, value_and_grad, value_and_grad_with_argnums},
        Array,
    };

    // The unit tests below are adapted from the mlx c++ codebase
    #[test]
    fn test_grad() {
        let x = &[Array::from_f32(1.0)];
        let fun = |argin: &[Array]| -> Vec<Array> { vec![&argin[0] + 1.0] };
        let argnums = &[0];

        // TODO: how to make this more "functional"?
        let grad_fn =
            move |args: &[Array]| -> Vec<Array> { grad_with_argnums(fun, argnums)(args).unwrap() };
        let (z, d2fdx2) = value_and_grad_with_argnums(grad_fn, argnums)(x).unwrap();

        assert_eq!(z[0].item_exact::<f32>(), 1.0);
        assert_eq!(d2fdx2[0].item_exact::<f32>(), 0.0);

        let grad_fn = move |args: &[Array]| -> Vec<Array> { grad(fun)(args).unwrap() };
        let (z, d2fdx2) = value_and_grad(grad_fn)(x).unwrap();

        assert_eq!(z[0].item_exact::<f32>(), 1.0);
        assert_eq!(d2fdx2[0].item_exact::<f32>(), 0.0);
    }
}