mlxrs 0.1.0

Safe Rust bindings for Apple's MLX array framework, with LM, VLM, audio, and embeddings support
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

//! M2 piece D — happy-path tests for random ops.
//!
//! Each sampler test fixes a seed via `key`, draws once, and asserts on shape
//! and dtype (and a sanity property such as "all values are bounded"); we
//! avoid direct value checks because the underlying counter-based PRNG is
//! mlx-internal and may shift across mlx-c versions.

use mlxrs::{Array, Dtype, ops::random};

#[test]
fn key_yields_u32_pair() {
  let k = random::key(0).unwrap();
  assert_eq!(k.shape(), vec![2]);
  assert_eq!(k.dtype().unwrap(), Dtype::U32);
}

#[test]
fn split_yields_two_keys() {
  let k = random::key(42).unwrap();
  let (a, b) = random::split(&k).unwrap();
  assert_eq!(a.shape(), vec![2]);
  assert_eq!(b.shape(), vec![2]);
}

#[test]
fn split_num_yields_n_keys() {
  let k = random::key(42).unwrap();
  let s = random::split_num(&k, 4).unwrap();
  assert_eq!(s.shape(), vec![4, 2]);
}

#[test]
fn split_method_form_matches_freefn() {
  let k = random::key(7).unwrap();
  let _ = k.split_key().unwrap();
  let s = k.split_key_num(3).unwrap();
  assert_eq!(s.shape(), vec![3, 2]);
}

#[test]
fn seed_succeeds() {
  random::seed(99).unwrap();
}

#[test]
fn uniform_draws_inside_range() {
  let key = random::key(0).unwrap();
  let low = Array::from_slice::<f32>(&[0.0], &[0i32; 0]).unwrap();
  let high = Array::from_slice::<f32>(&[1.0], &[0i32; 0]).unwrap();
  let mut u = random::uniform(&low, &high, &(64usize,), Dtype::F32, &key).unwrap();
  assert_eq!(u.shape(), vec![64]);
  let v = u.to_vec::<f32>().unwrap();
  for x in v {
    assert!((0.0..1.0).contains(&x), "uniform out of [0,1): {x}");
  }
}

#[test]
fn normal_draws_have_correct_shape_and_dtype() {
  let key = random::key(1).unwrap();
  let mut n = random::normal(&(8usize, 8), Dtype::F32, 0.0, 1.0, &key).unwrap();
  assert_eq!(n.shape(), vec![8, 8]);
  assert_eq!(n.dtype().unwrap(), Dtype::F32);
  // Force eval — sanity check that materialization works for normal output.
  n.eval().unwrap();
}

#[test]
fn normal_broadcast_uses_array_loc_scale() {
  let key = random::key(2).unwrap();
  let loc = Array::from_slice::<f32>(&[0.0], &[0i32; 0]).unwrap();
  let scale = Array::from_slice::<f32>(&[1.0], &[0i32; 0]).unwrap();
  let n = random::normal_broadcast(&(4usize,), Dtype::F32, &loc, &scale, &key).unwrap();
  assert_eq!(n.shape(), vec![4]);
}

#[test]
fn randint_draws_inside_range() {
  let key = random::key(3).unwrap();
  let low = Array::from_slice::<i32>(&[0], &[0i32; 0]).unwrap();
  let high = Array::from_slice::<i32>(&[10], &[0i32; 0]).unwrap();
  let mut r = random::randint(&low, &high, &(32usize,), Dtype::I32, &key).unwrap();
  assert_eq!(r.shape(), vec![32]);
  let v = r.to_vec::<i32>().unwrap();
  for x in v {
    assert!((0..10).contains(&x), "randint out of [0,10): {x}");
  }
}

#[test]
fn bernoulli_draws_bool() {
  let key = random::key(4).unwrap();
  let p = Array::from_slice::<f32>(&[0.5], &[0i32; 0]).unwrap();
  let b = random::bernoulli(&p, &(8usize,), &key).unwrap();
  assert_eq!(b.shape(), vec![8]);
  assert_eq!(b.dtype().unwrap(), Dtype::Bool);
}

#[test]
fn categorical_yields_u32_indices() {
  let key = random::key(5).unwrap();
  let logits = Array::from_slice::<f32>(&[0.1_f32, 0.2, 0.3, 0.4], &(1, 4)).unwrap();
  let c = random::categorical(&logits, -1, &key).unwrap();
  assert_eq!(c.shape(), vec![1]);
  assert_eq!(c.dtype().unwrap(), Dtype::U32);
}

#[test]
fn categorical_num_samples_yields_multiple_indices() {
  let key = random::key(5).unwrap();
  let logits = Array::from_slice::<f32>(&[0.1_f32, 0.2, 0.3, 0.4], &(1, 4)).unwrap();
  let c = random::categorical_num_samples(&logits, -1, 5, &key).unwrap();
  assert_eq!(c.shape(), vec![1, 5]);
  assert_eq!(c.dtype().unwrap(), Dtype::U32);
}

#[test]
fn categorical_shape_yields_explicit_shape() {
  let key = random::key(6).unwrap();
  let logits = Array::from_slice::<f32>(&[0.1_f32, 0.2, 0.3, 0.4], &(1, 4)).unwrap();
  let c = random::categorical_shape(&logits, -1, &(2usize, 3), &key).unwrap();
  assert_eq!(c.shape(), vec![2, 3]);
  assert_eq!(c.dtype().unwrap(), Dtype::U32);
}

#[test]
fn gumbel_draws_have_correct_shape() {
  let key = random::key(7).unwrap();
  let g = random::gumbel(&(16usize,), Dtype::F32, &key).unwrap();
  assert_eq!(g.shape(), vec![16]);
  assert_eq!(g.dtype().unwrap(), Dtype::F32);
}

#[test]
fn truncated_normal_inside_bounds() {
  let key = random::key(8).unwrap();
  let lower = Array::from_slice::<f32>(&[-1.0], &[0i32; 0]).unwrap();
  let upper = Array::from_slice::<f32>(&[1.0], &[0i32; 0]).unwrap();
  let mut t = random::truncated_normal(&lower, &upper, &(32usize,), Dtype::F32, &key).unwrap();
  let v = t.to_vec::<f32>().unwrap();
  for x in v {
    assert!(
      (-1.0..=1.0).contains(&x),
      "truncated_normal out of bounds: {x}"
    );
  }
}

#[test]
fn multivariate_normal_yields_correct_event_dim() {
  let key = random::key(9).unwrap();
  // mean: shape [2], cov: shape [2,2] identity.
  let mean = Array::zeros::<f32>(&[2i32]).unwrap();
  let cov = Array::eye::<f32>(2, None, 0).unwrap();
  let r = random::multivariate_normal(&mean, &cov, &(8usize,), Dtype::F32, &key).unwrap();
  // Output shape is (8, 2) — sample shape with event dim appended.
  assert_eq!(r.shape(), vec![8, 2]);
}

#[test]
fn multivariate_normal_rejects_empty_covariance() {
  // mlx implements `multivariate_normal` via `linalg::svd(cov, ...)` and only
  // checks `cov.ndim() < 2` / squareness — a `0×0` (or `0×n` / `m×0`) cov would
  // reach the SVD kernel's `a.size() / (m * n)` (`0 / 0`, UB / SIGFPE). The
  // shared SVD-input guard rejects it with `Error::EmptyInput` via a cheap shape
  // check, so the call returns `Err` WITHOUT entering mlx (no `eval`).
  let key = random::key(9).unwrap();
  let mean = Array::from_slice::<f32>(&[], &[0i32]).unwrap();
  for dims in [[0i32, 0], [0, 3], [3, 0]] {
    let cov = Array::from_slice::<f32>(&[], &dims).unwrap();
    match random::multivariate_normal(&mean, &cov, &(8usize,), Dtype::F32, &key) {
      Err(mlxrs::Error::EmptyInput(p)) => assert_eq!(
        p.context(),
        "multivariate_normal: covariance matrix has a zero-length row or column dimension"
      ),
      other => panic!("expected EmptyInput for cov {dims:?}, got {other:?}"),
    }
  }
}

#[test]
fn laplace_draws_have_correct_shape() {
  let key = random::key(10).unwrap();
  let l = random::laplace(&(8usize,), Dtype::F32, 0.0, 1.0, &key).unwrap();
  assert_eq!(l.shape(), vec![8]);
}

#[test]
fn bits_yields_uint_array() {
  let key = random::key(11).unwrap();
  let b = random::bits(&(4usize,), 4, &key).unwrap();
  assert_eq!(b.shape(), vec![4]);
}

#[test]
fn permutation_arange_yields_n_unique_indices() {
  let key = random::key(12).unwrap();
  let mut p = random::permutation_arange(8, &key).unwrap();
  assert_eq!(p.shape(), vec![8]);
  // Output dtype is U32 (mlx index-output convention).
  let mut v = p.to_vec::<u32>().unwrap();
  v.sort();
  assert_eq!(v, (0..8u32).collect::<Vec<u32>>());
}

#[test]
fn permutation_method_form_works() {
  let key = random::key(13).unwrap();
  let a = Array::arange::<f32>(0.0, 5.0, 1.0).unwrap();
  let p = a.permutation(0, &key).unwrap();
  assert_eq!(p.shape(), vec![5]);
}