mlx-native 0.1.1

Pure-Rust Metal GPU compute library for MLX-compatible inference on Apple Silicon
Documentation 
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
259
260
261
262
263
264
265
266
267
268
269
270
271
272
273
274
275
276
277
278
279
280

//! Tests for elementwise add, multiply, cast, and transpose GPU kernels.

#![allow(clippy::expect_used, clippy::unwrap_used, clippy::panic)]

use mlx_native::{DType, KernelRegistry, MlxDevice};

fn setup() -> (MlxDevice, KernelRegistry) {
    let device = MlxDevice::new().expect("MlxDevice::new");
    let registry = KernelRegistry::new();
    (device, registry)
}

// ---- elementwise_add f32 ----

#[test]
fn test_elementwise_add_f32_basic() {
    let (device, mut registry) = setup();

    let a_data: Vec<f32> = vec![1.0, 2.0, 3.0, 4.0, -1.0, 0.0, 0.5, 100.0];
    let b_data: Vec<f32> = vec![10.0, 20.0, 30.0, 40.0, 1.0, 0.0, -0.5, -100.0];
    let n = a_data.len();
    let byte_len = n * std::mem::size_of::<f32>();

    let mut a_buf = device.alloc_buffer(byte_len, DType::F32, vec![n]).expect("alloc a");
    let mut b_buf = device.alloc_buffer(byte_len, DType::F32, vec![n]).expect("alloc b");
    let out_buf = device.alloc_buffer(byte_len, DType::F32, vec![n]).expect("alloc out");

    a_buf.as_mut_slice::<f32>().expect("mut a").copy_from_slice(&a_data);
    b_buf.as_mut_slice::<f32>().expect("mut b").copy_from_slice(&b_data);

    let mut encoder = device.command_encoder().expect("encoder");
    mlx_native::ops::elementwise::elementwise_add(
        &mut encoder,
        &mut registry,
        device.metal_device(),
        &a_buf,
        &b_buf,
        &out_buf,
        n,
        DType::F32,
    )
    .expect("elementwise_add");
    encoder.commit_and_wait().expect("commit");

    let output: &[f32] = out_buf.as_slice().expect("read");
    for i in 0..n {
        let expected = a_data[i] + b_data[i];
        let diff = (output[i] - expected).abs();
        assert!(
            diff < 1e-6,
            "add f32 mismatch at {}: expected {}, got {}, diff {}",
            i, expected, output[i], diff
        );
    }
}

#[test]
fn test_elementwise_add_f32_zeros() {
    let (device, mut registry) = setup();

    let a_data: Vec<f32> = vec![0.0; 16];
    let b_data: Vec<f32> = vec![0.0; 16];
    let n = 16;
    let byte_len = n * 4;

    let mut a_buf = device.alloc_buffer(byte_len, DType::F32, vec![n]).expect("a");
    let mut b_buf = device.alloc_buffer(byte_len, DType::F32, vec![n]).expect("b");
    let out_buf = device.alloc_buffer(byte_len, DType::F32, vec![n]).expect("out");

    a_buf.as_mut_slice::<f32>().expect("a").copy_from_slice(&a_data);
    b_buf.as_mut_slice::<f32>().expect("b").copy_from_slice(&b_data);

    let mut encoder = device.command_encoder().expect("enc");
    mlx_native::ops::elementwise::elementwise_add(
        &mut encoder, &mut registry, device.metal_device(),
        &a_buf, &b_buf, &out_buf, n, DType::F32,
    ).expect("add");
    encoder.commit_and_wait().expect("commit");

    let output: &[f32] = out_buf.as_slice().expect("read");
    for &val in output {
        assert!(val.abs() < 1e-7, "expected 0, got {}", val);
    }
}

// ---- elementwise_mul f32 ----

#[test]
fn test_elementwise_mul_f32() {
    let (device, mut registry) = setup();

    let a_data: Vec<f32> = vec![2.0, 3.0, -1.0, 0.5];
    let b_data: Vec<f32> = vec![4.0, -2.0, 5.0, 10.0];
    let n = 4;
    let byte_len = n * 4;

    let mut a_buf = device.alloc_buffer(byte_len, DType::F32, vec![n]).expect("a");
    let mut b_buf = device.alloc_buffer(byte_len, DType::F32, vec![n]).expect("b");
    let out_buf = device.alloc_buffer(byte_len, DType::F32, vec![n]).expect("out");

    a_buf.as_mut_slice::<f32>().expect("a").copy_from_slice(&a_data);
    b_buf.as_mut_slice::<f32>().expect("b").copy_from_slice(&b_data);

    let mut encoder = device.command_encoder().expect("enc");
    mlx_native::ops::elementwise::elementwise_mul(
        &mut encoder, &mut registry, device.metal_device(),
        &a_buf, &b_buf, &out_buf, n, DType::F32,
    ).expect("mul");
    encoder.commit_and_wait().expect("commit");

    let output: &[f32] = out_buf.as_slice().expect("read");
    let expected: Vec<f32> = vec![8.0, -6.0, -5.0, 5.0];
    for i in 0..n {
        let diff = (output[i] - expected[i]).abs();
        assert!(diff < 1e-6, "mul mismatch at {}: expected {}, got {}", i, expected[i], output[i]);
    }
}

// ---- cast f32 <-> f16 ----

#[test]
fn test_cast_f32_to_f16_and_back() {
    let (device, mut registry) = setup();

    let data: Vec<f32> = vec![0.0, 1.0, -1.0, 0.5, 100.0, -100.0, 0.001, 65504.0];
    let n = data.len();

    let mut f32_buf = device.alloc_buffer(n * 4, DType::F32, vec![n]).expect("f32");
    let f16_buf = device.alloc_buffer(n * 2, DType::F16, vec![n]).expect("f16");
    let f32_back = device.alloc_buffer(n * 4, DType::F32, vec![n]).expect("f32 back");

    f32_buf.as_mut_slice::<f32>().expect("write").copy_from_slice(&data);

    // f32 -> f16
    let mut encoder = device.command_encoder().expect("enc");
    mlx_native::ops::elementwise::cast(
        &mut encoder, &mut registry, device.metal_device(),
        &f32_buf, &f16_buf, n,
        mlx_native::ops::elementwise::CastDirection::F32ToF16,
    ).expect("cast f32->f16");
    encoder.commit_and_wait().expect("commit");

    // f16 -> f32
    let mut encoder2 = device.command_encoder().expect("enc2");
    mlx_native::ops::elementwise::cast(
        &mut encoder2, &mut registry, device.metal_device(),
        &f16_buf, &f32_back, n,
        mlx_native::ops::elementwise::CastDirection::F16ToF32,
    ).expect("cast f16->f32");
    encoder2.commit_and_wait().expect("commit2");

    let output: &[f32] = f32_back.as_slice().expect("read");
    for i in 0..n {
        // f16 has limited precision, so we allow some tolerance
        let diff = (output[i] - data[i]).abs();
        let rel_tol = data[i].abs() * 1e-3 + 1e-4;
        assert!(
            diff <= rel_tol,
            "cast roundtrip mismatch at {}: original {}, got {}, diff {}",
            i, data[i], output[i], diff
        );
    }
}

// ---- transpose 2d f32 ----

#[test]
fn test_transpose_2d_f32() {
    let (device, mut registry) = setup();

    // 3x4 matrix
    let rows = 3;
    let cols = 4;
    let data: Vec<f32> = (0..12).map(|i| i as f32).collect();
    let n = rows * cols;
    let byte_len = n * 4;

    let mut input = device.alloc_buffer(byte_len, DType::F32, vec![rows, cols]).expect("in");
    let output = device.alloc_buffer(byte_len, DType::F32, vec![cols, rows]).expect("out");

    input.as_mut_slice::<f32>().expect("write").copy_from_slice(&data);

    let mut encoder = device.command_encoder().expect("enc");
    mlx_native::ops::transpose::transpose_2d(
        &mut encoder, &mut registry, device.metal_device(),
        &input, &output, rows, cols, DType::F32,
    ).expect("transpose");
    encoder.commit_and_wait().expect("commit");

    let result: &[f32] = output.as_slice().expect("read");

    // Verify: output[col][row] == input[row][col]
    for r in 0..rows {
        for c in 0..cols {
            let input_val = data[r * cols + c];
            let output_val = result[c * rows + r];
            assert!(
                (input_val - output_val).abs() < 1e-7,
                "transpose mismatch: input[{},{}]={} != output[{},{}]={}",
                r, c, input_val, c, r, output_val
            );
        }
    }
}

#[test]
fn test_transpose_2d_f32_square() {
    let (device, mut registry) = setup();

    let n = 4;
    let data: Vec<f32> = (0..16).map(|i| i as f32 * 0.5).collect();
    let byte_len = 16 * 4;

    let mut input = device.alloc_buffer(byte_len, DType::F32, vec![n, n]).expect("in");
    let output = device.alloc_buffer(byte_len, DType::F32, vec![n, n]).expect("out");

    input.as_mut_slice::<f32>().expect("write").copy_from_slice(&data);

    let mut encoder = device.command_encoder().expect("enc");
    mlx_native::ops::transpose::transpose_2d(
        &mut encoder, &mut registry, device.metal_device(),
        &input, &output, n, n, DType::F32,
    ).expect("transpose");
    encoder.commit_and_wait().expect("commit");

    let result: &[f32] = output.as_slice().expect("read");
    for r in 0..n {
        for c in 0..n {
            let expected = data[r * n + c];
            let actual = result[c * n + r];
            assert!(
                (expected - actual).abs() < 1e-7,
                "square transpose mismatch at [{},{}]",
                r, c
            );
        }
    }
}

// ---- Validation tests ----

#[test]
fn test_elementwise_add_zero_elements_error() {
    let (device, mut registry) = setup();

    let buf = device.alloc_buffer(16, DType::F32, vec![4]).expect("buf");
    let mut encoder = device.command_encoder().expect("enc");
    let result = mlx_native::ops::elementwise::elementwise_add(
        &mut encoder, &mut registry, device.metal_device(),
        &buf, &buf, &buf, 0, DType::F32,
    );
    assert!(result.is_err(), "zero elements should error");
}

#[test]
fn test_transpose_zero_rows_error() {
    let (device, mut registry) = setup();

    let buf = device.alloc_buffer(16, DType::F32, vec![4]).expect("buf");
    let mut encoder = device.command_encoder().expect("enc");
    let result = mlx_native::ops::transpose::transpose_2d(
        &mut encoder, &mut registry, device.metal_device(),
        &buf, &buf, 0, 4, DType::F32,
    );
    assert!(result.is_err(), "zero rows should error");
}

#[test]
fn test_cast_zero_elements_error() {
    let (device, mut registry) = setup();

    let buf = device.alloc_buffer(16, DType::F32, vec![4]).expect("buf");
    let mut encoder = device.command_encoder().expect("enc");
    let result = mlx_native::ops::elementwise::cast(
        &mut encoder, &mut registry, device.metal_device(),
        &buf, &buf, 0,
        mlx_native::ops::elementwise::CastDirection::F32ToF16,
    );
    assert!(result.is_err(), "zero elements should error");
}