zyx 0.16.0

Zyx machine learning library
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
281
282
283
284
285
286
287
288
289
290
291
292
293
294
295
296

// Copyright (C) 2025 zk4x
// SPDX-License-Identifier: LGPL-3.0-only

use zyx::{DType, ReduceOp, Scalar, Tape, Tensor, ZyxError};

#[test]
fn sin() -> Result<(), ZyxError> {
    let data: [f32; 10] = [-3.285, 0.001, 1.780, 5.675, -8.521, -0.456, 1.215, -3.474, -4.128, -7.657];
    let x = Tensor::from(data);
    let tape = Tape::new([&x])?;
    let z = x.sin();
    tape.realize([&z])?;
    let zdata: Vec<f32> = z.try_into()?;
    for (x, y) in data.iter().zip(zdata) {
        assert!(x.sin().is_equal(y), "{} != {y}", x.sin());
    }
    Ok(())
}

#[test]
fn relu() -> Result<(), ZyxError> {
    let data: [f32; 10] = [-3.285, 0.001, 1.780, 5.675, -8.521, -0.456, 1.215, -3.474, -4.128, -7.657];
    let x = Tensor::from(data);
    let tape = Tape::new([&x])?;
    let z = x.relu();
    tape.realize([&z])?;
    assert_eq!(z, [0.0f32, 0.001, 1.780, 5.675, 0.0, 0.0, 1.215, 0.0, 0.0, 0.0]);
    Ok(())
}

#[test]
fn matmul() -> Result<(), ZyxError> {
    let x = Tensor::from([[2, 4, 3], [1, 5, 1]]);
    let y = Tensor::from([[2, 4], [3, 1], [5, 1]]);
    let tape = Tape::new([&x, &y])?;
    let z = x.dot(y)?;
    tape.realize([&z])?;
    assert_eq!(z, [[31, 15], [22, 10]]);
    Ok(())
}

#[test]
fn matmul_f32() -> Result<(), ZyxError> {
    let x = Tensor::from([[2f32, 4., 3.], [1., 5., 1.]]);
    let y = Tensor::from([[2f32, 4.], [3., 1.], [5., 1.]]);
    let tape = Tape::new([&x, &y])?;
    let z = x.dot(y)?;
    tape.realize([&z])?;
    assert_eq!(z, [[31f32, 15.], [22., 10.]]);
    Ok(())
}

#[test]
fn softmax() -> Result<(), ZyxError> {
    let x = Tensor::from([2f32, 4., 3.]);
    let tape = Tape::new([&x])?;
    let y = x.softmax([])?;
    tape.realize([&y])?;
    assert_eq!(y, [0.09003056585788726807f32, 0.66524088382720947266, 0.24472846090793609619]);
    Ok(())
}

#[test]
fn causal_self_attention() -> Result<(), ZyxError> {
    let dtype = DType::F32;
    let n_embd = 4;
    let n_head = 4;
    let c_attn_weight = Tensor::from([
        [3, 1, 2, 3, 1, 2, 5, 4, 2, 3, 1, 3],
        [1, 1, 2, 3, 1, 2, 5, 4, 2, 3, 1, 3],
        [3, 1, 5, 3, 1, 2, 5, 4, 2, 3, 1, 3],
        [3, 1, 2, 3, 1, 2, 5, 8, 2, 3, 1, 3],
    ])
    .t()
    .cast(dtype);

    let x = Tensor::from([[[1, 0, 4, 2], [2, 5, 0, 1], [0, 8, 1, 0], [5, 1, 0, 0]]]).cast(dtype);

    let [b, t, c] = x.shape()[..] else {
        return Err(ZyxError::ShapeError("x must have exactly 3 dims, b, t, c".into()));
    };

    let tape = Tape::new([&x, &c_attn_weight])?;
    let mut splits = x.dot(c_attn_weight.t())?.split([n_embd, n_embd, n_embd], 2)?;
    let mut v = splits.pop().unwrap();
    let mut k = splits.pop().unwrap();
    let mut q = splits.pop().unwrap();

    k = k.reshape([b, t, n_head, c / n_head])?.transpose(1, 2)?;
    q = q.reshape([b, t, n_head, c / n_head])?.transpose(1, 2)?;
    v = v.reshape([b, t, n_head, c / n_head])?.transpose(1, 2)?;

    let mut att = q.dot(k.t())? * (1f32 / (*k.shape().last().unwrap() as f32).sqrt());
    att = att.softmax([-1])?;
    let mut y = att.dot(v)?;
    y = y.transpose(1, 2)?.reshape([b, t, c])?;

    tape.realize([&y])?;

    assert_eq!(
        y,
        [[
            [18f32, 27., 9., 24.],
            [18., 27., 9., 24.],
            [18., 27., 9., 24.],
            [18., 27., 9., 24.]
        ]]
    );

    Ok(())
}

// Reproducer for bug in promote_to_graph: promoting an eager tensor to Graph
// leaves a stale store entry in the original kernel. When gradient needs the
// value, materialize_kernel finds the store target in Graph state → panic.
// Gather through the full pipeline (kernelizer + fold_loops). The kernelizer
// fuses the arange produced by one_hot_along_dim into constants, making the
// mask's loop operand analyzable so the gather loop can fold to a direct gather.
#[test]
fn gather() -> Result<(), ZyxError> {
    let x = Tensor::from([10, 20, 30, 40, 50]);
    let indices = Tensor::from([0u32, 2, 4, 1]);
    let tape = Tape::new([&x, &indices])?;
    let gathered = x.gather(0, &indices)?;
    tape.realize([&gathered])?;
    assert_eq!(gathered, [10, 30, 50, 20]);
    Ok(())
}

#[test]
fn promote_and_gradient() -> Result<(), ZyxError> {
    let gt = Tensor::randn([2, 4, 1, 1], DType::F32)?;
    let tape = Tape::new([&gt])?;

    let a = Tensor::ones([4], DType::F32);
    let shape = [2, 4, 1, 1];
    let b = a.reshape([1, 4, 1, 1])?;
    let c = b.expand(shape)?;
    let d = &c + 1e-5f32;
    let e = d.rsqrt();

    let result = &e + &gt;
    let _g = tape.gradient(&result, [&gt]);
    Ok(())
}

// Reproducer for orphan kernel outputs: forward+backward through a 2-layer
// net with weight transpose, cross-entropy loss, and SGD update; then realize
// all params. This pattern triggers fill_remaining's force-seal bug where
// orphan kernel outputs not in the output_set prevent kernel sealing.
#[test]
fn small_net() -> Result<(), ZyxError> {
    let w1 = Tensor::randn([3, 4], DType::F32)?;
    let b1 = Tensor::randn([3], DType::F32)?;
    let w2 = Tensor::randn([2, 3], DType::F32)?;
    let b2 = Tensor::randn([2], DType::F32)?;

    for _ in 0..3 {
        let tape = Tape::new([&w1, &b1, &w2, &b2])?;
        let x = Tensor::randn([2, 4], DType::F32)?;
        let y = Tensor::from([0u32, 1]);
        let h = (x.dot(&w1.t())? + &b1).relu();
        let logits = h.dot(&w2.t())? + &b2;
        let loss = logits.cross_entropy(y, ReduceOp::Mean)?;
        let grads = tape.gradient(&loss, [&w1, &b1, &w2, &b2]);

        let lr = 0.01f32;
        let new_w1 = &w1 - &grads[0] * lr;
        let new_b1 = &b1 - &grads[1] * lr;
        let new_w2 = &w2 - &grads[2] * lr;
        let new_b2 = &b2 - &grads[3] * lr;

        tape.realize([&new_w1, &new_b1, &new_w2, &new_b2])?;
    }
    Ok(())
}

#[test]
fn tape_caching() -> Result<(), ZyxError> {
    let x = Tensor::from([1.0f32, 2.0, 3.0]);
    let y = Tensor::randn([3, 3], DType::F32)?;
    for _ in 0..10 {
        let tape = Tape::new([&x])?;
        let z = &x + y.t().sin();
        tape.realize([&z])?;
    }
    Ok(())
}

#[test]
fn tape_matmul() -> Result<(), ZyxError> {
    let x = Tensor::from([[1.0f32, 2.0], [3.0, 4.0], [5.0, 6.0]]);
    let w = Tensor::from([0.5f32, 1.5, 2.5, 3.5]).reshape([2, 2])?.relu();
    for _ in 0..3 {
        let tape = Tape::new([&x])?;
        let z = x.dot(&w)?.relu();
        tape.realize([&z])?;
    }
    Ok(())
}

#[test]
fn big_matmul() -> Result<(), ZyxError> {
    let x = Tensor::rand([256, 392], DType::F32)?;
    let w = Tensor::rand([392, 296], DType::F32)?;
    let tape = Tape::new([&x, &w])?;
    let z = x.dot(&w)?.relu();
    tape.realize([&z])?;
    let shape = z.shape();
    assert_eq!(shape, [256, 296]);
    Ok(())
}

#[test]
fn drop_without_realize_params_eager() -> Result<(), ZyxError> {
    let x = Tensor::from([1.0f32, 2.0, 3.0]);
    {
        let _tape = Tape::new([&x])?;
        let _z = x.sin();
    }
    let y = x + 1.0f32;
    let data: Vec<f32> = y.try_into()?;
    assert_eq!(data, [2.0f32, 3.0, 4.0]);
    Ok(())
}

#[test]
#[should_panic(expected = "tape scope has ended")]
fn use_intermediate_after_drop_panics() {
    let x = Tensor::from([1.0f32, 2.0, 3.0]);
    let z;
    {
        let _tape = Tape::new([&x]).unwrap();
        z = x.sin();
    }
    let _ = z + 1.0f32;
}

#[test]
fn realize_outputs_eager_leaves_eager_after_drop() -> Result<(), ZyxError> {
    let x = Tensor::from([1.0f32, 2.0, 3.0]);
    let z;
    {
        let tape = Tape::new([&x])?;
        z = x.sin();
        tape.realize([&z])?;
    }
    let zdata: Vec<f32> = z.try_into()?;
    for (a, b) in [1.0f32, 2.0, 3.0].iter().zip(zdata) {
        assert!(a.sin().is_equal(b));
    }
    let y = x + 1.0f32;
    let data: Vec<f32> = y.try_into()?;
    assert_eq!(data, [2.0f32, 3.0, 4.0]);
    Ok(())
}

#[test]
fn realized_tensor_promotes_as_leaf() -> Result<(), ZyxError> {
    let x = Tensor::from([1.0f32, 2.0, 3.0]);
    let tape = Tape::new([&x])?;
    let z = x.relu();
    tape.realize([&z])?;
    let data: Vec<f32> = z.try_into()?;
    assert_eq!(data, [1.0f32, 2.0, 3.0]);
    Ok(())
}

/*#[test]
fn frozen_tape_replay() -> Result<(), ZyxError> {
    let x = Tensor::from([1.0f32, 2.0, 3.0]);
    let frozen = {
        let tape = Tape::new([&x])?;
        let z = x.relu();
        tape.freeze([&z])?
    };
    let zs = frozen.replay([&x])?;
    assert_eq!(zs.len(), 1);
    let z = zs.into_iter().next().unwrap();
    let data: Vec<f32> = z.try_into()?;
    assert_eq!(data, [1.0f32, 2.0, 3.0]);
    Ok(())
}*/

#[test]
#[should_panic(expected = "tape scope has ended")]
fn use_frozen_output_panics() {
    let x = Tensor::from([1.0f32, 2.0, 3.0]);
    let z;
    {
        let tape = Tape::new([&x]).unwrap();
        z = x.relu();
        let _frozen = tape.freeze([&z]).unwrap();
    }
    let _ = z + 1.0f32;
}