aprender-core 0.31.2

Next-generation machine learning library in pure Rust
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

// CONTRACT: rope-kernel-v1.yaml
// HASH: sha256:f6e4c3b2a1d85970
// Generated by: pv probar --binding
// DO NOT EDIT — regenerate with `pv probar --binding`

use aprender::autograd::Tensor;
use aprender::nn::RotaryPositionEmbedding;
use proptest::prelude::*;

proptest! {
    /// Obligation: Norm preservation (invariant)
    /// Formal: |norm(RoPE(x, m)) - norm(x)| < eps
    #[test]
    fn prop_norm_preservation(
        data in proptest::collection::vec(-10.0f32..10.0, 4..5usize),
        pos in 0usize..32
    ) {
        // head_dim=4, batch=1, seq=1, heads=1
        let head_dim = 4;
        let rope = RotaryPositionEmbedding::new(head_dim, 64);
        let x = Tensor::new(&data[..head_dim], &[1, 1, 1, head_dim]);

        let y = rope.apply(&x, &[pos]);

        // Compute norms
        let x_data = x.data();
        let y_data = y.data();
        let x_norm: f32 = x_data.iter().map(|v| v * v).sum::<f32>().sqrt();
        let y_norm: f32 = y_data.iter().map(|v| v * v).sum::<f32>().sqrt();

        let diff = (x_norm - y_norm).abs();
        prop_assert!(
            diff < 1e-4,
            "norm not preserved: |{x_norm} - {y_norm}| = {diff}"
        );
    }

    /// Obligation: Relative position encoding (invariant)
    /// Formal: dot(RoPE(q,m), RoPE(k,n)) = f(q, k, m-n)
    #[test]
    fn prop_relative_position_encoding(
        q_data in proptest::collection::vec(-5.0f32..5.0, 4..5usize),
        k_data in proptest::collection::vec(-5.0f32..5.0, 4..5usize),
        m in 0usize..16,
        n in 0usize..16
    ) {
        let head_dim = 4;
        let max_seq = 64;
        let rope = RotaryPositionEmbedding::new(head_dim, max_seq);

        let q = Tensor::new(&q_data[..head_dim], &[1, 1, 1, head_dim]);
        let k = Tensor::new(&k_data[..head_dim], &[1, 1, 1, head_dim]);

        // dot(RoPE(q, m), RoPE(k, n))
        let q_rot = rope.apply(&q, &[m]);
        let k_rot = rope.apply(&k, &[n]);
        let dot1: f32 = q_rot.data().iter()
            .zip(k_rot.data().iter())
            .map(|(a, b)| a * b)
            .sum();

        // dot(RoPE(q, 0), RoPE(k, n-m)) if n >= m
        // dot(RoPE(q, m-n), RoPE(k, 0)) if m >= n
        let (pos_a, pos_b) = if n >= m { (0, n - m) } else { (m - n, 0) };
        let q_rot2 = rope.apply(&q, &[pos_a]);
        let k_rot2 = rope.apply(&k, &[pos_b]);
        let dot2: f32 = q_rot2.data().iter()
            .zip(k_rot2.data().iter())
            .map(|(a, b)| a * b)
            .sum();

        let diff = (dot1 - dot2).abs();
        prop_assert!(
            diff < 1e-3,
            "relative position: dot1={dot1}, dot2={dot2}, diff={diff}"
        );
    }

    /// Obligation: SIMD matches scalar (equivalence)
    #[test]
    #[ignore = "SIMD equivalence — trueno domain"]
    fn prop_simd_matches_scalar(
        _x in proptest::collection::vec(-5.0f32..5.0, 1..32usize)
    ) {
        // SIMD equivalence testing is trueno's responsibility
    }

    /// Obligation: Output bounded by input norm (bound)
    /// Formal: norm(RoPE(x, m)) <= norm(x) + eps
    #[test]
    fn prop_output_bounded_by_input_norm(
        data in proptest::collection::vec(-10.0f32..10.0, 4..5usize),
        pos in 0usize..32
    ) {
        let head_dim = 4;
        let rope = RotaryPositionEmbedding::new(head_dim, 64);
        let x = Tensor::new(&data[..head_dim], &[1, 1, 1, head_dim]);

        let y = rope.apply(&x, &[pos]);

        let x_norm: f32 = x.data().iter().map(|v| v * v).sum::<f32>().sqrt();
        let y_norm: f32 = y.data().iter().map(|v| v * v).sum::<f32>().sqrt();

        prop_assert!(
            y_norm <= x_norm + 1e-4,
            "output norm {y_norm} exceeds input norm {x_norm}"
        );
    }
}

/// RoPE at position 0 should be identity (cos(0)=1, sin(0)=0)
#[test]
fn rope_position_zero_is_identity() {
    let head_dim = 4;
    let rope = RotaryPositionEmbedding::new(head_dim, 64);
    let data = vec![1.0, 2.0, 3.0, 4.0f32];
    let x = Tensor::new(&data, &[1, 1, 1, head_dim]);

    let y = rope.apply(&x, &[0]);
    let y_data = y.data();

    for i in 0..head_dim {
        let diff = (data[i] - y_data[i]).abs();
        assert!(
            diff < 1e-5,
            "RoPE(x, 0)[{i}]={}, expected {}",
            y_data[i],
            data[i]
        );
    }
}