aprender-serve 0.64.0

Pure Rust ML inference engine built from scratch - model serving for GGUF and safetensors
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


#[test]
fn test_fused_q4_0_q8_0_dot_scalar_multiple_blocks() {
    // 2 blocks = 64 elements
    let q4_data = vec![0u8; 36]; // 2 * 18 bytes
    let q8_scales = vec![1.0f32; 2];
    let q8_quants = vec![1i8; 64];

    let result = fused_q4_0_q8_0_dot_scalar(&q4_data, &q8_scales, &q8_quants, 64);
    assert!(result.is_finite());
}

// =============================================================================
// fused_q8_0_q8_0_dot_scalar tests
// =============================================================================

#[test]
fn test_fused_q8_0_q8_0_dot_scalar_zero() {
    // Q8_0 block: 2 bytes f16 scale + 32 bytes quants = 34 bytes
    let q8_weight = vec![0u8; 34];
    let q8_scales = vec![0.0f32];
    let q8_quants = vec![0i8; 32];

    let result = fused_q8_0_q8_0_dot_scalar(&q8_weight, &q8_scales, &q8_quants, 32);
    assert_eq!(result, 0.0);
}

#[test]
fn test_fused_q8_0_q8_0_dot_scalar_basic() {
    // Create Q8_0 weight with scale = 1.0
    let mut q8_weight = vec![0u8; 34];
    q8_weight[0] = 0x00;
    q8_weight[1] = 0x3C; // f16 1.0

    // Set quantized values
    for i in 2..34 {
        q8_weight[i] = 1; // i8 value 1
    }

    let q8_scales = vec![1.0f32];
    let q8_quants = vec![1i8; 32];

    let result = fused_q8_0_q8_0_dot_scalar(&q8_weight, &q8_scales, &q8_quants, 32);
    assert!(result.is_finite());
    // Should be 32 * 1 * 1 * 1 * 1 = 32
    assert!((result - 32.0).abs() < 1.0);
}

#[test]
fn test_fused_q8_0_q8_0_dot_scalar_negative() {
    let mut q8_weight = vec![0u8; 34];
    q8_weight[0] = 0x00;
    q8_weight[1] = 0x3C; // f16 1.0

    // Set quantized values to -1 (0xFF as i8)
    for i in 2..34 {
        q8_weight[i] = 0xFF; // -1 as i8
    }

    let q8_scales = vec![1.0f32];
    let q8_quants = vec![1i8; 32];

    let result = fused_q8_0_q8_0_dot_scalar(&q8_weight, &q8_scales, &q8_quants, 32);
    assert!(result.is_finite());
    // Should be 32 * (-1) * 1 * 1 * 1 = -32
    assert!((result - (-32.0)).abs() < 1.0);
}

#[test]
fn test_fused_q8_0_q8_0_dot_scalar_multiple_blocks() {
    // 2 blocks = 64 elements
    let q8_weight = vec![0u8; 68]; // 2 * 34 bytes
    let q8_scales = vec![1.0f32; 2];
    let q8_quants = vec![1i8; 64];

    let result = fused_q8_0_q8_0_dot_scalar(&q8_weight, &q8_scales, &q8_quants, 64);
    assert!(result.is_finite());
}

// =============================================================================
// Additional edge case tests for high coverage
// =============================================================================

#[test]
fn test_q4_0_dot_scalar_partial_block() {
    // Test with in_dim not multiple of 32 (partial final block)
    let q4_data = vec![0u8; 36]; // 2 blocks worth
    let q8_scales = vec![1.0f32; 2];
    let q8_quants = vec![1i8; 50]; // Only 50 elements, not 64

    let result = fused_q4_0_q8_0_dot_scalar(&q4_data, &q8_scales, &q8_quants, 50);
    assert!(result.is_finite());
}

#[test]
fn test_q8_0_dot_scalar_partial_block() {
    let q8_weight = vec![0u8; 68];
    let q8_scales = vec![1.0f32; 2];
    let q8_quants = vec![1i8; 50];

    let result = fused_q8_0_q8_0_dot_scalar(&q8_weight, &q8_scales, &q8_quants, 50);
    assert!(result.is_finite());
}

#[test]
fn test_interleaved_q4k_three_superblocks() {
    let data = vec![0u8; 432]; // 3 super-blocks
    let interleaved = InterleavedQ4K::from_q4k(&data).expect("test value should be present");
    assert_eq!(interleaved.num_super_blocks, 3);
    assert_eq!(interleaved.num_values(), 768);
}

#[test]
fn test_q8k_into_exact_buffer_sizes() {
    let activations = vec![0.5f32; 256];
    let mut scales = vec![0.0f32; 1]; // Exactly 1 needed
    let mut quants = vec![0i8; 256]; // Exactly 256 needed

    let result = quantize_activations_q8k_into(&activations, &mut scales, &mut quants);
    assert!(result.is_ok());
}

#[test]
fn test_extract_scale_min_all_blocks() {
    // Test all 8 blocks to ensure both branches are covered
    let scales: [u8; 12] = [
        0xFF, 0xFF, 0xFF, 0xFF, 0xFF, 0xFF, 0xFF, 0xFF, 0xFF, 0xFF, 0xFF, 0xFF,
    ];

    for i in 0..8 {
        let (scale, min) = extract_scale_min(&scales, i);
        assert!(scale >= 0.0);
        assert!(min >= 0.0);
    }
}