realizar 0.8.5

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
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


#[test]
fn test_simd_bf16_to_f32_exact_8_values() {
    // Test exactly 8 values (one SIMD chunk on AVX2)
    let values: [f32; 8] = [1.0, 2.0, 3.0, 4.0, 5.0, 6.0, 7.0, 8.0];
    let mut bf16_bytes = Vec::with_capacity(16);
    for &v in &values {
        bf16_bytes.extend_from_slice(&half::bf16::from_f32(v).to_le_bytes());
    }

    let result = simd_bf16_to_f32(&bf16_bytes);
    assert_eq!(result.len(), 8);

    for (i, (&expected, &actual)) in values.iter().zip(result.iter()).enumerate() {
        assert!(
            (actual - expected).abs() < 0.01,
            "Index {}: expected {}, got {}",
            i,
            expected,
            actual
        );
    }
}

#[test]
fn test_simd_f16_to_f32_empty() {
    let result = simd_f16_to_f32(&[]);
    assert!(result.is_empty());
}

#[test]
fn test_simd_f16_to_f32_special_values() {
    let zero_bytes = half::f16::from_f32(0.0).to_le_bytes();
    let result = simd_f16_to_f32(&zero_bytes);
    assert!((result[0]).abs() < 1e-10);

    let inf_bytes = half::f16::from_f32(f32::INFINITY).to_le_bytes();
    let result = simd_f16_to_f32(&inf_bytes);
    assert!(result[0].is_infinite());
}

#[test]
fn test_simd_f16_to_f32_subnormal() {
    // Test a very small F16 value (subnormal)
    let small = half::f16::from_f32(1e-7);
    let bytes = small.to_le_bytes();
    let result = simd_f16_to_f32(&bytes);
    assert!(result[0] >= 0.0);
    assert!(result[0] < 1e-3);
}

// ============================================================================
// BF16 Dot Product Edge Cases
// ============================================================================

#[test]
fn test_simd_bf16_dot_empty() {
    let result = simd_bf16_dot(&[], &[]);
    assert!((result).abs() < 1e-5);
}

#[test]
fn test_simd_bf16_dot_single_value() {
    let a_bytes = half::bf16::from_f32(3.0).to_le_bytes().to_vec();
    let b_bytes = half::bf16::from_f32(4.0).to_le_bytes().to_vec();
    let result = simd_bf16_dot(&a_bytes, &b_bytes);
    assert!((result - 12.0).abs() < 0.1);
}

#[test]
fn test_simd_bf16_dot_mismatched_lengths() {
    let mut a_bytes = Vec::new();
    let mut b_bytes = Vec::new();

    // a has 3 values, b has 2 values
    for v in [1.0, 2.0, 3.0] {
        a_bytes.extend_from_slice(&half::bf16::from_f32(v).to_le_bytes());
    }
    for v in [1.0, 1.0] {
        b_bytes.extend_from_slice(&half::bf16::from_f32(v).to_le_bytes());
    }

    // Should use minimum length
    let result = simd_bf16_dot(&a_bytes, &b_bytes);
    // Only first 2 values used: 1*1 + 2*1 = 3
    assert!((result - 3.0).abs() < 0.1);
}

#[test]
fn test_simd_bf16_dot_chunk_boundary() {
    // Test with exactly 64 values (one chunk)
    let n = 64;
    let mut a_bytes = Vec::with_capacity(n * 2);
    let mut b_bytes = Vec::with_capacity(n * 2);

    for i in 0..n {
        let v = ((i % 5) as f32) * 0.2;
        a_bytes.extend_from_slice(&half::bf16::from_f32(v).to_le_bytes());
        b_bytes.extend_from_slice(&half::bf16::from_f32(1.0).to_le_bytes());
    }

    let result = simd_bf16_dot(&a_bytes, &b_bytes);
    // Sum of 0.0, 0.2, 0.4, 0.6, 0.8, 0.0, ... (13 complete cycles = 2.0 each, + 3 values)
    // 64/5 = 12 complete + 4 remainder (0.0, 0.2, 0.4, 0.6 = 1.2)
    let expected = 12.0 * 2.0 + 1.2;
    assert!(
        (result - expected).abs() < 1.0,
        "Expected ~{}, got {}",
        expected,
        result
    );
}

// ============================================================================
// BF16 Matmul Edge Cases
// ============================================================================

#[test]
fn test_simd_bf16_matmul_single_element() {
    let input = vec![2.0];
    let weight_bytes = half::bf16::from_f32(3.0).to_le_bytes().to_vec();

    let output = simd_bf16_matmul(&input, &weight_bytes, 1, 1);
    assert_eq!(output.len(), 1);
    assert!((output[0] - 6.0).abs() < 0.1);
}

#[test]
fn test_simd_bf16_matmul_zeros() {
    let input = vec![0.0; 4];
    let mut weight_bytes = Vec::new();
    for _ in 0..8 {
        weight_bytes.extend_from_slice(&half::bf16::from_f32(1.0).to_le_bytes());
    }

    let output = simd_bf16_matmul(&input, &weight_bytes, 4, 2);
    for &v in &output {
        assert!(v.abs() < 0.01);
    }
}

#[test]
fn test_simd_bf16_matmul_large() {
    let in_dim = 128;
    let out_dim = 64;
    let input: Vec<f32> = (0..in_dim).map(|i| (i % 10) as f32 * 0.1).collect();

    // Create identity-like weight in BF16
    let mut weight_bytes = Vec::with_capacity(out_dim * in_dim * 2);
    for row in 0..out_dim {
        for col in 0..in_dim {
            let val = if row == col { 1.0 } else { 0.0 };
            weight_bytes.extend_from_slice(&half::bf16::from_f32(val).to_le_bytes());
        }
    }

    let output = simd_bf16_matmul(&input, &weight_bytes, in_dim, out_dim);
    assert_eq!(output.len(), out_dim);

    // First out_dim elements should approximately match input
    for i in 0..out_dim {
        let expected = (i % 10) as f32 * 0.1;
        assert!(
            (output[i] - expected).abs() < 0.05,
            "Index {}: expected {}, got {}",
            i,
            expected,
            output[i]
        );
    }
}

// ============================================================================
// Integration: Activation Chains
// ============================================================================

#[test]
fn test_matmul_silu_chain() {
    let input = vec![1.0, 2.0];
    let weight = vec![
        1.0, 1.0, // sum: 3
        1.0, -1.0, // diff: -1
    ];
    let mut output = simd_matmul(&input, &weight, 2, 2);
    simd_silu(&mut output);

    // silu(3) and silu(-1)
    assert!(output[0] > 2.5); // silu(3) > 2.5
    assert!(output[1] < 0.0); // silu(-1) < 0
}

#[test]
fn test_matmul_gelu_softmax_chain() {
    let input = vec![0.5, 0.5, 0.5, 0.5];
    let weight = vec![
        1.0, 1.0, 1.0, 1.0, // sum: 2
        1.0, 0.0, 0.0, 0.0, // just first: 0.5
        0.0, 0.0, 0.0, 1.0, // just last: 0.5
    ];
    let mut output = simd_matmul(&input, &weight, 4, 3);
    simd_gelu(&mut output);
    simd_softmax(&mut output);

    // Output should sum to 1
    let sum: f32 = output.iter().sum();
    assert!((sum - 1.0).abs() < 1e-5);

    // First element (gelu(2)) should have highest probability
    assert!(output[0] > output[1]);
    assert!(output[0] > output[2]);
}

#[test]
fn test_residual_with_activation() {
    let input = vec![1.0, 2.0, 3.0, 4.0];
    let weight = vec![
        0.1, 0.0, 0.0, 0.0, 0.0, 0.1, 0.0, 0.0, 0.0, 0.0, 0.1, 0.0, 0.0, 0.0, 0.0, 0.1,
    ];

    let proj = simd_matmul(&input, &weight, 4, 4);
    let mut proj_activated = proj.clone();
    simd_gelu(&mut proj_activated);

    let mut residual = input.clone();
    simd_add(&mut residual, &proj_activated);

    // Should be approximately input + gelu(0.1 * input)
    for (&r, &i) in residual.iter().zip(input.iter()) {
        assert!(r > i); // Residual should be larger than input
        assert!(r < i * 1.5); // But not too much larger
    }
}