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
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


#[test]
fn test_fused_q4k_q8k_dot_negative_quants() {
    let q4k_data = create_q4k_test_block(1.0, 0.0);
    let q8k_scales = vec![1.0f32; 1];
    let q8k_quants = vec![-1i8; QK_K];

    let result = fused_q4k_q8k_dot(&q4k_data, &q8k_scales, &q8k_quants);
    assert!(result.is_ok());
    assert!(result.expect("test value should be present").is_finite());
}

#[test]
fn test_fused_q4k_q8k_dot_mixed_quants() {
    let q4k_data = create_q4k_test_block(1.0, 0.5);
    let q8k_scales = vec![0.5f32; 1];
    let mut q8k_quants = Vec::with_capacity(QK_K);
    for i in 0..QK_K {
        q8k_quants.push(if i % 2 == 0 { 1 } else { -1 });
    }

    let result = fused_q4k_q8k_dot(&q4k_data, &q8k_scales, &q8k_quants);
    assert!(result.is_ok());
}

// ============================================================================
// Part 7: fused_q4k_q8k_dot_simd Tests
// ============================================================================

#[test]
fn test_fused_q4k_q8k_dot_simd_invalid_length() {
    let q4k_data = vec![0u8; 100];
    let q8k_scales = vec![1.0f32; 1];
    let q8k_quants = vec![0i8; QK_K];

    let result = fused_q4k_q8k_dot_simd(&q4k_data, &q8k_scales, &q8k_quants);
    assert!(result.is_err());
}

#[test]
fn test_fused_q4k_q8k_dot_simd_empty_inputs() {
    let q4k_data: Vec<u8> = vec![];
    let q8k_scales: Vec<f32> = vec![];
    let q8k_quants: Vec<i8> = vec![];

    let result = fused_q4k_q8k_dot_simd(&q4k_data, &q8k_scales, &q8k_quants);
    assert!(result.is_ok());
    assert_eq!(result.expect("test value should be present"), 0.0);
}

#[test]
fn test_fused_q4k_q8k_dot_simd_single_block() {
    let q4k_data = create_q4k_test_block(1.0, 0.25);
    let q8k_scales = vec![0.5f32; 1];
    let q8k_quants = vec![1i8; QK_K];

    let result = fused_q4k_q8k_dot_simd(&q4k_data, &q8k_scales, &q8k_quants);
    assert!(result.is_ok());
    assert!(result.expect("test value should be present").is_finite());
}

#[test]
fn test_fused_q4k_q8k_dot_simd_multiple_blocks() {
    let num_blocks = 8;
    let mut q4k_data = Vec::with_capacity(num_blocks * Q4K_BLOCK_BYTES);
    for _ in 0..num_blocks {
        q4k_data.extend_from_slice(&create_q4k_test_block(0.25, 0.1));
    }
    let q8k_scales = vec![0.25f32; num_blocks];
    let q8k_quants = vec![3i8; num_blocks * QK_K];

    let result = fused_q4k_q8k_dot_simd(&q4k_data, &q8k_scales, &q8k_quants);
    assert!(result.is_ok());
    assert!(result.expect("test value should be present").is_finite());
}

// ============================================================================
// Part 8: Q4K x Q8K Scalar vs SIMD Equivalence
// ============================================================================

#[test]
fn test_fused_q4k_q8k_dot_scalar_simd_equivalence_zero() {
    let q4k_data = vec![0u8; Q4K_BLOCK_BYTES];
    let q8k_scales = vec![1.0f32; 1];
    let q8k_quants = vec![1i8; QK_K];

    let scalar = fused_q4k_q8k_dot(&q4k_data, &q8k_scales, &q8k_quants).expect("test value should be present");
    let simd = fused_q4k_q8k_dot_simd(&q4k_data, &q8k_scales, &q8k_quants).expect("test value should be present");

    assert!(
        (scalar - simd).abs() < 1e-5,
        "scalar={} simd={}",
        scalar,
        simd
    );
}

#[test]
fn test_fused_q4k_q8k_dot_scalar_simd_equivalence_basic() {
    let q4k_data = create_q4k_test_block(1.0, 0.5);
    let q8k_scales = vec![0.5f32; 1];
    let q8k_quants = vec![2i8; QK_K];

    let scalar = fused_q4k_q8k_dot(&q4k_data, &q8k_scales, &q8k_quants).expect("test value should be present");
    let simd = fused_q4k_q8k_dot_simd(&q4k_data, &q8k_scales, &q8k_quants).expect("test value should be present");

    assert!(scalar.is_finite());
    assert!(simd.is_finite());
}

#[test]
fn test_fused_q4k_q8k_dot_scalar_simd_equivalence_varied() {
    let mut q4k_data = Vec::with_capacity(Q4K_BLOCK_BYTES);
    for i in 0..Q4K_BLOCK_BYTES {
        q4k_data.push((i * 41 % 256) as u8);
    }
    // Fix d and dmin to valid f16
    q4k_data[0] = 0x00;
    q4k_data[1] = 0x3C;
    q4k_data[2] = 0x00;
    q4k_data[3] = 0x38;

    let q8k_scales: Vec<f32> = (0..8).map(|i| 0.1 + i as f32 * 0.1).collect();
    let q8k_quants: Vec<i8> = (0..QK_K)
        .map(|i| ((i % 256) as i8).wrapping_sub(64))
        .collect();

    let scalar = fused_q4k_q8k_dot(&q4k_data, &q8k_scales, &q8k_quants).expect("test value should be present");
    let simd = fused_q4k_q8k_dot_simd(&q4k_data, &q8k_scales, &q8k_quants).expect("test value should be present");

    assert!(scalar.is_finite());
    assert!(simd.is_finite());
}

#[test]
fn test_fused_q4k_q8k_dot_scalar_simd_equivalence_many_blocks() {
    let num_blocks = 16;
    let mut q4k_data = Vec::with_capacity(num_blocks * Q4K_BLOCK_BYTES);
    for b in 0..num_blocks {
        let mut block = create_q4k_test_block(0.5, 0.25);
        for i in 12..Q4K_BLOCK_BYTES {
            block[i] = ((b * 23 + i * 11) % 256) as u8;
        }
        q4k_data.extend_from_slice(&block);
    }

    let q8k_scales = vec![0.25f32; num_blocks];
    let q8k_quants: Vec<i8> = (0..num_blocks * QK_K)
        .map(|i| ((i % 127) as i8) - 63)
        .collect();

    let scalar = fused_q4k_q8k_dot(&q4k_data, &q8k_scales, &q8k_quants).expect("test value should be present");
    let simd = fused_q4k_q8k_dot_simd(&q4k_data, &q8k_scales, &q8k_quants).expect("test value should be present");

    assert!(scalar.is_finite());
    assert!(simd.is_finite());
}

// ============================================================================
// Part 9: Edge Cases
// ============================================================================

#[test]
fn test_fused_q4k_dot_large_scale() {
    // Very large scale value
    let q4k_data = create_q4k_test_block(1000.0, 500.0);
    // Ensure d is a valid f16 representation of ~1000
    // f16 max is 65504, so 1000 is fine
    let activations = vec![1.0f32; QK_K];

    let result = fused_q4k_dot(&q4k_data, &activations);
    assert!(result.is_ok());
    assert!(result.expect("test value should be present").is_finite());
}

#[test]
fn test_fused_q4k_dot_small_scale() {
    // Very small scale value
    let q4k_data = create_q4k_test_block(0.001, 0.0005);
    let activations = vec![1.0f32; QK_K];

    let result = fused_q4k_dot(&q4k_data, &activations);
    assert!(result.is_ok());
}

#[test]
fn test_fused_q4k_dot_large_activations() {
    let q4k_data = create_q4k_test_block(1.0, 0.5);
    let activations = vec![1000.0f32; QK_K];

    let result = fused_q4k_dot(&q4k_data, &activations);
    assert!(result.is_ok());
    assert!(result.expect("test value should be present").is_finite());
}

#[test]
fn test_fused_q4k_q8k_dot_extreme_quants() {
    let q4k_data = create_q4k_test_block(1.0, 0.0);
    let q8k_scales = vec![1.0f32; 1];
    let q8k_quants = vec![127i8; QK_K]; // Max i8 value

    let result = fused_q4k_q8k_dot(&q4k_data, &q8k_scales, &q8k_quants);
    assert!(result.is_ok());
    assert!(result.expect("test value should be present").is_finite());
}

#[test]
fn test_fused_q4k_q8k_dot_min_quants() {
    let q4k_data = create_q4k_test_block(1.0, 0.0);
    let q8k_scales = vec![1.0f32; 1];
    let q8k_quants = vec![-128i8; QK_K]; // Min i8 value

    let result = fused_q4k_q8k_dot(&q4k_data, &q8k_scales, &q8k_quants);
    assert!(result.is_ok());
    assert!(result.expect("test value should be present").is_finite());
}

// ============================================================================
// Helper Functions
// ============================================================================

/// Create a Q4_K test block with specified d and dmin values
///
/// Q4_K super-block format (144 bytes):
/// - d: 2 bytes (f16)
/// - dmin: 2 bytes (f16)
/// - scales: 12 bytes (6-bit scales for 8 sub-blocks)
/// - quants: 128 bytes (4-bit quants for 256 values)
fn create_q4k_test_block(d: f32, dmin: f32) -> Vec<u8> {
    let mut block = vec![0u8; Q4K_BLOCK_BYTES];

    // Write d as f16 (simplified - use known f16 bit patterns)
    let d_f16 = f32_to_f16_approx(d);
    block[0] = (d_f16 & 0xFF) as u8;
    block[1] = ((d_f16 >> 8) & 0xFF) as u8;

    // Write dmin as f16
    let dmin_f16 = f32_to_f16_approx(dmin);
    block[2] = (dmin_f16 & 0xFF) as u8;
    block[3] = ((dmin_f16 >> 8) & 0xFF) as u8;

    // Scales (12 bytes) - set to mid-range values
    for i in 4..16 {
        block[i] = 0x88; // Mid-range scale
    }

    // Quants (128 bytes) - set to alternating pattern
    for i in 16..Q4K_BLOCK_BYTES {
        block[i] = 0x55; // Alternating nibbles
    }

    block
}

/// Approximate f32 to f16 conversion
fn f32_to_f16_approx(f: f32) -> u16 {
    if f == 0.0 {
        return 0;
    }
    if f.is_nan() {
        return 0x7E00;
    }
    if f.is_infinite() {
        return if f > 0.0 { 0x7C00 } else { 0xFC00 };
    }

    let bits = f.to_bits();
    let sign = ((bits >> 31) & 1) as u16;
    let exp = ((bits >> 23) & 0xFF) as i32;
    let mantissa = bits & 0x7FFFFF;

    // Bias adjustment: f32 bias is 127, f16 bias is 15
    let new_exp = exp - 127 + 15;

    if new_exp <= 0 {
        // Subnormal or zero
        0
    } else if new_exp >= 31 {
        // Overflow to infinity
        (sign << 15) | 0x7C00
    } else {
        // Normal number
        let new_mantissa = (mantissa >> 13) as u16;
        (sign << 15) | ((new_exp as u16) << 10) | new_mantissa
    }
}