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
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
297
298
299
300
301
302
303
304
305
306
307
308
309
310
311
312
313
314
315
316
317
318
319
320
321
322
323
324
325
326
327
328
329
330
331
332
333
334
335
336
337
338
339
340
341
342
343
344
345
346
347
348
349
350
351
352
353
354
355
356
357
358
359
360
361
362
363
364
365
366
367
368
369
370
371
372
373
374
375
376
377
378
379
380
381
382
383
384
385
386
387
388
389
390
391
392
393
394
395
396
397
398
399
400
401
402
403
404
405

use super::*;

// ============================================================================
// FUSED Q4_K DOT PRODUCT TESTS (SCALAR)
// ============================================================================

#[test]
fn test_fused_q4k_dot_invalid_data_length() {
    // Q4_K super-block is 144 bytes
    let data = vec![0u8; 100]; // Not a multiple of 144
    let activations = vec![0.0f32; 256];

    let result = fused_q4k_dot(&data, &activations);
    assert!(result.is_err());
    assert!(result.unwrap_err().to_string().contains("not a multiple"));
}

#[test]
fn test_fused_q4k_dot_activation_length_mismatch() {
    // One super-block = 144 bytes = 256 values
    let data = vec![0u8; 144];
    let activations = vec![0.0f32; 128]; // Should be 256

    let result = fused_q4k_dot(&data, &activations);
    assert!(result.is_err());
    assert!(result.unwrap_err().to_string().contains("doesn't match"));
}

#[test]
fn test_fused_q4k_dot_zero_data() {
    // All zeros should produce zero dot product
    let data = vec![0u8; 144];
    let activations = vec![1.0f32; 256];

    let result = fused_q4k_dot(&data, &activations).unwrap();
    // d = 0, so all values dequantize to 0
    assert_eq!(result, 0.0);
}

#[test]
fn test_fused_q4k_dot_single_super_block() {
    // Create valid Q4_K data: d (2) + dmin (2) + scales (12) + qs (128)
    let mut data = vec![0u8; 144];

    // Set d = 1.0 (f16: 0x3C00)
    data[0..2].copy_from_slice(&0x3C00u16.to_le_bytes());

    // Set dmin = 0.5 (f16: 0x3800)
    data[2..4].copy_from_slice(&0x3800u16.to_le_bytes());

    // Set scales (12 bytes packed)
    for i in 0..12 {
        data[4 + i] = 0x11;
    }

    // Set qs values to a pattern
    for i in 0..128 {
        data[16 + i] = 0x55; // Pattern: low=5, high=5
    }

    let activations = vec![1.0f32; 256];

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

#[test]
fn test_fused_q4k_dot_multiple_super_blocks() {
    // Two super-blocks = 288 bytes = 512 values
    let mut data = vec![0u8; 288];

    // Set d = 0.5 for both blocks
    data[0..2].copy_from_slice(&0x3800u16.to_le_bytes());
    data[144..146].copy_from_slice(&0x3800u16.to_le_bytes());

    let activations = vec![0.5f32; 512];

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

#[test]
fn test_fused_q4k_dot_empty_data() {
    let data = vec![];
    let activations = vec![];

    let result = fused_q4k_dot(&data, &activations);
    assert!(result.is_ok());
    assert_eq!(result.unwrap(), 0.0);
}

// ============================================================================
// FUSED Q4_K DOT PRODUCT TESTS (SIMD)
// ============================================================================

#[test]
fn test_fused_q4k_dot_simd_invalid_input() {
    let data = vec![0u8; 100]; // Invalid length
    let activations = vec![0.0f32; 256];

    let result = fused_q4k_dot_simd(&data, &activations);
    assert!(result.is_err());
}

#[test]
fn test_fused_q4k_dot_simd_matches_scalar() {
    // Create valid Q4_K data
    let mut data = vec![0u8; 144];

    // Set d = 1.0, dmin = 0.5
    data[0..2].copy_from_slice(&0x3C00u16.to_le_bytes());
    data[2..4].copy_from_slice(&0x3800u16.to_le_bytes());

    // Set scales
    for i in 0..12 {
        data[4 + i] = 0x11;
    }

    // Set qs to deterministic pattern
    for i in 0..128 {
        data[16 + i] = ((i * 3) % 256) as u8;
    }

    let activations: Vec<f32> = (0..256).map(|i| (i as f32) * 0.01).collect();

    let scalar_result = fused_q4k_dot(&data, &activations).unwrap();
    let simd_result = fused_q4k_dot_simd(&data, &activations).unwrap();

    // Allow small tolerance for SIMD vs scalar (FMA ordering differences)
    let rel_err = if scalar_result.abs() > 1e-6 {
        (simd_result - scalar_result).abs() / scalar_result.abs()
    } else {
        (simd_result - scalar_result).abs()
    };
    assert!(
        rel_err < 0.001,
        "scalar={} simd={} rel_err={}",
        scalar_result,
        simd_result,
        rel_err
    );
}

#[test]
fn test_fused_q4k_dot_simd_zero_activations() {
    let mut data = vec![0u8; 144];
    data[0..2].copy_from_slice(&0x3C00u16.to_le_bytes());

    let activations = vec![0.0f32; 256];

    let result = fused_q4k_dot_simd(&data, &activations).unwrap();
    // Product of anything with zero should be zero
    assert_eq!(result, 0.0);
}

// ============================================================================
// FUSED Q4_K × Q8_K DOT PRODUCT TESTS
// ============================================================================

#[test]
fn test_fused_q4k_q8k_dot_invalid_data_length() {
    let data = vec![0u8; 100]; // Not a multiple of 144
    let q8k_scales = vec![1.0f32; 1];
    let q8k_quants = vec![10i8; 256];

    let result = fused_q4k_q8k_dot(&data, &q8k_scales, &q8k_quants);
    assert!(result.is_err());
}

#[test]
fn test_fused_q4k_q8k_dot_scales_too_small() {
    let data = vec![0u8; 144]; // 1 super-block
    let q8k_scales = vec![]; // Should be >= 1
    let q8k_quants = vec![10i8; 256];

    let result = fused_q4k_q8k_dot(&data, &q8k_scales, &q8k_quants);
    assert!(result.is_err());
}

#[test]
fn test_fused_q4k_q8k_dot_quants_too_small() {
    let data = vec![0u8; 144]; // 1 super-block = 256 values
    let q8k_scales = vec![1.0f32; 1];
    let q8k_quants = vec![10i8; 128]; // Should be 256

    let result = fused_q4k_q8k_dot(&data, &q8k_scales, &q8k_quants);
    assert!(result.is_err());
}

#[test]
fn test_fused_q4k_q8k_dot_zero_data() {
    let data = vec![0u8; 144];
    let q8k_scales = vec![0.0f32; 1]; // Zero scale
    let q8k_quants = vec![10i8; 256];

    let result = fused_q4k_q8k_dot(&data, &q8k_scales, &q8k_quants).unwrap();
    // With zero Q4_K d and zero Q8_K scale, result should be 0
    assert_eq!(result, 0.0);
}

#[test]
fn test_fused_q4k_q8k_dot_basic() {
    let mut data = vec![0u8; 144];

    // Set d = 1.0, dmin = 0.5
    data[0..2].copy_from_slice(&0x3C00u16.to_le_bytes());
    data[2..4].copy_from_slice(&0x3800u16.to_le_bytes());

    // Set scales
    for i in 0..12 {
        data[4 + i] = 0x11;
    }

    // Set qs
    for i in 0..128 {
        data[16 + i] = 0x55;
    }

    let q8k_scales = vec![0.1f32; 1];
    let q8k_quants = vec![10i8; 256];

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

#[test]
fn test_fused_q4k_q8k_dot_simd_invalid_input() {
    let data = vec![0u8; 100]; // Invalid
    let q8k_scales = vec![1.0f32; 1];
    let q8k_quants = vec![10i8; 256];

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

#[test]
fn test_fused_q4k_q8k_dot_simd_matches_scalar() {
    let mut data = vec![0u8; 144];

    // Set d = 1.0, dmin = 0.5
    data[0..2].copy_from_slice(&0x3C00u16.to_le_bytes());
    data[2..4].copy_from_slice(&0x3800u16.to_le_bytes());

    // Set scales
    for i in 0..12 {
        data[4 + i] = 0x11;
    }

    // Set qs to pattern
    for i in 0..128 {
        data[16 + i] = ((i * 7) % 256) as u8;
    }

    let q8k_scales = vec![0.1f32; 1];
    let q8k_quants: Vec<i8> = (0..256).map(|i| ((i % 64) - 32) as i8).collect();

    let scalar_result = fused_q4k_q8k_dot(&data, &q8k_scales, &q8k_quants).unwrap();
    let simd_result = fused_q4k_q8k_dot_simd(&data, &q8k_scales, &q8k_quants).unwrap();

    // Allow tolerance for SIMD vs scalar
    let rel_err = if scalar_result.abs() > 1e-6 {
        (simd_result - scalar_result).abs() / scalar_result.abs()
    } else {
        (simd_result - scalar_result).abs()
    };
    assert!(
        rel_err < 0.02,
        "scalar={} simd={} rel_err={}",
        scalar_result,
        simd_result,
        rel_err
    );
}

#[test]
fn test_fused_q4k_q8k_dot_multiple_super_blocks() {
    // Two super-blocks = 288 bytes = 512 values
    let mut data = vec![0u8; 288];

    // Set d for both blocks
    data[0..2].copy_from_slice(&0x3C00u16.to_le_bytes());
    data[144..146].copy_from_slice(&0x3C00u16.to_le_bytes());

    let q8k_scales = vec![0.1f32; 2];
    let q8k_quants = vec![5i8; 512];

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

// ============================================================================
// EDGE CASE TESTS
// ============================================================================

#[test]
fn test_fused_q4k_dot_max_nibble_values() {
    let mut data = vec![0u8; 144];

    // Set d = 1.0, dmin = 0
    data[0..2].copy_from_slice(&0x3C00u16.to_le_bytes());

    // Set qs to max nibble (0xFF = 15 low, 15 high)
    for i in 0..128 {
        data[16 + i] = 0xFF;
    }

    let activations = vec![1.0f32; 256];

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

#[test]
fn test_fused_q4k_q8k_dot_negative_quants() {
    let mut data = vec![0u8; 144];
    data[0..2].copy_from_slice(&0x3C00u16.to_le_bytes());

    let q8k_scales = vec![0.1f32; 1];
    let q8k_quants = vec![-10i8; 256]; // All negative

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

#[test]
fn test_fused_q4k_dot_simd_large_input() {
    // 8 super-blocks = 1152 bytes = 2048 values
    let mut data = vec![0u8; 1152];

    for sb in 0..8 {
        let offset = sb * 144;
        // Set d = 0.1
        data[offset..offset + 2].copy_from_slice(&0x2E66u16.to_le_bytes());
    }

    let activations = vec![0.5f32; 2048];

    let scalar_result = fused_q4k_dot(&data, &activations).unwrap();
    let simd_result = fused_q4k_dot_simd(&data, &activations).unwrap();

    let rel_err = if scalar_result.abs() > 1e-6 {
        (simd_result - scalar_result).abs() / scalar_result.abs()
    } else {
        (simd_result - scalar_result).abs()
    };
    assert!(
        rel_err < 0.001,
        "scalar={} simd={} rel_err={}",
        scalar_result,
        simd_result,
        rel_err
    );
}

#[test]
fn test_fused_q4k_q8k_dot_mixed_signs() {
    let mut data = vec![0u8; 144];
    data[0..2].copy_from_slice(&0x3C00u16.to_le_bytes());
    data[2..4].copy_from_slice(&0x3800u16.to_le_bytes());

    // Set qs to alternating pattern
    for i in 0..128 {
        data[16 + i] = if i % 2 == 0 { 0x0F } else { 0xF0 };
    }

    let q8k_scales = vec![0.1f32; 1];
    // Alternating positive and negative quants
    let q8k_quants: Vec<i8> = (0..256)
        .map(|i| if i % 2 == 0 { 10 } else { -10 })
        .collect();

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

#[test]
fn test_fused_q4k_dot_scale_extraction() {
    // Test that scales are correctly extracted and applied
    let mut data = vec![0u8; 144];

    // Set d = 2.0 (f16: 0x4000)
    data[0..2].copy_from_slice(&0x4000u16.to_le_bytes());

    // Set dmin = 0 (no offset)
    data[2..4].copy_from_slice(&[0x00, 0x00]);

    // Set first scale to max (63) - packed format
    data[4] = 0x3F; // First scale = 63

    // Set qs to 1 (low nibble only)
    for i in 0..128 {
        data[16 + i] = 0x01;
    }

    let activations = vec![1.0f32; 256];

    let result = fused_q4k_dot(&data, &activations);
    assert!(result.is_ok());
    // Result should be non-zero since d > 0, scale > 0, qs > 0
    assert!(result.unwrap().abs() > 0.0);
}

include!("fused_k_tests_q4k.rs");
include!("fused_k_tests_dot_errors.rs");
include!("fused_k_tests_avx2_q4k.rs");