realizar 0.8.4

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
406
407
408
409
410
411
412
413
414
415
416
417
418
419
420
421
422
423
424
425
426
427
428
429
430
431
432
433

//! Part 02: SIMD operations and BF16/F16 conversion tests
//!
//! Additional coverage for simd.rs including:
//! - Edge cases for SIMD operations
//! - BF16/F16 conversion edge cases
//! - Streaming matmul tests
//! - Large input handling

use crate::inference::{
    simd_add, simd_bf16_dot, simd_bf16_matmul, simd_bf16_to_f32, simd_dot, simd_f16_to_f32,
    simd_gelu, simd_matmul, simd_mul, simd_silu, simd_softmax,
};

// ============================================================================
// simd_matmul Additional Edge Cases
// ============================================================================

#[test]
fn test_simd_matmul_single_element() {
    let input = vec![3.0];
    let weight = vec![2.0];
    let output = simd_matmul(&input, &weight, 1, 1);

    assert_eq!(output.len(), 1);
    assert!((output[0] - 6.0).abs() < 1e-5);
}

#[test]
fn test_simd_matmul_row_vector_output() {
    // 1x4 input, 4x1 weight -> 1x1 output
    let input = vec![1.0, 2.0, 3.0, 4.0];
    let weight = vec![1.0, 1.0, 1.0, 1.0]; // Single row that sums
    let output = simd_matmul(&input, &weight, 4, 1);

    assert_eq!(output.len(), 1);
    assert!((output[0] - 10.0).abs() < 1e-5);
}

#[test]
fn test_simd_matmul_zeros_input() {
    let input = vec![0.0; 4];
    let weight = vec![1.0; 8]; // 2x4 matrix
    let output = simd_matmul(&input, &weight, 4, 2);

    assert_eq!(output.len(), 2);
    for &v in &output {
        assert!((v).abs() < 1e-5);
    }
}

#[test]
fn test_simd_matmul_zeros_weight() {
    let input = vec![1.0, 2.0, 3.0, 4.0];
    let weight = vec![0.0; 8]; // 2x4 matrix of zeros
    let output = simd_matmul(&input, &weight, 4, 2);

    assert_eq!(output.len(), 2);
    for &v in &output {
        assert!((v).abs() < 1e-5);
    }
}

#[test]
fn test_simd_matmul_negative_values() {
    let input = vec![-1.0, -2.0];
    let weight = vec![
        -1.0, -1.0, // row 0: sum = -(-1) + -(-2) = 3
        1.0, 1.0, // row 1: sum = -1 + -2 = -3
    ];
    let output = simd_matmul(&input, &weight, 2, 2);

    assert_eq!(output.len(), 2);
    assert!((output[0] - 3.0).abs() < 1e-5);
    assert!((output[1] - (-3.0)).abs() < 1e-5);
}

#[test]
fn test_simd_matmul_non_square() {
    // 3 input -> 5 output
    let input = vec![1.0, 2.0, 3.0];
    let weight = vec![
        1.0, 0.0, 0.0, // row 0
        0.0, 1.0, 0.0, // row 1
        0.0, 0.0, 1.0, // row 2
        1.0, 1.0, 0.0, // row 3
        0.0, 1.0, 1.0, // row 4
    ];
    let output = simd_matmul(&input, &weight, 3, 5);

    assert_eq!(output.len(), 5);
    assert!((output[0] - 1.0).abs() < 1e-5); // x
    assert!((output[1] - 2.0).abs() < 1e-5); // y
    assert!((output[2] - 3.0).abs() < 1e-5); // z
    assert!((output[3] - 3.0).abs() < 1e-5); // x + y
    assert!((output[4] - 5.0).abs() < 1e-5); // y + z
}

#[test]
fn test_simd_matmul_large_dimensions() {
    // Test with dimensions that span multiple tiles
    let in_dim = 257; // Just over TILE_SIZE of 64 * 4
    let out_dim = 129; // Just over TILE_SIZE of 64 * 2
    let input: Vec<f32> = (0..in_dim).map(|i| (i % 10) as f32 * 0.1).collect();
    let weight = vec![0.0; out_dim * in_dim];

    let output = simd_matmul(&input, &weight, in_dim, out_dim);

    assert_eq!(output.len(), out_dim);
    for &v in &output {
        assert!(v.is_finite());
    }
}

#[test]
fn test_simd_matmul_tile_boundary() {
    // Test exactly at tile boundary (64)
    let in_dim = 64;
    let out_dim = 64;
    let input: Vec<f32> = vec![1.0; in_dim];

    // Weight: identity matrix at position 0
    let mut weight = vec![0.0; out_dim * in_dim];
    for i in 0..out_dim {
        weight[i * in_dim + i] = 1.0;
    }

    let output = simd_matmul(&input, &weight, in_dim, out_dim);

    assert_eq!(output.len(), out_dim);
    for &v in &output {
        assert!((v - 1.0).abs() < 1e-5);
    }
}

// ============================================================================
// simd_dot Additional Edge Cases
// ============================================================================

#[test]
fn test_simd_dot_empty() {
    let a: Vec<f32> = vec![];
    let b: Vec<f32> = vec![];
    let result = simd_dot(&a, &b);
    assert!((result).abs() < 1e-5);
}

#[test]
fn test_simd_dot_single() {
    let a = vec![3.0];
    let b = vec![4.0];
    let result = simd_dot(&a, &b);
    assert!((result - 12.0).abs() < 1e-5);
}

#[test]
fn test_simd_dot_alternating_signs() {
    let a = vec![1.0, -1.0, 1.0, -1.0];
    let b = vec![1.0, 1.0, 1.0, 1.0];
    let result = simd_dot(&a, &b);
    assert!((result).abs() < 1e-5); // Should cancel out
}

#[test]
fn test_simd_dot_large_values() {
    let a = vec![1e10, 1e10];
    let b = vec![1e10, 1e10];
    let result = simd_dot(&a, &b);
    assert!(result.is_finite());
    assert!((result - 2e20).abs() < 1e15); // Some tolerance for large values
}

#[test]
fn test_simd_dot_small_values() {
    let a = vec![1e-10, 1e-10];
    let b = vec![1e-10, 1e-10];
    let result = simd_dot(&a, &b);
    assert!(result.is_finite());
    assert!((result - 2e-20).abs() < 1e-25);
}

// ============================================================================
// simd_add/simd_mul Edge Cases
// ============================================================================

#[test]
fn test_simd_add_empty() {
    let mut a: Vec<f32> = vec![];
    let b: Vec<f32> = vec![];
    simd_add(&mut a, &b);
    assert!(a.is_empty());
}

#[test]
fn test_simd_add_single() {
    let mut a = vec![1.5];
    let b = vec![2.5];
    simd_add(&mut a, &b);
    assert!((a[0] - 4.0).abs() < 1e-5);
}

#[test]
fn test_simd_add_large_vector() {
    let mut a: Vec<f32> = (0..1000).map(|i| i as f32).collect();
    let b: Vec<f32> = vec![1.0; 1000];
    simd_add(&mut a, &b);

    for (i, &v) in a.iter().enumerate() {
        assert!((v - (i as f32 + 1.0)).abs() < 1e-5);
    }
}

#[test]
fn test_simd_mul_empty() {
    let mut a: Vec<f32> = vec![];
    let b: Vec<f32> = vec![];
    simd_mul(&mut a, &b);
    assert!(a.is_empty());
}

#[test]
fn test_simd_mul_single() {
    let mut a = vec![3.0];
    let b = vec![4.0];
    simd_mul(&mut a, &b);
    assert!((a[0] - 12.0).abs() < 1e-5);
}

#[test]
fn test_simd_mul_large_vector() {
    let mut a: Vec<f32> = vec![2.0; 1000];
    let b: Vec<f32> = vec![3.0; 1000];
    simd_mul(&mut a, &b);

    for &v in &a {
        assert!((v - 6.0).abs() < 1e-5);
    }
}

// ============================================================================
// simd_silu Edge Cases
// ============================================================================

#[test]
fn test_simd_silu_empty() {
    let mut data: Vec<f32> = vec![];
    simd_silu(&mut data);
    assert!(data.is_empty());
}

#[test]
fn test_simd_silu_single() {
    let mut data = vec![0.0];
    simd_silu(&mut data);
    assert!((data[0]).abs() < 1e-5);
}

#[test]
fn test_simd_silu_symmetry() {
    // SiLU is NOT symmetric: silu(-x) != -silu(x)
    let mut pos = vec![2.0];
    let mut neg = vec![-2.0];
    simd_silu(&mut pos);
    simd_silu(&mut neg);

    // |silu(2)| should be much larger than |silu(-2)|
    assert!(pos[0].abs() > neg[0].abs() * 5.0);
}

#[test]
fn test_simd_silu_gradient_like() {
    // Test that silu(x) increases as x increases for positive x
    let mut data: Vec<f32> = (0..10).map(|i| i as f32 * 0.5).collect();
    simd_silu(&mut data);

    // For positive x, silu should be monotonically increasing
    for i in 1..data.len() {
        assert!(data[i] >= data[i - 1]);
    }
}

// ============================================================================
// simd_gelu Edge Cases
// ============================================================================

#[test]
fn test_simd_gelu_empty() {
    let mut data: Vec<f32> = vec![];
    simd_gelu(&mut data);
    assert!(data.is_empty());
}

#[test]
fn test_simd_gelu_single() {
    let mut data = vec![0.0];
    simd_gelu(&mut data);
    assert!((data[0]).abs() < 1e-5);
}

#[test]
fn test_simd_gelu_large_positive() {
    let mut data = vec![5.0];
    simd_gelu(&mut data);
    // gelu(5) should be very close to 5
    assert!((data[0] - 5.0).abs() < 0.01);
}

#[test]
fn test_simd_gelu_large_negative() {
    let mut data = vec![-5.0];
    simd_gelu(&mut data);
    // gelu(-5) should be very close to 0
    assert!((data[0]).abs() < 0.01);
}

#[test]
fn test_simd_gelu_batch_consistency() {
    // Verify batch processing gives same result as individual
    let mut batch = vec![0.0, 1.0, -1.0, 2.0, -2.0];
    let mut individual_results = vec![0.0; 5];

    for (i, &v) in [0.0, 1.0, -1.0, 2.0, -2.0].iter().enumerate() {
        let mut single = vec![v];
        simd_gelu(&mut single);
        individual_results[i] = single[0];
    }

    simd_gelu(&mut batch);

    for (batch_val, ind_val) in batch.iter().zip(individual_results.iter()) {
        assert!((batch_val - ind_val).abs() < 1e-5);
    }
}

// ============================================================================
// simd_softmax Edge Cases
// ============================================================================

#[test]
fn test_simd_softmax_two_elements() {
    let mut data = vec![0.0, 0.0];
    simd_softmax(&mut data);

    // Uniform input -> uniform output
    assert!((data[0] - 0.5).abs() < 1e-5);
    assert!((data[1] - 0.5).abs() < 1e-5);
}

#[test]
fn test_simd_softmax_dominant_element() {
    let mut data = vec![0.0, 100.0, 0.0];
    simd_softmax(&mut data);

    // Middle element should dominate
    assert!(data[1] > 0.99);
    assert!(data[0] < 0.01);
    assert!(data[2] < 0.01);
}

#[test]
fn test_simd_softmax_negative_dominant() {
    let mut data = vec![-100.0, 0.0, -100.0];
    simd_softmax(&mut data);

    // Middle element should dominate
    assert!(data[1] > 0.99);
}

#[test]
fn test_simd_softmax_all_same_large() {
    let mut data = vec![1000.0; 4];
    simd_softmax(&mut data);

    // All same -> uniform distribution
    for &v in &data {
        assert!((v - 0.25).abs() < 1e-5);
    }
}

#[test]
fn test_simd_softmax_preserves_sum() {
    let mut data = vec![0.5, -0.5, 1.0, -1.0, 0.0];
    simd_softmax(&mut data);

    let sum: f32 = data.iter().sum();
    assert!((sum - 1.0).abs() < 1e-5);
}

// ============================================================================
// BF16/F16 Conversion Edge Cases
// ============================================================================

#[test]
fn test_simd_bf16_to_f32_odd_count() {
    // Odd number of bytes (incomplete last value)
    let bf16_bytes = vec![0u8; 5]; // 2 complete values + 1 incomplete byte
    let result = simd_bf16_to_f32(&bf16_bytes);

    // Should only convert complete pairs (5/2 = 2)
    assert_eq!(result.len(), 2);
}

#[test]
fn test_simd_bf16_to_f32_special_values() {
    // Test zero
    let zero_bytes = half::bf16::from_f32(0.0).to_le_bytes();
    let result = simd_bf16_to_f32(&zero_bytes);
    assert!((result[0]).abs() < 1e-10);

    // Test negative zero
    let neg_zero_bytes = half::bf16::from_f32(-0.0).to_le_bytes();
    let result = simd_bf16_to_f32(&neg_zero_bytes);
    assert!((result[0]).abs() < 1e-10);
}

#[test]
fn test_simd_bf16_to_f32_infinity() {
    let inf_bytes = half::bf16::from_f32(f32::INFINITY).to_le_bytes();
    let result = simd_bf16_to_f32(&inf_bytes);
    assert!(result[0].is_infinite() && result[0].is_sign_positive());

    let neg_inf_bytes = half::bf16::from_f32(f32::NEG_INFINITY).to_le_bytes();
    let result = simd_bf16_to_f32(&neg_inf_bytes);
    assert!(result[0].is_infinite() && result[0].is_sign_negative());
}

#[test]
fn test_simd_bf16_to_f32_nan() {
    let nan_bytes = half::bf16::from_f32(f32::NAN).to_le_bytes();
    let result = simd_bf16_to_f32(&nan_bytes);
    assert!(result[0].is_nan());
}

include!("simd_bf16.rs");