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
434
435
436
437
438
439
440
441
442

//! Tests for batched forward pass implementations
//!
//! Coverage targets for batch.rs uncovered paths:
//! - Empty tokens/prompts error paths
//! - Bias handling in attention and FFN
//! - Temperature and top-k sampling branches
//! - batch_throughput_factor ranges
//! - Softmax variants (standard, online, tiled)
//! - Tiled attention with various tile sizes

use crate::gguf::test_helpers::create_test_model_with_config;
use crate::gguf::{
    GGUFConfig, OwnedQuantizedKVCache, OwnedQuantizedModel, QuantizedGenerateConfig,
};

fn test_config() -> GGUFConfig {
    GGUFConfig {
        architecture: "test".to_string(),
        constraints: crate::gguf::ArchConstraints::from_architecture("test"),
        hidden_dim: 64,
        intermediate_dim: 128,
        num_heads: 4,
        num_kv_heads: 4,
        num_layers: 1,
        vocab_size: 100,
        rope_theta: 10000.0,
        context_length: 512,
        eps: 1e-5,
        rope_type: 0,
        explicit_head_dim: None,
        bos_token_id: None,
        eos_token_id: None,
    }
}

// ============================================================================
// generate_with_smallvec tests
// ============================================================================

#[test]
fn test_generate_with_smallvec_empty_prompt_error() {
    let config = test_config();
    let model = create_test_model_with_config(&config);
    let gen_config = QuantizedGenerateConfig::deterministic(5);

    let result = model.generate_with_smallvec(&[], &gen_config);
    assert!(result.is_err());
    let err = result.unwrap_err();
    assert!(format!("{:?}", err).contains("empty"));
}

#[test]
fn test_generate_with_smallvec_greedy_sampling() {
    let config = test_config();
    let model = create_test_model_with_config(&config);
    let gen_config = QuantizedGenerateConfig::deterministic(3);

    let result = model.generate_with_smallvec(&[1], &gen_config);
    assert!(result.is_ok());
    let tokens = result.unwrap();
    // Should have at least the prompt token
    assert!(!tokens.is_empty());
    assert_eq!(tokens[0], 1);
}

#[test]
fn test_generate_with_smallvec_temperature_sampling() {
    let config = test_config();
    let model = create_test_model_with_config(&config);
    let gen_config = QuantizedGenerateConfig::default()
        .with_max_tokens(2)
        .with_temperature(1.0)
        .with_top_k(5);

    let result = model.generate_with_smallvec(&[1], &gen_config);
    assert!(result.is_ok());
}

#[test]
fn test_generate_with_smallvec_stop_token() {
    let config = test_config();
    let model = create_test_model_with_config(&config);
    // Use token 0 as stop token (likely to be generated from zero weights)
    let gen_config = QuantizedGenerateConfig::deterministic(10).with_stop_tokens(vec![0]);

    let result = model.generate_with_smallvec(&[1], &gen_config);
    assert!(result.is_ok());
}

// ============================================================================
// batch_generate tests
// ============================================================================

#[test]
fn test_batch_generate_empty_prompts_error() {
    let config = test_config();
    let model = create_test_model_with_config(&config);
    let gen_config = QuantizedGenerateConfig::deterministic(5);

    let result = model.batch_generate(&[], &gen_config);
    assert!(result.is_err());
    let err = result.unwrap_err();
    assert!(format!("{:?}", err).contains("empty"));
}

#[test]
fn test_batch_generate_single_prompt_optimization() {
    let config = test_config();
    let model = create_test_model_with_config(&config);
    let gen_config = QuantizedGenerateConfig::deterministic(2);

    let prompt: &[u32] = &[1, 2];
    let result = model.batch_generate(&[prompt], &gen_config);
    assert!(result.is_ok());
    let outputs = result.unwrap();
    assert_eq!(outputs.len(), 1);
    assert!(outputs[0].len() >= 2);
}

#[test]
fn test_batch_generate_multiple_prompts() {
    let config = test_config();
    let model = create_test_model_with_config(&config);
    let gen_config = QuantizedGenerateConfig::deterministic(2);

    let prompt1: &[u32] = &[1];
    let prompt2: &[u32] = &[2, 3];
    let result = model.batch_generate(&[prompt1, prompt2], &gen_config);
    assert!(result.is_ok());
    let outputs = result.unwrap();
    assert_eq!(outputs.len(), 2);
}

#[test]
fn test_batch_generate_with_stop_tokens() {
    let config = test_config();
    let model = create_test_model_with_config(&config);
    let gen_config = QuantizedGenerateConfig::deterministic(10).with_stop_tokens(vec![0]);

    let prompt1: &[u32] = &[1];
    let prompt2: &[u32] = &[2];
    let result = model.batch_generate(&[prompt1, prompt2], &gen_config);
    assert!(result.is_ok());
}

#[test]
fn test_batch_generate_with_temperature() {
    let config = test_config();
    let model = create_test_model_with_config(&config);
    let gen_config = QuantizedGenerateConfig::default()
        .with_max_tokens(2)
        .with_temperature(0.8)
        .with_top_k(10);

    let prompt1: &[u32] = &[1];
    let prompt2: &[u32] = &[2];
    let result = model.batch_generate(&[prompt1, prompt2], &gen_config);
    assert!(result.is_ok());
}

// ============================================================================
// batch_throughput_factor tests
// ============================================================================

#[test]
fn test_batch_throughput_factor_all_ranges() {
    // 0 or 1 => 1.0
    assert!((OwnedQuantizedModel::batch_throughput_factor(0) - 1.0).abs() < f64::EPSILON);
    assert!((OwnedQuantizedModel::batch_throughput_factor(1) - 1.0).abs() < f64::EPSILON);

    // 2..=4 => 1.8
    assert!((OwnedQuantizedModel::batch_throughput_factor(2) - 1.8).abs() < f64::EPSILON);
    assert!((OwnedQuantizedModel::batch_throughput_factor(4) - 1.8).abs() < f64::EPSILON);

    // 5..=8 => 2.5
    assert!((OwnedQuantizedModel::batch_throughput_factor(5) - 2.5).abs() < f64::EPSILON);
    assert!((OwnedQuantizedModel::batch_throughput_factor(8) - 2.5).abs() < f64::EPSILON);

    // 9..=16 => 3.5
    assert!((OwnedQuantizedModel::batch_throughput_factor(9) - 3.5).abs() < f64::EPSILON);
    assert!((OwnedQuantizedModel::batch_throughput_factor(16) - 3.5).abs() < f64::EPSILON);

    // 17..=32 => 5.0
    assert!((OwnedQuantizedModel::batch_throughput_factor(17) - 5.0).abs() < f64::EPSILON);
    assert!((OwnedQuantizedModel::batch_throughput_factor(32) - 5.0).abs() < f64::EPSILON);

    // >32 => 6.0
    assert!((OwnedQuantizedModel::batch_throughput_factor(33) - 6.0).abs() < f64::EPSILON);
    assert!((OwnedQuantizedModel::batch_throughput_factor(100) - 6.0).abs() < f64::EPSILON);
}

// ============================================================================
// forward_batch tests
// ============================================================================

#[test]
fn test_forward_batch_single_token() {
    let config = test_config();
    let model = create_test_model_with_config(&config);

    let result = model.forward_batch(&[1]);
    assert!(result.is_ok());
    let logits = result.unwrap();
    assert_eq!(logits.len(), config.vocab_size);
}

#[test]
fn test_forward_batch_multiple_tokens() {
    let config = test_config();
    let model = create_test_model_with_config(&config);

    let result = model.forward_batch(&[1, 2, 3]);
    assert!(result.is_ok());
    let logits = result.unwrap();
    // batch_size * vocab_size
    assert_eq!(logits.len(), 3 * config.vocab_size);
}

// ============================================================================
// prefill_batch tests
// ============================================================================

#[test]
fn test_prefill_batch_empty_prompt_error() {
    let config = test_config();
    let model = create_test_model_with_config(&config);
    let mut cache = OwnedQuantizedKVCache::from_config(&config, 100);

    let result = model.prefill_batch(&[], &mut cache);
    assert!(result.is_err());
    let err = result.unwrap_err();
    assert!(format!("{:?}", err).contains("empty"));
}

#[test]
fn test_prefill_batch_single_token() {
    let config = test_config();
    let model = create_test_model_with_config(&config);
    let mut cache = OwnedQuantizedKVCache::from_config(&config, 100);

    let result = model.prefill_batch(&[1], &mut cache);
    assert!(result.is_ok());
    let logits = result.unwrap();
    assert_eq!(logits.len(), config.vocab_size);
}

#[test]
fn test_prefill_batch_multiple_tokens() {
    let config = test_config();
    let model = create_test_model_with_config(&config);
    let mut cache = OwnedQuantizedKVCache::from_config(&config, 100);

    let result = model.prefill_batch(&[1, 2, 3], &mut cache);
    assert!(result.is_ok());
    let logits = result.unwrap();
    assert_eq!(logits.len(), config.vocab_size);
}

// ============================================================================
// standard_softmax tests
// ============================================================================

#[test]
fn test_standard_softmax_empty() {
    let config = test_config();
    let model = create_test_model_with_config(&config);

    let result = model.standard_softmax(&[]);
    assert!(result.is_empty());
}

#[test]
fn test_standard_softmax_single_element() {
    let config = test_config();
    let model = create_test_model_with_config(&config);

    let result = model.standard_softmax(&[1.0]);
    assert_eq!(result.len(), 1);
    assert!((result[0] - 1.0).abs() < 1e-6);
}

#[test]
fn test_standard_softmax_sums_to_one() {
    let config = test_config();
    let model = create_test_model_with_config(&config);

    let result = model.standard_softmax(&[1.0, 2.0, 3.0]);
    let sum: f32 = result.iter().sum();
    assert!((sum - 1.0).abs() < 1e-6);
}

#[test]
fn test_standard_softmax_ordering() {
    let config = test_config();
    let model = create_test_model_with_config(&config);

    let result = model.standard_softmax(&[1.0, 2.0, 3.0]);
    // Larger input should have larger output
    assert!(result[2] > result[1]);
    assert!(result[1] > result[0]);
}

// ============================================================================
// online_softmax tests
// ============================================================================

#[test]
fn test_online_softmax_empty() {
    let config = test_config();
    let model = create_test_model_with_config(&config);

    let result = model.online_softmax(&[], 4);
    assert!(result.is_ok());
    assert!(result.unwrap().is_empty());
}

#[test]
fn test_online_softmax_matches_standard() {
    let config = test_config();
    let model = create_test_model_with_config(&config);

    let scores = vec![1.0, 2.0, 3.0, 4.0, 5.0, 6.0];
    let standard = model.standard_softmax(&scores);
    let online = model.online_softmax(&scores, 2).unwrap();

    for (s, o) in standard.iter().zip(online.iter()) {
        assert!((s - o).abs() < 1e-5, "standard={}, online={}", s, o);
    }
}

#[test]
fn test_online_softmax_various_tile_sizes() {
    let config = test_config();
    let model = create_test_model_with_config(&config);

    let scores = vec![1.0, 2.0, 3.0, 4.0, 5.0, 6.0, 7.0, 8.0];
    let standard = model.standard_softmax(&scores);

    for tile_size in [1, 2, 3, 4, 8, 16] {
        let online = model.online_softmax(&scores, tile_size).unwrap();
        for (s, o) in standard.iter().zip(online.iter()) {
            assert!((s - o).abs() < 1e-5, "tile_size={}", tile_size);
        }
    }
}

#[test]
fn test_online_softmax_tile_size_zero() {
    let config = test_config();
    let model = create_test_model_with_config(&config);

    // tile_size 0 should be treated as 1
    let result = model.online_softmax(&[1.0, 2.0], 0);
    assert!(result.is_ok());
    let sum: f32 = result.unwrap().iter().sum();
    assert!((sum - 1.0).abs() < 1e-6);
}

// ============================================================================
// standard_single_head_attention tests
// ============================================================================

#[test]
fn test_standard_single_head_attention_basic() {
    let config = test_config();
    let model = create_test_model_with_config(&config);

    let seq_len = 2;
    let head_dim = 4;
    let scale = 1.0 / (head_dim as f32).sqrt();

    // Q, K, V: [seq_len, head_dim]
    let q = vec![1.0; seq_len * head_dim];
    let k = vec![1.0; seq_len * head_dim];
    let v = vec![1.0; seq_len * head_dim];

    let result = model.standard_single_head_attention(&q, &k, &v, seq_len, head_dim, scale);
    assert!(result.is_ok());
    let output = result.unwrap();
    assert_eq!(output.len(), seq_len * head_dim);
}

#[test]
fn test_standard_single_head_attention_identity() {
    let config = test_config();
    let model = create_test_model_with_config(&config);

    // Single position: attention weights will be 1.0, so output = V
    let seq_len = 1;
    let head_dim = 4;
    let scale = 1.0 / (head_dim as f32).sqrt();

    let q = vec![1.0; head_dim];
    let k = vec![1.0; head_dim];
    let v = vec![2.0, 3.0, 4.0, 5.0];

    let result = model.standard_single_head_attention(&q, &k, &v, seq_len, head_dim, scale);
    assert!(result.is_ok());
    let output = result.unwrap();
    // Output should be V (since softmax of single element is 1.0)
    for (o, expected) in output.iter().zip(v.iter()) {
        assert!((o - expected).abs() < 1e-5);
    }
}

// ============================================================================
// tiled_single_head_attention tests
// ============================================================================

#[test]
fn test_tiled_single_head_attention_matches_standard() {
    let config = test_config();
    let model = create_test_model_with_config(&config);

    let seq_len = 4;
    let head_dim = 4;
    let scale = 1.0 / (head_dim as f32).sqrt();

    // Random-ish values
    let q: Vec<f32> = (0..seq_len * head_dim)
        .map(|i| (i % 7) as f32 * 0.1)
        .collect();
    let k: Vec<f32> = (0..seq_len * head_dim)
        .map(|i| (i % 5) as f32 * 0.1)
        .collect();
    let v: Vec<f32> = (0..seq_len * head_dim)
        .map(|i| (i % 3) as f32 * 0.1)
        .collect();

    let standard = model
        .standard_single_head_attention(&q, &k, &v, seq_len, head_dim, scale)
        .unwrap();
    let tiled = model
        .tiled_single_head_attention(&q, &k, &v, seq_len, head_dim, scale, 2)
        .unwrap();

    for (s, t) in standard.iter().zip(tiled.iter()) {
        assert!((s - t).abs() < 1e-4, "standard={}, tiled={}", s, t);
    }
}

include!("batch_tests_tiled_single.rs");