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
406
407
408
409
410
411
412
413
414
415
416
417
418
419
420
421
422
423
424


#[test]
#[serial]
fn test_cov004_init_batched_kv_cache_without_regular() {
    if !CudaExecutor::is_available() {
        return;
    }
    let mut executor = CudaExecutor::new(0).expect("CUDA executor");

    // Without init_kv_cache_gpu, should fail
    let result = executor.init_batched_kv_cache_gpu(2, 4);
    assert!(result.is_err());
}

#[test]
#[serial]
fn test_cov004_init_batched_kv_cache_valid() {
    if !CudaExecutor::is_available() {
        return;
    }
    let mut executor = CudaExecutor::new(0).expect("CUDA executor");

    // Init regular first
    let _ = executor.init_kv_cache_gpu(2, 4, 4, 64, 128);

    // Now batched should work
    let result = executor.init_batched_kv_cache_gpu(2, 4);
    assert!(result.is_ok());
}

#[test]
#[serial]
fn test_cov004_reset_batched_kv_cache() {
    if !CudaExecutor::is_available() {
        return;
    }
    let mut executor = CudaExecutor::new(0).expect("CUDA executor");

    let _ = executor.init_kv_cache_gpu(2, 4, 4, 64, 128);
    let _ = executor.init_batched_kv_cache_gpu(2, 4);

    // Reset batched should work
    executor.reset_batched_kv_cache_gpu();
}

#[test]
#[serial]
fn test_cov004_flash_attention_cached_dimension_mismatch() {
    if !CudaExecutor::is_available() {
        return;
    }
    let mut executor = CudaExecutor::new(0).expect("CUDA executor");

    let num_heads = 4;
    let head_dim = 64;
    let hidden_dim = num_heads * head_dim; // 256

    let _ = executor.init_kv_cache_gpu(2, num_heads, num_heads, head_dim, 128);

    // Wrong Q dimension should fail
    let q_wrong = vec![0.0f32; 128]; // Should be 256
    let k = vec![0.0f32; hidden_dim];
    let v = vec![0.0f32; hidden_dim];
    let mut output = vec![0.0f32; hidden_dim];

    let result = executor.flash_attention_cached(0, &q_wrong, &k, &v, &mut output);
    assert!(result.is_err());
}

#[test]
#[serial]
fn test_cov004_flash_attention_cached_valid() {
    if !CudaExecutor::is_available() {
        return;
    }
    let mut executor = CudaExecutor::new(0).expect("CUDA executor");

    // Use small dimensions known to work with flash_attention_multi_head
    let num_heads = 4;
    let head_dim = 8; // Reduced from 64
    let hidden_dim = num_heads * head_dim; // 32

    let _ = executor.init_kv_cache_gpu(2, num_heads, num_heads, head_dim, 16);

    let q = vec![1.0f32; hidden_dim];
    let k = vec![1.0f32; hidden_dim];
    let v = vec![1.0f32; hidden_dim];
    let mut output = vec![0.0f32; hidden_dim];

    let result = executor.flash_attention_cached(0, &q, &k, &v, &mut output);
    assert!(
        result.is_ok(),
        "flash_attention_cached failed: {:?}",
        result.err()
    );
    assert_eq!(result.unwrap(), 1); // New sequence length is 1
    assert_eq!(executor.kv_cache_len(0), 1);
}

#[test]
#[serial]
fn test_cov004_flash_attention_cached_multiple_tokens() {
    if !CudaExecutor::is_available() {
        return;
    }
    let mut executor = CudaExecutor::new(0).expect("CUDA executor");

    let num_heads = 4;
    let head_dim = 8;
    let hidden_dim = num_heads * head_dim;

    let _ = executor.init_kv_cache_gpu(2, num_heads, num_heads, head_dim, 16);

    let q = vec![1.0f32; hidden_dim];
    let k = vec![1.0f32; hidden_dim];
    let v = vec![1.0f32; hidden_dim];
    let mut output = vec![0.0f32; hidden_dim];

    // Add 3 tokens
    for i in 1..=3 {
        let result = executor.flash_attention_cached(0, &q, &k, &v, &mut output);
        assert!(result.is_ok(), "Token {} failed: {:?}", i, result.err());
        assert_eq!(result.unwrap(), i);
    }
    assert_eq!(executor.kv_cache_len(0), 3);
}

#[test]
#[serial]
fn test_cov004_flash_attention_cached_overflow() {
    if !CudaExecutor::is_available() {
        return;
    }
    let mut executor = CudaExecutor::new(0).expect("CUDA executor");

    let num_heads = 2;
    let head_dim = 8;
    let hidden_dim = num_heads * head_dim;
    let max_len = 4; // Very small for fast overflow

    let _ = executor.init_kv_cache_gpu(1, num_heads, num_heads, head_dim, max_len);

    let q = vec![1.0f32; hidden_dim];
    let k = vec![1.0f32; hidden_dim];
    let v = vec![1.0f32; hidden_dim];
    let mut output = vec![0.0f32; hidden_dim];

    // Fill cache
    for i in 0..max_len {
        let result = executor.flash_attention_cached(0, &q, &k, &v, &mut output);
        assert!(
            result.is_ok(),
            "Token {} failed during fill: {:?}",
            i,
            result.err()
        );
    }

    // Next should overflow
    let result = executor.flash_attention_cached(0, &q, &k, &v, &mut output);
    assert!(result.is_err());
}

#[test]
#[serial]
fn test_cov004_incremental_attention_gpu_dimension_mismatch() {
    if !CudaExecutor::is_available() {
        return;
    }
    let mut executor = CudaExecutor::new(0).expect("CUDA executor");

    let num_heads = 4;
    let num_kv_heads = 2; // GQA: fewer KV heads
    let head_dim = 8;
    let q_dim = num_heads * head_dim; // 32
    let kv_dim = num_kv_heads * head_dim; // 16

    let _ = executor.init_kv_cache_gpu(2, num_heads, num_kv_heads, head_dim, 16);

    // Wrong Q dimension
    let q_wrong = vec![0.0f32; 16]; // Should be 32
    let k = vec![0.0f32; kv_dim];
    let v = vec![0.0f32; kv_dim];
    let mut output = vec![0.0f32; q_dim];

    let result = executor.incremental_attention_gpu(0, &q_wrong, &k, &v, &mut output);
    assert!(result.is_err());
}

#[test]
#[serial]
fn test_cov004_incremental_attention_gpu_kv_mismatch() {
    if !CudaExecutor::is_available() {
        return;
    }
    let mut executor = CudaExecutor::new(0).expect("CUDA executor");

    let num_heads = 4;
    let num_kv_heads = 2;
    let head_dim = 8;
    let q_dim = num_heads * head_dim;
    let kv_dim = num_kv_heads * head_dim;

    let _ = executor.init_kv_cache_gpu(2, num_heads, num_kv_heads, head_dim, 16);

    let q = vec![0.0f32; q_dim];
    let k_wrong = vec![0.0f32; 8]; // Wrong size
    let v = vec![0.0f32; kv_dim];
    let mut output = vec![0.0f32; q_dim];

    let result = executor.incremental_attention_gpu(0, &q, &k_wrong, &v, &mut output);
    assert!(result.is_err());
}

#[test]
#[serial]
fn test_cov004_incremental_attention_gpu_valid() {
    if !CudaExecutor::is_available() {
        return;
    }
    let mut executor = CudaExecutor::new(0).expect("CUDA executor");

    let num_heads = 4;
    let num_kv_heads = 4; // MHA (not GQA)
    let head_dim = 8;
    let hidden_dim = num_heads * head_dim;

    let _ = executor.init_kv_cache_gpu(2, num_heads, num_kv_heads, head_dim, 16);

    let q = vec![1.0f32; hidden_dim];
    let k = vec![1.0f32; hidden_dim];
    let v = vec![1.0f32; hidden_dim];
    let mut output = vec![0.0f32; hidden_dim];

    let result = executor.incremental_attention_gpu(0, &q, &k, &v, &mut output);
    assert!(
        result.is_ok(),
        "incremental_attention_gpu failed: {:?}",
        result.err()
    );
    assert_eq!(result.unwrap(), 1);
}

#[test]
#[serial]
fn test_cov004_incremental_attention_gpu_gqa() {
    if !CudaExecutor::is_available() {
        return;
    }
    let mut executor = CudaExecutor::new(0).expect("CUDA executor");

    // GQA: 4 Q heads, 2 KV heads
    let num_heads = 4;
    let num_kv_heads = 2;
    let head_dim = 8;
    let q_dim = num_heads * head_dim; // 32
    let kv_dim = num_kv_heads * head_dim; // 16

    let _ = executor.init_kv_cache_gpu(2, num_heads, num_kv_heads, head_dim, 16);

    let q = vec![1.0f32; q_dim];
    let k = vec![1.0f32; kv_dim];
    let v = vec![1.0f32; kv_dim];
    let mut output = vec![0.0f32; q_dim];

    let result = executor.incremental_attention_gpu(0, &q, &k, &v, &mut output);
    assert!(
        result.is_ok(),
        "GQA incremental attention failed: {:?}",
        result.err()
    );
}

#[test]
#[serial]
fn test_cov004_incremental_attention_overflow() {
    if !CudaExecutor::is_available() {
        return;
    }
    let mut executor = CudaExecutor::new(0).expect("CUDA executor");

    let num_heads = 2;
    let head_dim = 8;
    let hidden_dim = num_heads * head_dim;
    let max_len = 4;

    let _ = executor.init_kv_cache_gpu(1, num_heads, num_heads, head_dim, max_len);

    let q = vec![1.0f32; hidden_dim];
    let k = vec![1.0f32; hidden_dim];
    let v = vec![1.0f32; hidden_dim];
    let mut output = vec![0.0f32; hidden_dim];

    // Fill cache
    for i in 0..max_len {
        let result = executor.incremental_attention_gpu(0, &q, &k, &v, &mut output);
        assert!(
            result.is_ok(),
            "Fill token {} failed: {:?}",
            i,
            result.err()
        );
    }

    // Next should overflow
    let result = executor.incremental_attention_gpu(0, &q, &k, &v, &mut output);
    assert!(result.is_err());
}

#[test]
#[serial]
fn test_cov004_rollback_preserves_earlier_state() {
    if !CudaExecutor::is_available() {
        return;
    }
    let mut executor = CudaExecutor::new(0).expect("CUDA executor");

    let num_heads = 2;
    let head_dim = 8;
    let hidden_dim = num_heads * head_dim;

    let _ = executor.init_kv_cache_gpu(2, num_heads, num_heads, head_dim, 16);

    let q = vec![1.0f32; hidden_dim];
    let k = vec![1.0f32; hidden_dim];
    let v = vec![1.0f32; hidden_dim];
    let mut output = vec![0.0f32; hidden_dim];

    // Add 5 tokens
    for i in 0..5 {
        let result = executor.flash_attention_cached(0, &q, &k, &v, &mut output);
        assert!(result.is_ok(), "Token {} failed: {:?}", i, result.err());
    }
    assert_eq!(executor.kv_cache_len(0), 5);

    // Rollback to position 2
    executor.rollback_kv_cache_gpu(2);
    assert_eq!(executor.kv_cache_len(0), 2);

    // Can add more tokens from position 2
    let result = executor.flash_attention_cached(0, &q, &k, &v, &mut output);
    assert!(
        result.is_ok(),
        "Token after rollback failed: {:?}",
        result.err()
    );
    assert_eq!(executor.kv_cache_len(0), 3);
}

#[test]
#[serial]
fn test_cov004_reset_after_tokens() {
    if !CudaExecutor::is_available() {
        return;
    }
    let mut executor = CudaExecutor::new(0).expect("CUDA executor");

    let num_heads = 2;
    let head_dim = 8;
    let hidden_dim = num_heads * head_dim;

    let _ = executor.init_kv_cache_gpu(2, num_heads, num_heads, head_dim, 16);

    let q = vec![1.0f32; hidden_dim];
    let k = vec![1.0f32; hidden_dim];
    let v = vec![1.0f32; hidden_dim];
    let mut output = vec![0.0f32; hidden_dim];

    // Add tokens
    for i in 0..5 {
        let result = executor.flash_attention_cached(0, &q, &k, &v, &mut output);
        assert!(result.is_ok(), "Token {} failed: {:?}", i, result.err());
    }

    // Reset
    executor.reset_kv_cache_gpu();
    assert_eq!(executor.kv_cache_len(0), 0);

    // Can start fresh
    let result = executor.flash_attention_cached(0, &q, &k, &v, &mut output);
    assert!(
        result.is_ok(),
        "Fresh token after reset failed: {:?}",
        result.err()
    );
    assert_eq!(result.unwrap(), 1);
}

// =============================================================================
// COV-005: cuda/executor/attention.rs coverage tests
// Target: 16.19% → 50%+
// Tests for: incremental_attention_async, incremental_attention_into,
//            batched_incremental_attention_into, init_flash_decoding,
//            tensor_core_attention, gemm_fp16, flash_attention_memory_bytes
// =============================================================================

#[test]
fn test_cov005_flash_attention_memory_bytes() {
    // Static function - no CUDA needed
    let (naive, flash) = CudaExecutor::flash_attention_memory_bytes(128, 64);

    // Naive: 128 * 128 * 4 = 65536 bytes
    assert_eq!(naive, 128 * 128 * 4);

    // Flash: block_size(64) * block_size(64) * 4 * 2 = 32768 bytes
    assert_eq!(flash, 64 * 64 * 4 * 2);

    // Flash should always be smaller for reasonable seq_len
    assert!(flash < naive);
}

#[test]
fn test_cov005_flash_attention_memory_bytes_large() {
    let (naive, flash) = CudaExecutor::flash_attention_memory_bytes(4096, 128);

    // Naive: 4096 * 4096 * 4 = 67MB
    assert_eq!(naive, 4096 * 4096 * 4);

    // Flash is constant regardless of seq_len
    assert_eq!(flash, 64 * 64 * 4 * 2);

    // Huge difference for long sequences
    assert!(naive > flash * 1000);
}