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


// ============================================================================
// Tests (PMAT-802: T-COV-95)
// ============================================================================

#[cfg(test)]
mod tests {
    use super::*;

    // =========================================================================
    // StreamingKVCache Tests
    // =========================================================================

    #[test]
    fn test_streaming_kv_cache_new() {
        let cache = StreamingKVCache::new(4, 128, 8, 64);
        assert!(cache.is_empty());
        assert_eq!(cache.len(), 0);
        assert_eq!(cache.max_positions(), 128);
    }

    #[test]
    fn test_streaming_kv_cache_append() {
        let mut cache = StreamingKVCache::new(2, 16, 4, 32);
        let kv_dim = 4 * 32;
        let key = vec![1.0f32; kv_dim];
        let value = vec![2.0f32; kv_dim];

        cache.append(0, &key, &value);
        cache.append(1, &key, &value);

        assert_eq!(cache.len(), 1);
        assert!(!cache.is_empty());
    }

    #[test]
    fn test_streaming_kv_cache_append_multiple() {
        let mut cache = StreamingKVCache::new(2, 16, 2, 4);
        let kv_dim = 2 * 4;
        let key1 = vec![1.0f32; kv_dim];
        let value1 = vec![2.0f32; kv_dim];
        let key2 = vec![3.0f32; kv_dim];
        let value2 = vec![4.0f32; kv_dim];

        // First position
        cache.append(0, &key1, &value1);
        cache.append(1, &key1, &value1);
        assert_eq!(cache.len(), 1);

        // Second position
        cache.append(0, &key2, &value2);
        cache.append(1, &key2, &value2);
        assert_eq!(cache.len(), 2);
    }

    #[test]
    fn test_streaming_kv_cache_get_range() {
        let mut cache = StreamingKVCache::new(1, 16, 2, 4);
        let kv_dim = 2 * 4;
        let key = vec![1.5f32; kv_dim];
        let value = vec![2.5f32; kv_dim];

        cache.append(0, &key, &value);

        let (keys, values) = cache.get_range(0, 0, 1);
        assert_eq!(keys.len(), kv_dim);
        assert_eq!(values.len(), kv_dim);
        assert!((keys[0] - 1.5).abs() < 0.01);
        assert!((values[0] - 2.5).abs() < 0.01);
    }

    #[test]
    fn test_streaming_kv_cache_get_valid() {
        let mut cache = StreamingKVCache::new(1, 16, 2, 4);
        let kv_dim = 2 * 4;
        let key = vec![1.0f32; kv_dim];
        let value = vec![2.0f32; kv_dim];

        cache.append(0, &key, &value);
        cache.append(0, &key, &value);

        let (keys, values) = cache.get_valid(0);
        assert_eq!(keys.len(), 2 * kv_dim);
        assert_eq!(values.len(), 2 * kv_dim);
    }

    #[test]
    fn test_streaming_kv_cache_circular_buffer() {
        // Test that cache wraps around when full
        let mut cache = StreamingKVCache::new(1, 4, 1, 2);
        let kv_dim = 2;

        // Fill cache
        for i in 0..4 {
            let key = vec![i as f32; kv_dim];
            let value = vec![(i * 10) as f32; kv_dim];
            cache.append(0, &key, &value);
        }
        assert_eq!(cache.len(), 4);

        // Overflow: should wrap around
        let key = vec![100.0f32; kv_dim];
        let value = vec![200.0f32; kv_dim];
        cache.append(0, &key, &value);
        assert_eq!(cache.len(), 4); // Still 4, max_positions

        // First position should now have the overwritten value
        let (keys, _) = cache.get_range(0, 0, 1);
        assert!((keys[0] - 100.0).abs() < 0.01);
    }

    #[test]
    fn test_streaming_kv_cache_clear() {
        let mut cache = StreamingKVCache::new(1, 16, 2, 4);
        let kv_dim = 2 * 4;
        let key = vec![1.0f32; kv_dim];
        let value = vec![2.0f32; kv_dim];

        cache.append(0, &key, &value);
        assert_eq!(cache.len(), 1);

        cache.clear();
        assert_eq!(cache.len(), 0);
        assert!(cache.is_empty());
    }

    #[test]
    fn test_streaming_kv_cache_memory_bytes() {
        // 2 layers, 128 positions, 4 heads, 32 dim
        // kv_size = 128 * 4 * 32 = 16384
        // memory = 2 * 16384 * 2 * 4 = 262144 bytes = 256 KB
        let cache = StreamingKVCache::new(2, 128, 4, 32);
        assert_eq!(cache.memory_bytes(), 262_144);
    }

    #[test]
    fn test_streaming_kv_cache_memory_mb() {
        let cache = StreamingKVCache::new(2, 128, 4, 32);
        let mb = cache.memory_mb();
        assert!((mb - 0.25).abs() < 0.01); // 256 KB = 0.25 MB
    }

    #[test]
    #[should_panic(expected = "Layer index out of bounds")]
    fn test_streaming_kv_cache_invalid_layer() {
        let mut cache = StreamingKVCache::new(2, 16, 2, 4);
        let kv_dim = 2 * 4;
        let key = vec![1.0f32; kv_dim];
        let value = vec![2.0f32; kv_dim];
        cache.append(2, &key, &value); // layer 2 is out of bounds (0, 1)
    }

    #[test]
    #[should_panic(expected = "Key dimension mismatch")]
    fn test_streaming_kv_cache_key_dimension_mismatch() {
        let mut cache = StreamingKVCache::new(1, 16, 2, 4);
        let key = vec![1.0f32; 4]; // Wrong dimension (should be 8)
        let value = vec![2.0f32; 8];
        cache.append(0, &key, &value);
    }

    // =========================================================================
    // StreamingKVCacheFp16 Tests
    // =========================================================================

    #[test]
    fn test_streaming_kv_cache_fp16_new() {
        let cache = StreamingKVCacheFp16::new(4, 128, 8, 64);
        assert!(cache.is_empty());
        assert_eq!(cache.len(), 0);
        assert_eq!(cache.max_positions(), 128);
    }

    #[test]
    fn test_streaming_kv_cache_fp16_conversion() {
        // Test f32 -> f16 -> f32 round trip
        let val = 1.5f32;
        let fp16 = StreamingKVCacheFp16::f32_to_f16(val);
        let back = StreamingKVCacheFp16::f16_to_f32(fp16);
        assert!((back - val).abs() < 0.001);
    }

    #[test]
    fn test_streaming_kv_cache_fp16_append() {
        let mut cache = StreamingKVCacheFp16::new(2, 16, 4, 32);
        let kv_dim = 4 * 32;
        let key = vec![1.0f32; kv_dim];
        let value = vec![2.0f32; kv_dim];

        cache.append(0, &key, &value);
        cache.append(1, &key, &value);

        assert_eq!(cache.len(), 1);
        assert!(!cache.is_empty());
    }

    #[test]
    fn test_streaming_kv_cache_fp16_get_range_f32() {
        let mut cache = StreamingKVCacheFp16::new(1, 16, 2, 4);
        let kv_dim = 2 * 4;
        let key = vec![1.5f32; kv_dim];
        let value = vec![2.5f32; kv_dim];

        cache.append(0, &key, &value);

        let (keys, values) = cache.get_range_f32(0, 0, 1);
        assert_eq!(keys.len(), kv_dim);
        assert_eq!(values.len(), kv_dim);
        // FP16 has some loss of precision
        assert!((keys[0] - 1.5).abs() < 0.01);
        assert!((values[0] - 2.5).abs() < 0.01);
    }

    #[test]
    fn test_streaming_kv_cache_fp16_get_range_raw() {
        let mut cache = StreamingKVCacheFp16::new(1, 16, 2, 4);
        let kv_dim = 2 * 4;
        let key = vec![1.0f32; kv_dim];
        let value = vec![2.0f32; kv_dim];

        cache.append(0, &key, &value);

        let (keys_raw, values_raw) = cache.get_range_raw(0, 0, 1);
        assert_eq!(keys_raw.len(), kv_dim);
        assert_eq!(values_raw.len(), kv_dim);
    }

    #[test]
    fn test_streaming_kv_cache_fp16_get_valid_f32() {
        let mut cache = StreamingKVCacheFp16::new(1, 16, 2, 4);
        let kv_dim = 2 * 4;
        let key = vec![1.0f32; kv_dim];
        let value = vec![2.0f32; kv_dim];

        cache.append(0, &key, &value);
        cache.append(0, &key, &value);

        let (keys, values) = cache.get_valid_f32(0);
        assert_eq!(keys.len(), 2 * kv_dim);
        assert_eq!(values.len(), 2 * kv_dim);
    }

    #[test]
    fn test_streaming_kv_cache_fp16_clear() {
        let mut cache = StreamingKVCacheFp16::new(1, 16, 2, 4);
        let kv_dim = 2 * 4;
        let key = vec![1.0f32; kv_dim];
        let value = vec![2.0f32; kv_dim];

        cache.append(0, &key, &value);
        assert_eq!(cache.len(), 1);

        cache.clear();
        assert_eq!(cache.len(), 0);
        assert!(cache.is_empty());
    }

    #[test]
    fn test_streaming_kv_cache_fp16_memory_bytes() {
        // 2 layers, 128 positions, 4 heads, 32 dim
        // kv_size = 128 * 4 * 32 = 16384
        // memory = 2 * 16384 * 2 * 2 = 131072 bytes = 128 KB
        // (half of FP32 version)
        let cache = StreamingKVCacheFp16::new(2, 128, 4, 32);
        assert_eq!(cache.memory_bytes(), 131_072);
    }

    #[test]
    fn test_streaming_kv_cache_fp16_memory_mb() {
        let cache = StreamingKVCacheFp16::new(2, 128, 4, 32);
        let mb = cache.memory_mb();
        assert!((mb - 0.125).abs() < 0.01); // 128 KB = 0.125 MB
    }

    #[test]
    fn test_streaming_kv_cache_fp16_half_memory_of_fp32() {
        let fp32_cache = StreamingKVCache::new(4, 256, 8, 64);
        let fp16_cache = StreamingKVCacheFp16::new(4, 256, 8, 64);

        // FP16 should use exactly half the memory of FP32
        assert_eq!(fp16_cache.memory_bytes(), fp32_cache.memory_bytes() / 2);
    }

    #[test]
    fn test_streaming_kv_cache_fp16_circular_buffer() {
        let mut cache = StreamingKVCacheFp16::new(1, 4, 1, 2);
        let kv_dim = 2;

        // Fill cache
        for i in 0..4 {
            let key = vec![i as f32; kv_dim];
            let value = vec![(i * 10) as f32; kv_dim];
            cache.append(0, &key, &value);
        }
        assert_eq!(cache.len(), 4);

        // Overflow: should wrap around
        let key = vec![100.0f32; kv_dim];
        let value = vec![200.0f32; kv_dim];
        cache.append(0, &key, &value);
        assert_eq!(cache.len(), 4);

        // First position should now have the overwritten value
        let (keys, _) = cache.get_range_f32(0, 0, 1);
        assert!((keys[0] - 100.0).abs() < 0.1); // FP16 precision
    }

    #[test]
    #[should_panic(expected = "Layer index out of bounds")]
    fn test_streaming_kv_cache_fp16_invalid_layer() {
        let mut cache = StreamingKVCacheFp16::new(2, 16, 2, 4);
        let kv_dim = 2 * 4;
        let key = vec![1.0f32; kv_dim];
        let value = vec![2.0f32; kv_dim];
        cache.append(2, &key, &value);
    }

    #[test]
    #[should_panic(expected = "Value dimension mismatch")]
    fn test_streaming_kv_cache_fp16_value_dimension_mismatch() {
        let mut cache = StreamingKVCacheFp16::new(1, 16, 2, 4);
        let key = vec![1.0f32; 8];
        let value = vec![2.0f32; 4]; // Wrong dimension (should be 8)
        cache.append(0, &key, &value);
    }
}