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

//! Part 3: Quantization read/write tests, prefix cache, and ID tests
//!
//! Focus areas:
//! 1. Quantization write/read across block boundaries
//! 2. QuantizedKvPage token operations
//! 3. Prefix cache additional tests
//! 4. ID generation and hashing

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

    // =========================================================================
    // Quantization Write/Read Across Block Boundaries
    // =========================================================================

    #[test]
    fn test_q8_write_read_across_block_boundary() {
        let mut data = QuantizedKvData::new(KvQuantType::Q8, 16, 8, 64);

        // Write data that spans multiple quant blocks (block_size = 32)
        let test_data: Vec<f32> = (0..100).map(|i| i as f32 * 0.1).collect();
        data.write_keys(0, &test_data);

        let read = data.read_keys(0, 100);
        assert_eq!(read.len(), 100);

        // Q8 should preserve values with small error
        for (orig, read_val) in test_data.iter().zip(read.iter()) {
            assert!(
                (orig - read_val).abs() < 0.05,
                "Q8 error: {} vs {}",
                orig,
                read_val
            );
        }
    }

    #[test]
    fn test_q4_write_read_across_block_boundary() {
        let mut data = QuantizedKvData::new(KvQuantType::Q4, 16, 8, 64);

        let test_data: Vec<f32> = (0..100).map(|i| (i as f32 - 50.0) * 0.1).collect();
        data.write_values(0, &test_data);

        let read = data.read_values(0, 100);
        assert_eq!(read.len(), 100);

        // Q4 has larger error tolerance
        for (orig, read_val) in test_data.iter().zip(read.iter()) {
            assert!(
                (orig - read_val).abs() < 0.5,
                "Q4 error: {} vs {}",
                orig,
                read_val
            );
        }
    }

    #[test]
    fn test_read_with_offset_in_middle_of_block() {
        let mut data = QuantizedKvData::new(KvQuantType::FP32, 16, 8, 64);

        // Write full data
        let full_data: Vec<f32> = (0..128).map(|i| i as f32).collect();
        data.write_keys(0, &full_data);

        // Read from middle of block
        let read = data.read_keys(10, 20);
        assert_eq!(read.len(), 20);
        for i in 0..20 {
            assert_eq!(read[i], (i + 10) as f32);
        }
    }

    #[test]
    fn test_write_partial_at_end_of_data() {
        let mut data = QuantizedKvData::new(KvQuantType::FP32, 16, 8, 64);

        let total_size = 16 * 8 * 64; // 8192 elements
        let small_data = vec![99.0f32; 10];

        // Write near the end
        data.write_keys(total_size - 5, &small_data);

        // Only 5 elements should be written (truncated)
        let read = data.read_keys(total_size - 5, 10);
        assert_eq!(read.len(), 5);
        for val in &read {
            assert_eq!(*val, 99.0);
        }
    }

    // =========================================================================
    // QuantizedKvPage Token Position Operations
    // =========================================================================

    #[test]
    fn test_quantized_page_write_read_multiple_tokens() {
        let mut page = QuantizedKvPage::new(PageId::new(0), KvQuantType::FP32, 16, 8, 64);

        // Write keys/values for token 0
        let k0: Vec<f32> = (0..512).map(|i| i as f32 * 0.1).collect();
        let v0: Vec<f32> = (0..512).map(|i| -i as f32 * 0.1).collect();
        page.write_keys(0, &k0);
        page.write_values(0, &v0);
        page.num_tokens = 1;

        // Write keys/values for token 1
        let k1: Vec<f32> = (0..512).map(|i| (i + 1000) as f32 * 0.1).collect();
        let v1: Vec<f32> = (0..512).map(|i| -(i + 1000) as f32 * 0.1).collect();
        page.write_keys(1, &k1);
        page.write_values(1, &v1);
        page.num_tokens = 2;

        // Verify reads
        let rk0 = page.read_keys(0);
        let rv0 = page.read_values(0);
        let rk1 = page.read_keys(1);
        let rv1 = page.read_values(1);

        assert_eq!(rk0, k0);
        assert_eq!(rv0, v0);
        assert_eq!(rk1, k1);
        assert_eq!(rv1, v1);
    }

    #[test]
    fn test_quantized_page_remaining_capacity() {
        let mut page = QuantizedKvPage::new(PageId::new(0), KvQuantType::Q8, 16, 8, 64);

        assert_eq!(page.remaining_capacity(), 16);
        page.num_tokens = 5;
        assert_eq!(page.remaining_capacity(), 11);
        page.num_tokens = 16;
        assert_eq!(page.remaining_capacity(), 0);
        page.num_tokens = 20; // Over capacity
        assert_eq!(page.remaining_capacity(), 0);
    }

    // =========================================================================
    // Prefix Cache Additional Tests
    // =========================================================================

    #[test]
    #[ignore = "Test expectation needs adjustment"]
    fn test_prefix_cache_eviction_with_zero_ref_count() {
        let mut cache = PrefixCache::new(2);

        // Insert first prefix
        let mut p1 = CachedPrefix::new(1, 10, vec![]);
        p1.ref_count = 0; // Evictable
        cache.insert(p1);

        // Insert second prefix (also evictable)
        let mut p2 = CachedPrefix::new(2, 20, vec![]);
        p2.ref_count = 0;
        cache.insert(p2);

        assert_eq!(cache.len(), 2);

        // Insert third - should evict LRU
        let p3 = CachedPrefix::new(3, 30, vec![]);
        let success = cache.insert(p3);
        assert!(success);
        assert_eq!(cache.len(), 2);

        // First prefix (hash=1) should be evicted (LRU)
        assert!(!cache.contains(1));
        assert!(cache.contains(2) || cache.contains(3));
    }

    #[test]
    fn test_prefix_cache_update_access_time() {
        let mut cache = PrefixCache::new(10);

        // Insert two prefixes
        let p1 = CachedPrefix::new(1, 10, vec![]);
        let p2 = CachedPrefix::new(2, 20, vec![]);
        cache.insert(p1);
        cache.insert(p2);

        // Lookup hash 1 multiple times to update access time
        cache.lookup(1);
        cache.lookup(1);
        cache.lookup(1);

        let stats = cache.stats();
        assert_eq!(stats.hits, 3);
    }

    #[test]
    fn test_prefix_cache_add_ref_nonexistent() {
        let mut cache = PrefixCache::new(10);

        // Add ref to nonexistent entry
        let result = cache.add_ref(99999);
        assert!(!result);
    }

    #[test]
    fn test_prefix_cache_remove_ref_nonexistent() {
        let mut cache = PrefixCache::new(10);

        // Remove ref from nonexistent entry
        let result = cache.remove_ref(99999);
        assert!(!result);
    }

    #[test]
    fn test_prefix_cache_zero_capacity() {
        let mut cache = PrefixCache::new(0);
        assert_eq!(cache.utilization(), 0.0);

        // Can't insert anything
        let p = CachedPrefix::new(1, 10, vec![]);
        let success = cache.insert(p);
        assert!(!success);
    }

    // =========================================================================
    // ID Generation and Hashing
    // =========================================================================

    #[test]
    fn test_seq_id_monotonic() {
        let mut prev = SeqId::new().value();
        for _ in 0..100 {
            let current = SeqId::new().value();
            assert!(current > prev);
            prev = current;
        }
    }

    #[test]
    fn test_page_id_serialization() {
        let id = PageId::new(42);
        let json = serde_json::to_string(&id).expect("serialize");
        let parsed: PageId = serde_json::from_str(&json).expect("deserialize");
        assert_eq!(parsed.value(), 42);
    }

    #[test]
    fn test_seq_id_serialization() {
        let id = SeqId::new();
        let json = serde_json::to_string(&id).expect("serialize");
        let parsed: SeqId = serde_json::from_str(&json).expect("deserialize");
        assert_eq!(parsed.value(), id.value());
    }

    #[test]
    fn test_prefix_hash_single_token() {
        let hash1 = compute_prefix_hash(&[0]);
        let hash2 = compute_prefix_hash(&[1]);
        let hash3 = compute_prefix_hash(&[u32::MAX]);

        // All should be different
        assert_ne!(hash1, hash2);
        assert_ne!(hash2, hash3);
        assert_ne!(hash1, hash3);
    }

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

    #[test]
    fn test_q8_quantize_extreme_values() {
        let mut values = [0.0f32; KV_QUANT_BLOCK_SIZE];
        // Set some extreme values
        values[0] = 1000.0;
        values[1] = -1000.0;
        values[2] = 0.0001;
        values[3] = -0.0001;

        let block = Q8KvBlock::quantize(&values);
        let restored = block.dequantize();

        // Extremes should be preserved (within quantization error)
        assert!((restored[0] - 1000.0).abs() < 10.0);
        assert!((restored[1] + 1000.0).abs() < 10.0);
    }

    #[test]
    fn test_q4_quantize_extreme_values() {
        let mut values = [0.0f32; KV_QUANT_BLOCK_SIZE];
        // Set some extreme values
        values[0] = 100.0;
        values[1] = -100.0;

        let block = Q4KvBlock::quantize(&values);
        let restored = block.dequantize();

        // Q4 has higher quantization error
        assert!((restored[0] - 100.0).abs() < 20.0);
        assert!((restored[1] + 100.0).abs() < 20.0);
    }

    #[test]
    fn test_quantized_kv_data_q4_read_write() {
        let mut data = QuantizedKvData::new(KvQuantType::Q4, 16, 8, 64);

        let test_values: Vec<f32> = (0..64).map(|i| (i as f32 - 32.0) * 0.1).collect();
        data.write_values(0, &test_values);
        let read_values = data.read_values(0, 64);

        // Q4 has larger error tolerance
        for (orig, read) in test_values.iter().zip(read_values.iter()) {
            assert!((orig - read).abs() < 0.5);
        }
    }

    #[test]
    fn test_quantized_cache_total_pages() {
        let cache = QuantizedPagedKvCache::new(50, 16, 8, 64, KvQuantType::Q8);
        assert_eq!(cache.total_pages(), 50);
    }

    #[test]
    fn test_kv_page_data_clone() {
        let mut page = KvPage::new(PageId::new(0), 16, 8, 64);
        page.keys[0] = 1.0;
        page.values[0] = 2.0;
        page.num_tokens = 5;
        page.ref_count = 2;

        let cloned = page.clone();
        assert_eq!(cloned.keys[0], 1.0);
        assert_eq!(cloned.values[0], 2.0);
        assert_eq!(cloned.num_tokens, 5);
        assert_eq!(cloned.ref_count, 2);
    }

    #[test]
    fn test_cached_prefix_clone() {
        let prefix = CachedPrefix::new(12345, 100, vec![PageId::new(0), PageId::new(1)]);
        let cloned = prefix.clone();

        assert_eq!(cloned.hash, 12345);
        assert_eq!(cloned.num_tokens, 100);
        assert_eq!(cloned.page_ids.len(), 2);
    }

    #[test]
    fn test_fragmentation_stats_default() {
        let stats = FragmentationStats::default();
        assert_eq!(stats.holes, 0);
        assert_eq!(stats.wasted_capacity, 0);
        assert_eq!(stats.fragmentation_ratio, 0.0);
        assert_eq!(stats.largest_free_region, 0);
        assert_eq!(stats.avg_tokens_per_page, 0.0);
    }

    #[test]
    fn test_paged_cache_stats_default() {
        let stats = PagedCacheStats::default();
        assert_eq!(stats.sequences_allocated, 0);
        assert_eq!(stats.sequences_freed, 0);
        assert_eq!(stats.pages_allocated, 0);
        assert_eq!(stats.pages_freed, 0);
        assert_eq!(stats.active_sequences, 0);
        assert_eq!(stats.used_pages, 0);
        assert_eq!(stats.cow_operations, 0);
        assert_eq!(stats.defrag_operations, 0);
        assert_eq!(stats.pages_moved, 0);
    }
}