boostr 0.1.0

ML framework built on numr - attention, quantization, model architectures
Documentation 
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

use boostr::inference::memory::CpuBlockAllocator;
use boostr::inference::prefix_cache::{
    CacheResult, PrefixCache, PrefixCacheConfig, SequencePrefixState,
};

fn create_cache() -> PrefixCache<CpuBlockAllocator> {
    let allocator = CpuBlockAllocator::new(100, 16);
    let config = PrefixCacheConfig {
        enabled: true,
        max_cached_blocks: 50,
        min_prefix_tokens: 4,
        block_size: 16,
    };
    PrefixCache::new(allocator, config)
}

#[test]
fn test_prefix_cache_create() {
    let cache = create_cache();
    assert!(cache.config().enabled);
    assert_eq!(cache.config().block_size, 16);
}

#[test]
fn test_cache_miss() {
    let mut cache = create_cache();
    let tokens: Vec<u32> = (0..32).collect();

    let result = cache.get_or_allocate_blocks(1, &tokens).unwrap();

    assert!(!result.is_hit());
    assert_eq!(result.blocks().len(), 2);

    let stats = cache.stats();
    assert_eq!(stats.cache_misses, 1);
    assert_eq!(stats.blocks_allocated, 2);
}

#[test]
fn test_cache_hit() {
    let mut cache = create_cache();
    let tokens: Vec<u32> = (0..32).collect();

    let result1 = cache.get_or_allocate_blocks(1, &tokens).unwrap();
    assert!(!result1.is_hit());

    let result2 = cache.get_or_allocate_blocks(2, &tokens).unwrap();
    assert!(result2.is_hit());
    assert_eq!(result2.cached_count(), 2);

    assert_eq!(result1.blocks(), result2.blocks());

    let stats = cache.stats();
    assert_eq!(stats.cache_hits, 1);
    assert_eq!(stats.cache_misses, 1);
    assert_eq!(stats.blocks_from_cache, 2);
}

#[test]
fn test_partial_cache_hit() {
    let mut cache = create_cache();

    let tokens1: Vec<u32> = (0..32).collect();
    cache.get_or_allocate_blocks(1, &tokens1).unwrap();

    let tokens2: Vec<u32> = (0..48).collect();
    let result = cache.get_or_allocate_blocks(2, &tokens2).unwrap();

    assert!(result.is_hit());
    assert_eq!(result.blocks().len(), 3);
    match result {
        CacheResult::Hit {
            cached_blocks,
            new_blocks,
            ..
        } => {
            assert_eq!(cached_blocks, 2);
            assert_eq!(new_blocks, 1);
        }
        _ => panic!("Expected hit"),
    }
}

#[test]
fn test_different_prefix_no_hit() {
    let mut cache = create_cache();

    let tokens1: Vec<u32> = (0..32).collect();
    cache.get_or_allocate_blocks(1, &tokens1).unwrap();

    let tokens2: Vec<u32> = (100..132).collect();
    let result = cache.get_or_allocate_blocks(2, &tokens2).unwrap();

    assert!(!result.is_hit());
}

#[test]
fn test_release_blocks() {
    let mut cache = create_cache();
    let tokens: Vec<u32> = (0..32).collect();

    let result = cache.get_or_allocate_blocks(1, &tokens).unwrap();
    let blocks = result.blocks().to_vec();

    cache.release_blocks(1, &blocks).unwrap();
    assert!(cache.has_cached_prefix(&tokens));
}

#[test]
fn test_force_free_blocks() {
    let mut cache = create_cache();
    let tokens: Vec<u32> = (0..32).collect();

    let result = cache.get_or_allocate_blocks(1, &tokens).unwrap();
    let blocks = result.blocks().to_vec();

    cache.force_free_blocks(&blocks).unwrap();
    assert!(!cache.has_cached_prefix(&tokens));
}

#[test]
fn test_has_cached_prefix() {
    let mut cache = create_cache();
    let tokens: Vec<u32> = (0..32).collect();

    assert!(!cache.has_cached_prefix(&tokens));
    cache.get_or_allocate_blocks(1, &tokens).unwrap();
    assert!(cache.has_cached_prefix(&tokens));
}

#[test]
fn test_cached_block_count() {
    let mut cache = create_cache();
    let tokens: Vec<u32> = (0..32).collect();

    assert_eq!(cache.cached_block_count(&tokens), 0);
    cache.get_or_allocate_blocks(1, &tokens).unwrap();
    assert_eq!(cache.cached_block_count(&tokens), 2);

    let longer: Vec<u32> = (0..48).collect();
    assert_eq!(cache.cached_block_count(&longer), 2);
}

#[test]
fn test_min_prefix_tokens() {
    let mut cache = create_cache();

    let short_tokens: Vec<u32> = vec![1, 2, 3];
    let result1 = cache.get_or_allocate_blocks(1, &short_tokens).unwrap();
    let result2 = cache.get_or_allocate_blocks(2, &short_tokens).unwrap();

    assert!(!result1.is_hit());
    assert!(!result2.is_hit());
}

#[test]
fn test_cache_disabled() {
    let allocator = CpuBlockAllocator::new(100, 16);
    let config = PrefixCacheConfig {
        enabled: false,
        ..Default::default()
    };
    let mut cache = PrefixCache::new(allocator, config);

    let tokens: Vec<u32> = (0..32).collect();
    cache.get_or_allocate_blocks(1, &tokens).unwrap();
    let result = cache.get_or_allocate_blocks(2, &tokens).unwrap();
    assert!(!result.is_hit());
}

#[test]
fn test_eviction() {
    let allocator = CpuBlockAllocator::new(10, 16);
    let config = PrefixCacheConfig {
        enabled: true,
        max_cached_blocks: 10,
        min_prefix_tokens: 4,
        block_size: 16,
    };
    let mut cache = PrefixCache::new(allocator, config);

    for i in 0..4 {
        let tokens: Vec<u32> = ((i * 100)..((i * 100) + 32)).map(|x| x as u32).collect();
        let result = cache.get_or_allocate_blocks(i as u64, &tokens).unwrap();
        cache.release_blocks(i as u64, result.blocks()).unwrap();
    }

    let tokens: Vec<u32> = (1000..1032).collect();
    let result = cache.get_or_allocate_blocks(100, &tokens);
    assert!(result.is_ok());
}

#[test]
fn test_stats() {
    let mut cache = create_cache();
    let tokens: Vec<u32> = (0..32).collect();

    cache.get_or_allocate_blocks(1, &tokens).unwrap();
    cache.get_or_allocate_blocks(2, &tokens).unwrap();
    cache.get_or_allocate_blocks(3, &tokens).unwrap();

    let stats = cache.stats();
    assert_eq!(stats.total_lookups, 3);
    assert_eq!(stats.cache_hits, 2);
    assert_eq!(stats.cache_misses, 1);
    assert_eq!(stats.blocks_from_cache, 4);
    assert_eq!(stats.blocks_allocated, 2);
    assert!(stats.hit_rate > 0.6);
}

#[test]
fn test_reset() {
    let mut cache = create_cache();
    let tokens: Vec<u32> = (0..32).collect();

    let result = cache.get_or_allocate_blocks(1, &tokens).unwrap();
    let blocks = result.blocks().to_vec();
    cache.release_blocks(1, &blocks).ok();

    assert!(cache.has_cached_prefix(&tokens));
    cache.reset().unwrap();
    assert!(!cache.has_cached_prefix(&tokens));
    assert_eq!(cache.stats().total_lookups, 0);
}

#[test]
fn test_sequence_prefix_state() {
    let mut cache = create_cache();
    let tokens: Vec<u32> = (0..32).collect();

    let result1 = cache.get_or_allocate_blocks(1, &tokens).unwrap();
    let state1 = SequencePrefixState::from_cache_result(&result1, tokens.len());
    assert_eq!(state1.cached_count(), 0);
    assert_eq!(state1.new_count(), 2);

    let result2 = cache.get_or_allocate_blocks(2, &tokens).unwrap();
    let state2 = SequencePrefixState::from_cache_result(&result2, tokens.len());
    assert_eq!(state2.cached_count(), 2);
    assert_eq!(state2.new_count(), 0);
}

#[test]
fn test_refcount_tracking() {
    let mut cache = create_cache();
    let tokens: Vec<u32> = (0..32).collect();

    let result1 = cache.get_or_allocate_blocks(1, &tokens).unwrap();
    cache.get_or_allocate_blocks(2, &tokens).unwrap();
    cache.get_or_allocate_blocks(3, &tokens).unwrap();

    let blocks = result1.blocks().to_vec();

    cache.release_blocks(1, &blocks).unwrap();
    assert!(cache.has_cached_prefix(&tokens));

    cache.release_blocks(2, &blocks).unwrap();
    assert!(cache.has_cached_prefix(&tokens));

    cache.release_blocks(3, &blocks).unwrap();
    assert!(cache.has_cached_prefix(&tokens));
}

/// max_cached_blocks must be enforced: once sequences release their blocks
/// (ref_count → 0), new allocations must evict LRU entries to stay within cap.
#[test]
fn test_max_cached_blocks_enforced() {
    let block_size = 4;
    // Physical pool: 20 blocks; cache cap: 4 blocks.
    let allocator = CpuBlockAllocator::new(20, block_size);
    let config = PrefixCacheConfig {
        enabled: true,
        max_cached_blocks: 4,
        min_prefix_tokens: 1,
        block_size,
    };
    let mut cache = PrefixCache::new(allocator, config);

    // Allocate 4 sequences and immediately release them so their blocks are
    // eligible for eviction (ref_count == 0).
    let mut all_blocks = Vec::new();
    for seq_id in 1u64..=4 {
        let tokens: Vec<u32> = (seq_id as u32 * 10..(seq_id as u32 * 10 + 4)).collect();
        let result = cache.get_or_allocate_blocks(seq_id, &tokens).unwrap();
        let blocks = result.blocks().to_vec();
        cache.release_blocks(seq_id, &blocks).unwrap();
        all_blocks.push(blocks);
    }

    // At this point the cache holds exactly 4 entries, all ref_count == 0.
    assert_eq!(cache.stats().cached_blocks, 4);

    // A 5th allocation must evict the LRU entry, keeping the count at ≤ 4.
    let tokens5: Vec<u32> = (50..54).collect();
    cache.get_or_allocate_blocks(5, &tokens5).unwrap();

    let stats = cache.stats();
    assert!(
        stats.cached_blocks <= 4,
        "cached_blocks {} exceeds max_cached_blocks 4",
        stats.cached_blocks
    );
    assert!(
        stats.blocks_evicted >= 1,
        "at least one block must have been evicted"
    );
}

/// has_cached_prefix and cached_block_count must validate token content,
/// not just hash presence, to guard against hash collisions.
#[test]
fn test_cached_block_count_validates_tokens() {
    let block_size = 4;
    let allocator = CpuBlockAllocator::new(20, block_size);
    let config = PrefixCacheConfig {
        enabled: true,
        max_cached_blocks: 100,
        min_prefix_tokens: 1,
        block_size,
    };
    let mut cache = PrefixCache::new(allocator, config);

    let tokens: Vec<u32> = vec![1, 2, 3, 4];
    cache.get_or_allocate_blocks(1, &tokens).unwrap();

    // Correct tokens → should find the cached block.
    assert!(cache.has_cached_prefix(&tokens));
    assert_eq!(cache.cached_block_count(&tokens), 1);

    // Different tokens with the same prefix length → should NOT hit.
    let different: Vec<u32> = vec![9, 9, 9, 9];
    // (These are very unlikely to collide; the test verifies the token check.)
    if !cache.has_cached_prefix(&different) {
        assert_eq!(cache.cached_block_count(&different), 0);
    }
    // If by extreme chance hashes collide, the token check must still return 0.
    // We can't force a collision in a unit test, but we verify the path compiles
    // and the non-colliding case is correct.
}