trustformers-core 0.2.1

Core traits and utilities for TrustformeRS
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

/// Tests for generation cache types: KVCache and Beam
#[cfg(test)]
mod tests {
    use crate::generation::cache::{Beam, KVCache};
    use crate::tensor::Tensor;
    use std::sync::Arc;

    // ---- KVCache tests ----

    #[test]
    fn test_kvcache_new_is_empty() {
        let cache = KVCache::new();
        assert!(cache.keys.is_empty());
        assert!(cache.values.is_empty());
        assert_eq!(cache.seq_len, 0);
    }

    #[test]
    fn test_kvcache_default_matches_new() {
        let a = KVCache::new();
        let b = KVCache::default();
        assert_eq!(a.seq_len, b.seq_len);
        assert_eq!(a.keys.len(), b.keys.len());
    }

    #[test]
    fn test_kvcache_clear_resets_all_fields() {
        let mut cache = KVCache::new();
        // Start with seq_len > 0 by manipulating directly (no append needed for clearing test)
        cache.seq_len = 5;
        cache.clear();
        assert_eq!(cache.seq_len, 0);
        assert!(cache.keys.is_empty());
        assert!(cache.values.is_empty());
    }

    #[test]
    fn test_kvcache_get_layer_out_of_bounds_returns_none() {
        let cache = KVCache::new();
        let result = cache.get_layer(0);
        assert!(result.is_none());
    }

    #[test]
    fn test_kvcache_get_layer_large_index_returns_none() {
        let cache = KVCache::new();
        assert!(cache.get_layer(999).is_none());
    }

    #[test]
    fn test_kvcache_seq_len_starts_at_zero() {
        let cache = KVCache::new();
        assert_eq!(cache.seq_len, 0);
    }

    #[test]
    fn test_kvcache_clear_idempotent() {
        let mut cache = KVCache::new();
        cache.clear();
        cache.clear();
        assert_eq!(cache.seq_len, 0);
        assert!(cache.keys.is_empty());
    }

    // ---- Beam tests ----

    #[test]
    fn test_beam_new_creates_unfinished_beam() {
        let beam = Beam::new(vec![1, 2, 3], -0.5);
        assert_eq!(beam.tokens, vec![1, 2, 3]);
        assert!((beam.score - (-0.5)).abs() < 1e-6);
        assert!(!beam.finished);
        assert!(beam.cache.is_none());
    }

    #[test]
    fn test_beam_new_empty_tokens() {
        let beam = Beam::new(vec![], 0.0);
        assert!(beam.tokens.is_empty());
        assert_eq!(beam.score, 0.0);
        assert!(!beam.finished);
    }

    #[test]
    fn test_beam_extend_appends_token_and_accumulates_score() {
        let beam = Beam::new(vec![10, 20], -1.0);
        let extended = beam.extend(30, -0.5);
        assert_eq!(extended.tokens, vec![10, 20, 30]);
        assert!((extended.score - (-1.5)).abs() < 1e-6);
        assert!(!extended.finished);
    }

    #[test]
    fn test_beam_extend_does_not_mutate_original() {
        let beam = Beam::new(vec![1], 0.0);
        let _extended = beam.extend(2, -1.0);
        assert_eq!(beam.tokens, vec![1]);
        assert_eq!(beam.score, 0.0);
    }

    #[test]
    fn test_beam_finalize_sets_finished_flag() {
        let mut beam = Beam::new(vec![5], -0.3);
        assert!(!beam.finished);
        beam.finalize();
        assert!(beam.finished);
    }

    #[test]
    fn test_beam_finalize_idempotent() {
        let mut beam = Beam::new(vec![1], 0.0);
        beam.finalize();
        beam.finalize();
        assert!(beam.finished);
    }

    #[test]
    fn test_beam_normalized_score_single_token() {
        let beam = Beam::new(vec![42], -2.0);
        let normalized = beam.get_normalized_score();
        assert!((normalized - (-2.0)).abs() < 1e-6);
    }

    #[test]
    fn test_beam_normalized_score_multiple_tokens() {
        let beam = Beam::new(vec![1, 2, 3, 4], -8.0);
        let normalized = beam.get_normalized_score();
        assert!((normalized - (-2.0)).abs() < 1e-6);
    }

    #[test]
    fn test_beam_normalized_score_empty_tokens_returns_zero() {
        let beam = Beam::new(vec![], 0.0);
        let normalized = beam.get_normalized_score();
        assert_eq!(normalized, 0.0);
    }

    #[test]
    fn test_beam_extend_chain() {
        let mut beam = Beam::new(vec![], 0.0);
        // Use LCG for deterministic "random" scores
        let mut s = 42u64;
        for i in 0..5usize {
            s = s.wrapping_mul(6364136223846793005).wrapping_add(1442695040888963407);
            let score = -((s % 1000) as f32) / 1000.0;
            beam = beam.extend(i, score);
        }
        assert_eq!(beam.tokens.len(), 5);
    }

    #[test]
    fn test_beam_clone_is_independent() {
        let beam = Beam::new(vec![1, 2], -0.5);
        let mut cloned = beam.clone();
        cloned.finalize();
        assert!(!beam.finished);
        assert!(cloned.finished);
    }

    #[test]
    fn test_beam_extend_preserves_cache_none() {
        let beam = Beam::new(vec![1], 0.0);
        let extended = beam.extend(2, -1.0);
        assert!(extended.cache.is_none());
    }

    #[test]
    fn test_beam_normalized_score_positive() {
        let beam = Beam::new(vec![1, 2], 4.0);
        let normalized = beam.get_normalized_score();
        assert!((normalized - 2.0).abs() < 1e-6);
    }

    #[test]
    fn test_beam_score_accumulation_zero() {
        let beam = Beam::new(vec![1], 0.0);
        let extended = beam.extend(2, 0.0);
        assert_eq!(extended.score, 0.0);
    }

    #[test]
    fn test_kvcache_multiple_clears() {
        let mut cache = KVCache::new();
        for _ in 0..10 {
            cache.clear();
        }
        assert_eq!(cache.seq_len, 0);
    }

    #[test]
    fn test_kvcache_push_layer_does_not_move_the_sequence_forward() {
        // Regression: `append` used to bump `seq_len` once per *layer*, so a
        // 32-layer model reported 32 cached positions after a single token.
        let mut cache = KVCache::new();
        for _ in 0..4 {
            let key = Tensor::from_vec(vec![0.0_f32; 2], &[2]).expect("tensor");
            let value = Tensor::from_vec(vec![0.0_f32; 2], &[2]).expect("tensor");
            cache.push_layer(key, value).expect("push layer");
        }
        assert_eq!(cache.num_layers(), 4);
        assert_eq!(cache.seq_len, 0, "layers are not token positions");

        cache.advance(1);
        assert_eq!(cache.seq_len, 1);
    }

    #[test]
    fn test_kvcache_set_layer_out_of_range_is_error() {
        let mut cache = KVCache::new();
        let key = Tensor::from_vec(vec![0.0_f32], &[1]).expect("tensor");
        let value = Tensor::from_vec(vec![0.0_f32], &[1]).expect("tensor");
        assert!(cache.set_layer(0, key, value).is_err());
    }

    #[test]
    fn test_beam_extend_shares_the_cache_instead_of_deep_copying() {
        // Regression: `extend` deep-cloned every cached key/value tensor of
        // every layer for every candidate expansion.
        let mut cache = KVCache::new();
        let key = Tensor::from_vec(vec![1.0_f32; 64], &[64]).expect("tensor");
        let value = Tensor::from_vec(vec![2.0_f32; 64], &[64]).expect("tensor");
        cache.push_layer(key, value).expect("push layer");
        cache.advance(1);

        let beam = Beam::new(vec![1], 0.0).with_cache(cache);
        let child = beam.extend(2, -0.5);

        let parent_cache = beam.cache.as_ref().expect("parent cache");
        let child_cache = child.cache.as_ref().expect("child cache");
        assert!(
            Arc::ptr_eq(parent_cache, child_cache),
            "extend must share the cache allocation"
        );
        assert_eq!(child.tokens, vec![1, 2]);
        assert!((child.score - (-0.5)).abs() < 1e-6);
    }

    #[test]
    fn test_beam_cache_mut_is_copy_on_write() {
        let mut cache = KVCache::new();
        cache.advance(3);
        let beam = Beam::new(vec![1], 0.0).with_cache(cache);
        let mut child = beam.extend(2, 0.0);

        child.cache_mut().expect("cache").advance(1);

        assert_eq!(beam.cache.as_ref().expect("parent").seq_len, 3);
        assert_eq!(child.cache.as_ref().expect("child").seq_len, 4);
        assert!(!Arc::ptr_eq(
            beam.cache.as_ref().expect("parent"),
            child.cache.as_ref().expect("child")
        ));
    }

    #[test]
    fn test_beam_length_normalized_score_uses_the_penalty() {
        let short = Beam::new(vec![1, 2], -2.0);
        let long = Beam::new(vec![1, 2, 3, 4], -4.0);
        assert!(short.length_normalized_score(0.5) > long.length_normalized_score(0.5));
        assert!(long.length_normalized_score(2.0) > short.length_normalized_score(2.0));
    }
}