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


// ============================================================================
// Tests
// ============================================================================

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

    #[test]
    fn test_should_accept_ratio_exactly_half() {
        let draft = MockModel::new(100, 1);
        let target = MockModel::new(100, 1);
        let decoder = SpeculativeDecoder::new(draft, target, 4).expect("create");

        // We need ratio = 0.5 exactly
        // target_prob / draft_prob = 0.5
        // exp(target_log) / exp(draft_log) = 0.5
        // exp(target_log - draft_log) = 0.5
        // target_log - draft_log = ln(0.5) ≈ -0.693
        // If draft_log = -1.0, target_log = -1.693
        let draft_prob = TokenProb::new(5, -1.0);
        let target_prob = TokenProb::new(10, -1.0 + (-0.5_f32).ln());

        // ratio = 0.5, condition is ratio > 0.5, so should NOT accept
        // But ratio >= 1.0 OR ratio > 0.5, so 0.5 exactly should NOT match > 0.5
        assert!(!decoder.should_accept(&draft_prob, &target_prob));
    }

    #[test]
    fn test_should_accept_ratio_below_half() {
        let draft = MockModel::new(100, 1);
        let target = MockModel::new(100, 1);
        let decoder = SpeculativeDecoder::new(draft, target, 4).expect("create");

        // ratio needs to be < 0.5
        // target_prob / draft_prob < 0.5
        // If draft is high and target is low
        let draft_prob = TokenProb::new(5, 0.0); // prob = 1.0
        let target_prob = TokenProb::new(10, -2.0); // prob ≈ 0.135

        // ratio = 0.135 / 1.0 = 0.135 < 0.5, should NOT accept
        assert!(!decoder.should_accept(&draft_prob, &target_prob));
    }

    // === Error Handling Edge Cases ===

    #[test]
    fn test_speculative_error_verification_failed_fields() {
        let err = SpeculativeError::VerificationFailed { position: 42 };
        if let SpeculativeError::VerificationFailed { position } = err {
            assert_eq!(position, 42);
        } else {
            panic!("Expected VerificationFailed variant");
        }
    }

    #[test]
    fn test_speculative_error_draft_model_error_message() {
        let err = SpeculativeError::DraftModelError("custom error message".to_string());
        let display = format!("{}", err);
        assert!(display.contains("Draft model error"));
        assert!(display.contains("custom error message"));
    }

    #[test]
    fn test_speculative_error_target_model_error_message() {
        let err = SpeculativeError::TargetModelError("target failed".to_string());
        let display = format!("{}", err);
        assert!(display.contains("Target model error"));
        assert!(display.contains("target failed"));
    }

    #[test]
    fn test_speculative_error_invalid_spec_length_message() {
        let err = SpeculativeError::InvalidSpecLength(100);
        let display = format!("{}", err);
        assert!(display.contains("Invalid speculation length"));
        assert!(display.contains("100"));
    }

    // === Decoder with Rejection Scenarios ===

    /// Model where draft and target always disagree with very low target probability
    struct AlwaysRejectModel {
        token: u32,
        log_prob: f32,
    }

    impl SpeculativeModel for AlwaysRejectModel {
        fn forward(&self, _input_ids: &[u32]) -> Result<Vec<f32>, SpeculativeError> {
            Ok(vec![0.0; 100])
        }

        fn sample(&self, _logits: &[f32]) -> Result<TokenProb, SpeculativeError> {
            Ok(TokenProb::new(self.token, self.log_prob))
        }

        fn vocab_size(&self) -> usize {
            100
        }

        fn eos_token(&self) -> u32 {
            0
        }
    }

    #[test]
    fn test_speculative_decoder_immediate_rejection() {
        // Draft has token 5 with high prob, target has token 10 with very low prob
        let draft = AlwaysRejectModel {
            token: 5,
            log_prob: 0.0, // prob = 1.0
        };
        let target = AlwaysRejectModel {
            token: 10,
            log_prob: -10.0, // prob ≈ 0.00005
        };

        let mut decoder = SpeculativeDecoder::new(draft, target, 4).expect("create");
        let result = decoder.decode_iteration(&[1]).expect("decode");

        // First token should be rejected and resampled
        assert!(result.resampled_token.is_some());
        // Only 1 token accepted (the resampled one)
        assert_eq!(result.num_accepted, 1);
    }

    // === Stats Accumulation Verification ===

    #[test]
    fn test_speculative_stats_time_accumulation() {
        let mut stats = SpeculativeStats::default();

        stats.record_iteration(4, 3, 1.5, 15.0);
        assert!((stats.draft_time_ms - 1.5).abs() < 0.001);
        assert!((stats.target_time_ms - 15.0).abs() < 0.001);

        stats.record_iteration(4, 4, 2.0, 20.0);
        assert!((stats.draft_time_ms - 3.5).abs() < 0.001);
        assert!((stats.target_time_ms - 35.0).abs() < 0.001);
    }

    #[test]
    fn test_speculative_stats_avg_spec_length_calculation() {
        let mut stats = SpeculativeStats::default();

        stats.record_iteration(2, 2, 1.0, 10.0);
        assert!((stats.avg_spec_length - 2.0).abs() < 0.001);

        stats.record_iteration(6, 4, 1.0, 10.0);
        // Total speculated = 8, iterations = 2, avg = 4.0
        assert!((stats.avg_spec_length - 4.0).abs() < 0.001);

        stats.record_iteration(4, 3, 1.0, 10.0);
        // Total speculated = 12, iterations = 3, avg = 4.0
        assert!((stats.avg_spec_length - 4.0).abs() < 0.001);
    }

    // === Config with Custom max_spec_length ===

    #[test]
    fn test_speculative_config_custom_max_spec_length() {
        let mut config = SpeculativeConfig {
            spec_length: 10,
            min_acceptance_rate: 0.5,
            adaptive: true,
            max_spec_length: 12,
        };

        config.adapt_spec_length(0.95);
        assert_eq!(config.spec_length, 11);

        config.adapt_spec_length(0.95);
        assert_eq!(config.spec_length, 12);

        config.adapt_spec_length(0.95);
        assert_eq!(config.spec_length, 12); // Capped at max
    }

    // === Deserialization Tests ===

    #[test]
    fn test_speculative_stats_deserialize_from_json() {
        let json = r#"{
            "iterations": 5,
            "tokens_speculated": 20,
            "tokens_accepted": 15,
            "acceptance_rate": 0.75,
            "avg_spec_length": 4.0,
            "time_saved_ms": 50.0,
            "draft_time_ms": 5.0,
            "target_time_ms": 50.0
        }"#;

        let stats: SpeculativeStats = serde_json::from_str(json).expect("deserialize");
        assert_eq!(stats.iterations, 5);
        assert_eq!(stats.tokens_speculated, 20);
        assert_eq!(stats.tokens_accepted, 15);
        assert!((stats.acceptance_rate - 0.75).abs() < 0.001);
    }

    #[test]
    fn test_speculative_config_deserialize_from_json() {
        let json = r#"{
            "spec_length": 8,
            "min_acceptance_rate": 0.6,
            "adaptive": false,
            "max_spec_length": 16
        }"#;

        let config: SpeculativeConfig = serde_json::from_str(json).expect("deserialize");
        assert_eq!(config.spec_length, 8);
        assert!((config.min_acceptance_rate - 0.6).abs() < 0.001);
        assert!(!config.adaptive);
        assert_eq!(config.max_spec_length, 16);
    }

    // === Decoder Stats Integration ===

    #[test]
    fn test_speculative_decoder_stats_after_multiple_iterations() {
        let draft = MockModel::new(100, 1);
        let target = MockModel::new(100, 1);

        let mut decoder = SpeculativeDecoder::new(draft, target, 3).expect("create");

        for _ in 0..10 {
            let _ = decoder.decode_iteration(&[1, 2, 3]).expect("decode");
        }

        let stats = decoder.stats();
        assert_eq!(stats.iterations, 10);
        assert!(stats.tokens_speculated >= 10);
        assert!(stats.draft_time_ms >= 0.0);
        assert!(stats.target_time_ms >= 0.0);
    }

    #[test]
    fn test_speculative_decoder_reset_clears_all_stats() {
        let draft = MockModel::new(100, 1);
        let target = MockModel::new(100, 1);

        let mut decoder = SpeculativeDecoder::new(draft, target, 4).expect("create");

        // Run some iterations
        for _ in 0..5 {
            let _ = decoder.decode_iteration(&[1]).expect("decode");
        }

        // Reset
        decoder.reset_stats();

        let stats = decoder.stats();
        assert_eq!(stats.iterations, 0);
        assert_eq!(stats.tokens_speculated, 0);
        assert_eq!(stats.tokens_accepted, 0);
        assert_eq!(stats.acceptance_rate, 0.0);
        assert_eq!(stats.avg_spec_length, 0.0);
        assert_eq!(stats.time_saved_ms, 0.0);
        assert_eq!(stats.draft_time_ms, 0.0);
        assert_eq!(stats.target_time_ms, 0.0);
    }
include!("speculative_mock_tests.rs");
include!("speculative_config.rs");
include!("speculative_tests.rs");
}