entrenar 0.7.9

Training & Optimization library with autograd, LoRA, quantization, and model merging
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

//! Tests for benchmarks module

use super::super::granularity::{
    calibrate_per_channel, calibrate_per_tensor, dequantize_with_params, quantization_mse,
    quantize_with_params, QuantGranularity, QuantMode,
};
use super::generators::{
    generate_gaussian_weights, generate_multi_channel_weights, generate_uniform_weights,
    generate_weights_with_outliers,
};
use super::runners::{
    accuracy_retention, compare_bit_width_degradation, run_benchmark, run_full_benchmark_suite,
};
use super::types::{BenchmarkSuite, QuantBenchmarkResult};
use proptest::prelude::*;

#[test]
fn test_run_benchmark() {
    let values = generate_gaussian_weights(1000, 0.0, 1.0, 42);
    let result =
        run_benchmark("test", &values, 8, QuantGranularity::PerTensor, QuantMode::Symmetric);

    assert_eq!(result.name, "test");
    assert_eq!(result.num_elements, 1000);
    assert_eq!(result.bits, 8);
    assert!(result.mse >= 0.0);
    assert!(result.sqnr_db > 0.0);
    assert!(result.compression_ratio > 1.0);
}

#[test]
fn test_benchmark_suite() {
    let mut suite = BenchmarkSuite::default();

    suite.add(QuantBenchmarkResult {
        name: "a".to_string(),
        num_elements: 100,
        bits: 8,
        granularity: QuantGranularity::PerTensor,
        mode: QuantMode::Symmetric,
        mse: 0.01,
        max_error: 0.1,
        sqnr_db: 40.0,
        compression_ratio: 4.0,
    });

    suite.add(QuantBenchmarkResult {
        name: "b".to_string(),
        num_elements: 100,
        bits: 4,
        granularity: QuantGranularity::PerTensor,
        mode: QuantMode::Symmetric,
        mse: 0.1,
        max_error: 0.5,
        sqnr_db: 20.0,
        compression_ratio: 8.0,
    });

    assert_eq!(suite.results.len(), 2);
    assert_eq!(suite.best_by_sqnr().expect("operation should succeed").name, "a");
    assert_eq!(suite.best_by_mse().expect("operation should succeed").name, "a");
}

#[test]
fn test_quality_score() {
    let result = QuantBenchmarkResult {
        name: "test".to_string(),
        num_elements: 100,
        bits: 8,
        granularity: QuantGranularity::PerTensor,
        mode: QuantMode::Symmetric,
        mse: 0.01,
        max_error: 0.1,
        sqnr_db: 40.0,
        compression_ratio: 4.0,
    };

    assert_eq!(result.quality_score(), 10.0); // 40 / 4
}

#[test]
fn test_generate_gaussian_weights() {
    let weights = generate_gaussian_weights(1000, 0.0, 1.0, 42);

    assert_eq!(weights.len(), 1000);

    // Mean should be approximately 0
    let mean: f32 = weights.iter().sum::<f32>() / weights.len() as f32;
    assert!(mean.abs() < 0.2, "Mean {mean} should be close to 0");

    // Std dev should be approximately 1
    let variance: f32 =
        weights.iter().map(|x| (x - mean).powi(2)).sum::<f32>() / weights.len() as f32;
    let std_dev = variance.sqrt();
    assert!((std_dev - 1.0).abs() < 0.3, "Std dev {std_dev} should be close to 1");
}

#[test]
fn test_generate_uniform_weights() {
    let weights = generate_uniform_weights(1000, -1.0, 1.0, 42);

    assert_eq!(weights.len(), 1000);

    for &w in &weights {
        assert!((-1.0..=1.0).contains(&w), "Weight {w} out of range");
    }
}

#[test]
fn test_generate_weights_with_outliers() {
    let weights = generate_weights_with_outliers(1000, 0.01, 10.0, 42);

    assert_eq!(weights.len(), 1000);

    // Should have some large values
    let large_count = weights.iter().filter(|&&w| w.abs() > 5.0).count();
    assert!(large_count > 0, "Should have outliers");
}

#[test]
fn test_generate_multi_channel_weights() {
    let weights = generate_multi_channel_weights(16, 64, 5.0, 42);

    assert_eq!(weights.len(), 16 * 64);
}

#[test]
fn test_full_benchmark_suite() {
    let suite = run_full_benchmark_suite(256);

    // Should have multiple results
    assert!(suite.results.len() >= 5);

    // All results should be valid
    for result in &suite.results {
        assert!(result.mse >= 0.0);
        assert!(result.compression_ratio >= 1.0);
    }
}

#[test]
fn test_bit_width_comparison() {
    let values = generate_gaussian_weights(1000, 0.0, 1.0, 42);
    let results = compare_bit_width_degradation(&values);

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

    // 8-bit should have lower MSE than 4-bit
    let (_, mse_4bit, _) =
        results.iter().find(|(b, _, _)| *b == 4).expect("operation should succeed");
    let (_, mse_8bit, _) =
        results.iter().find(|(b, _, _)| *b == 8).expect("operation should succeed");

    assert!(mse_8bit <= mse_4bit, "8-bit MSE ({mse_8bit}) should be <= 4-bit MSE ({mse_4bit})");
}

#[test]
fn test_accuracy_retention() {
    assert_eq!(accuracy_retention(0.0, 0.0), 100.0);
    assert!(accuracy_retention(0.0, 0.01) >= 0.0);
    assert!(accuracy_retention(0.0, 0.01) <= 100.0);
}

// Property tests

proptest! {
    #![proptest_config(ProptestConfig::with_cases(100))]

    #[test]
    fn prop_benchmark_compression_positive(
        size in 100usize..500,
        bits in prop::sample::select(vec![4u8, 8])
    ) {
        let values = generate_gaussian_weights(size, 0.0, 1.0, 42);
        let result = run_benchmark(
            "test",
            &values,
            bits,
            QuantGranularity::PerTensor,
            QuantMode::Symmetric,
        );

        prop_assert!(result.compression_ratio > 0.0);
    }

    #[test]
    fn prop_8bit_better_than_4bit(size in 100usize..500) {
        let values = generate_gaussian_weights(size, 0.0, 1.0, 42);

        let result_4bit = run_benchmark(
            "4bit",
            &values,
            4,
            QuantGranularity::PerTensor,
            QuantMode::Symmetric,
        );
        let result_8bit = run_benchmark(
            "8bit",
            &values,
            8,
            QuantGranularity::PerTensor,
            QuantMode::Symmetric,
        );

        prop_assert!(
            result_8bit.mse <= result_4bit.mse * 1.01,
            "8-bit MSE ({}) should be <= 4-bit MSE ({})",
            result_8bit.mse,
            result_4bit.mse
        );
    }

    #[test]
    fn prop_per_channel_helps_multi_scale(
        num_channels in 4usize..16,
        scale_variance in 5.0f32..20.0 // Higher variance to ensure per-channel helps
    ) {
        let features = 64;
        let values = generate_multi_channel_weights(num_channels, features, scale_variance, 42);

        // Use calibration directly with correct num_channels
        let params_tensor = calibrate_per_tensor(&values, 8, QuantMode::Symmetric);
        let params_channel = calibrate_per_channel(&values, num_channels, 8, QuantMode::Symmetric);

        let q_tensor = quantize_with_params(&values, &params_tensor);
        let q_channel = quantize_with_params(&values, &params_channel);

        let d_tensor = dequantize_with_params(&q_tensor, &params_tensor);
        let d_channel = dequantize_with_params(&q_channel, &params_channel);

        let mse_tensor = quantization_mse(&values, &d_tensor);
        let mse_channel = quantization_mse(&values, &d_channel);

        // Per-channel should be at least as good when scales vary significantly
        prop_assert!(
            mse_channel <= mse_tensor * 1.01,
            "Per-channel MSE ({}) should be <= per-tensor MSE ({})",
            mse_channel,
            mse_tensor
        );
    }

    #[test]
    fn prop_benchmark_deterministic(size in 100usize..500) {
        let values1 = generate_gaussian_weights(size, 0.0, 1.0, 42);
        let values2 = generate_gaussian_weights(size, 0.0, 1.0, 42);

        // Same seed should produce same weights
        prop_assert_eq!(values1, values2);
    }

    #[test]
    fn prop_sqnr_positive_for_signal(size in 100usize..500) {
        let values = generate_gaussian_weights(size, 0.0, 1.0, 42);
        let result = run_benchmark(
            "test",
            &values,
            8,
            QuantGranularity::PerTensor,
            QuantMode::Symmetric,
        );

        prop_assert!(result.sqnr_db > 0.0, "SQNR must be positive");
    }
}