entrenar 0.7.13

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
266
267
268
269
270
271
272
273
274
275
276
277
278
279
280
281
282
283
284

//! Tests for LoRA adapter serialization

use super::*;
use crate::lora::LoRALayer;
use crate::Tensor;
use approx::assert_abs_diff_eq;
use proptest::prelude::*;
use std::fs;

// ========================================================================
// PROPERTY TESTS - Serialization correctness validation
// ========================================================================

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

    /// Round-trip serialization should preserve all data exactly
    #[test]
    fn prop_round_trip_preserves_data(
        d_out in 2usize..16,
        d_in in 2usize..16,
        rank in 1usize..4,
        alpha in 1.0f32..32.0,
    ) {
        let size = d_out * d_in;
        let base_weight = Tensor::from_vec(vec![1.0; size], false);
        let layer = LoRALayer::new(base_weight.clone(), d_out, d_in, rank, alpha);

        // Create adapter and save
        let adapter = LoRAAdapter::from_layer(&layer, rank, alpha);
        let path = format!("/tmp/prop_test_adapter_{d_out}_{d_in}_{rank}.json");
        adapter.save(&path).expect("save should succeed");

        // Load and reconstruct
        let loaded = LoRAAdapter::load(&path).expect("load should succeed");
        let loaded_layer = loaded.to_layer(base_weight).expect("load should succeed");

        // Verify all fields preserved
        prop_assert_eq!(loaded_layer.d_out(), d_out);
        prop_assert_eq!(loaded_layer.d_in(), d_in);
        prop_assert_eq!(loaded_layer.rank(), rank);
        prop_assert!((loaded_layer.scale() - alpha / rank as f32).abs() < 1e-5);

        // Cleanup
        fs::remove_file(&path).ok();
    }

    /// Metadata calculation should be correct for any dimensions
    #[test]
    fn prop_metadata_correct(
        d_out in 2usize..32,
        d_in in 2usize..32,
        rank in 1usize..8,
        alpha in 1.0f32..64.0,
    ) {
        let size = d_out * d_in;
        let base_weight = Tensor::from_vec(vec![1.0; size], false);
        let layer = LoRALayer::new(base_weight, d_out, d_in, rank, alpha);

        let adapter = LoRAAdapter::from_layer(&layer, rank, alpha);
        let metadata = adapter.metadata();

        // Verify metadata
        prop_assert_eq!(metadata.d_out, d_out);
        prop_assert_eq!(metadata.d_in, d_in);
        prop_assert_eq!(metadata.rank, rank);
        prop_assert!((metadata.alpha - alpha).abs() < 1e-6);
        prop_assert_eq!(metadata.num_params, rank * d_in + d_out * rank);
        prop_assert_eq!(metadata.version, "1.0");
    }

    /// Forward output should be identical after save/load
    #[test]
    fn prop_forward_invariant_after_save_load(
        d in 4usize..12,
        rank in 1usize..3,
    ) {
        let size = d * d;
        let base_weight = Tensor::from_vec(vec![0.5; size], false);
        let layer = LoRALayer::new(base_weight.clone(), d, d, rank, 4.0);

        let x = Tensor::from_vec(vec![0.1; d], true);
        let original_output = layer.forward(&x);

        // Save and load
        let path = format!("/tmp/prop_forward_test_{d}_{rank}.json");
        save_adapter(&layer, rank, 4.0, &path).expect("save should succeed");
        let loaded_layer = load_adapter(base_weight, &path).expect("load should succeed");
        let loaded_output = loaded_layer.forward(&x);

        // Forward outputs must match
        prop_assert_eq!(original_output.len(), loaded_output.len());
        for i in 0..original_output.len() {
            prop_assert!(
                (original_output.data()[i] - loaded_output.data()[i]).abs() < 1e-5,
                "Forward output mismatch at index {}", i
            );
        }

        fs::remove_file(&path).ok();
    }

    /// Invalid dimension should be caught during to_layer
    #[test]
    fn prop_dimension_validation_catches_mismatches(
        d_out in 2usize..8,
        d_in in 2usize..8,
        rank in 1usize..3,
    ) {
        let size = d_out * d_in;
        let base_weight = Tensor::from_vec(vec![1.0; size], false);
        let layer = LoRALayer::new(base_weight, d_out, d_in, rank, 4.0);

        let adapter = LoRAAdapter::from_layer(&layer, rank, 4.0);

        // Try loading with wrong size base weight
        let wrong_size = size + 1;
        let wrong_base = Tensor::from_vec(vec![1.0; wrong_size], false);
        let result = adapter.to_layer(wrong_base);

        // Should fail with dimension error
        prop_assert!(result.is_err());
    }
}

// ========================================================================
// UNIT TESTS
// ========================================================================

#[test]
fn test_adapter_serialization_round_trip() {
    // Create LoRA layer
    let base_weight = Tensor::from_vec(vec![1.0, 2.0, 3.0, 4.0], false);
    let mut layer = LoRALayer::new(base_weight.clone(), 2, 2, 2, 4.0);

    // Set known weights
    *layer.lora_a_mut().data_mut() = ndarray::arr1(&[0.1, 0.2, 0.3, 0.4]);
    *layer.lora_b_mut().data_mut() = ndarray::arr1(&[0.5, 0.6, 0.7, 0.8]);

    // Save adapter
    let path = "/tmp/test_adapter.json";
    let adapter = LoRAAdapter::from_layer(&layer, 2, 4.0);
    adapter.save(path).expect("save should succeed");

    // Load adapter
    let loaded_adapter = LoRAAdapter::load(path).expect("load should succeed");
    let loaded_layer = loaded_adapter.to_layer(base_weight).expect("load should succeed");

    // Verify weights match
    for (&orig, &loaded) in layer.lora_a().data().iter().zip(loaded_layer.lora_a().data().iter()) {
        assert_abs_diff_eq!(orig, loaded, epsilon = 1e-6);
    }

    for (&orig, &loaded) in layer.lora_b().data().iter().zip(loaded_layer.lora_b().data().iter()) {
        assert_abs_diff_eq!(orig, loaded, epsilon = 1e-6);
    }

    // Verify metadata
    assert_eq!(loaded_layer.rank(), 2);
    assert_eq!(loaded_layer.d_out(), 2);
    assert_eq!(loaded_layer.d_in(), 2);
    assert_abs_diff_eq!(loaded_layer.scale(), 2.0, epsilon = 1e-6); // 4.0 / 2

    // Cleanup
    fs::remove_file(path).expect("operation should succeed");
}

#[test]
fn test_adapter_forward_consistency() {
    // Test that forward pass gives same result after save/load
    let base_weight = Tensor::from_vec(vec![1.0, 0.0, 0.0, 1.0], false);
    let mut layer = LoRALayer::new(base_weight.clone(), 2, 2, 1, 1.0);

    *layer.lora_a_mut().data_mut() = ndarray::arr1(&[0.5, 0.5]);
    *layer.lora_b_mut().data_mut() = ndarray::arr1(&[0.3, 0.3]);

    let x = Tensor::from_vec(vec![2.0, 3.0], true);
    let output_original = layer.forward(&x);

    // Save and load
    let path = "/tmp/test_adapter_forward.json";
    save_adapter(&layer, 1, 1.0, path).expect("save should succeed");
    let loaded_layer = load_adapter(base_weight, path).expect("load should succeed");

    let output_loaded = loaded_layer.forward(&x);

    // Verify outputs match
    assert_eq!(output_original.len(), output_loaded.len());
    for i in 0..output_original.len() {
        assert_abs_diff_eq!(output_original.data()[i], output_loaded.data()[i], epsilon = 1e-5);
    }

    // Cleanup
    fs::remove_file(path).expect("operation should succeed");
}

#[test]
fn test_adapter_dimension_validation() {
    let base_weight_2x2 = Tensor::from_vec(vec![1.0, 2.0, 3.0, 4.0], false);
    let layer = LoRALayer::new(base_weight_2x2, 2, 2, 2, 4.0);

    let adapter = LoRAAdapter::from_layer(&layer, 2, 4.0);

    // Try to load with wrong size base weight
    let wrong_base = Tensor::from_vec(vec![1.0, 2.0, 3.0, 4.0, 5.0, 6.0], false); // 3x2
    let result = adapter.to_layer(wrong_base);

    assert!(result.is_err());
    match result {
        Err(AdapterError::DimensionMismatch { .. }) => {}
        _ => panic!("Expected DimensionMismatch error"),
    }
}

#[test]
fn test_adapter_metadata() {
    let base_weight = Tensor::from_vec(vec![1.0, 2.0, 3.0, 4.0, 5.0, 6.0], false);
    let layer = LoRALayer::new(base_weight, 3, 2, 2, 8.0);

    let adapter = LoRAAdapter::from_layer(&layer, 2, 8.0);
    let metadata = adapter.metadata();

    assert_eq!(metadata.rank, 2);
    assert_abs_diff_eq!(metadata.alpha, 8.0, epsilon = 1e-6);
    assert_eq!(metadata.d_out, 3);
    assert_eq!(metadata.d_in, 2);
    assert_abs_diff_eq!(metadata.scale, 4.0, epsilon = 1e-6); // 8.0 / 2
    assert_eq!(metadata.num_params, 4 + 6); // A: 2*2=4, B: 3*2=6
    assert_eq!(metadata.version, "1.0");
}

#[test]
fn test_multiple_adapters_same_base() {
    // Test loading multiple different adapters on the same base weight
    let base_weight = Tensor::from_vec(vec![1.0, 0.0, 0.0, 1.0], false);

    // Create and save adapter 1
    let mut layer1 = LoRALayer::new(base_weight.clone(), 2, 2, 1, 1.0);
    *layer1.lora_a_mut().data_mut() = ndarray::arr1(&[1.0, 1.0]);
    *layer1.lora_b_mut().data_mut() = ndarray::arr1(&[1.0, 1.0]);
    save_adapter(&layer1, 1, 1.0, "/tmp/adapter1.json").expect("save should succeed");

    // Create and save adapter 2
    let mut layer2 = LoRALayer::new(base_weight.clone(), 2, 2, 1, 1.0);
    *layer2.lora_a_mut().data_mut() = ndarray::arr1(&[2.0, 2.0]);
    *layer2.lora_b_mut().data_mut() = ndarray::arr1(&[2.0, 2.0]);
    save_adapter(&layer2, 1, 1.0, "/tmp/adapter2.json").expect("save should succeed");

    // Load both adapters
    let loaded1 =
        load_adapter(base_weight.clone(), "/tmp/adapter1.json").expect("load should succeed");
    let loaded2 = load_adapter(base_weight, "/tmp/adapter2.json").expect("load should succeed");

    // Verify they're different
    assert_abs_diff_eq!(loaded1.lora_a().data()[0], 1.0, epsilon = 1e-6);
    assert_abs_diff_eq!(loaded2.lora_a().data()[0], 2.0, epsilon = 1e-6);

    // Cleanup
    fs::remove_file("/tmp/adapter1.json").expect("operation should succeed");
    fs::remove_file("/tmp/adapter2.json").expect("operation should succeed");
}

#[test]
fn test_adapter_file_format_readable() {
    // Verify JSON is human-readable
    let base_weight = Tensor::from_vec(vec![1.0, 2.0], false);
    let layer = LoRALayer::new(base_weight, 2, 1, 1, 2.0);

    let path = "/tmp/test_readable.json";
    save_adapter(&layer, 1, 2.0, path).expect("save should succeed");

    // Read raw file content
    let content = fs::read_to_string(path).expect("file read should succeed");

    // Should contain key fields
    assert!(content.contains("\"version\""));
    assert!(content.contains("\"rank\""));
    assert!(content.contains("\"alpha\""));
    assert!(content.contains("\"lora_a\""));
    assert!(content.contains("\"lora_b\""));

    // Cleanup
    fs::remove_file(path).expect("operation should succeed");
}