tensorlogic-train 0.1.0

Training loops, loss composition, and optimization schedules for TensorLogic
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

//! Unit tests for the LoRA module.

use super::adapter::LoraAdapter;
use super::config::LoraConfig;
use super::error::{LoraError, LoraResult};
use super::layer::LoraLayer;

fn identity_weight(n: usize) -> Vec<Vec<f64>> {
    (0..n)
        .map(|i| {
            let mut row = vec![0.0; n];
            row[i] = 1.0;
            row
        })
        .collect()
}

fn constant_weight(d: usize, k: usize, val: f64) -> Vec<Vec<f64>> {
    vec![vec![val; k]; d]
}

fn cfg(rank: usize) -> LoraConfig {
    LoraConfig {
        rank,
        alpha: rank as f64,
        dropout: 0.0,
        target_modules: Vec::new(),
        seed: 42,
    }
}

// -----------------------------------------------------------------------
// 1. Valid creation
// -----------------------------------------------------------------------

#[test]
fn test_layer_creation_valid() -> LoraResult<()> {
    let w = identity_weight(8);
    let layer = LoraLayer::new(w, cfg(4))?;
    assert!(!layer.merged);
    assert_eq!(layer.weight_a.len(), 4);
    assert_eq!(layer.weight_a[0].len(), 8);
    assert_eq!(layer.weight_b.len(), 8);
    assert_eq!(layer.weight_b[0].len(), 4);
    Ok(())
}

// -----------------------------------------------------------------------
// 2. Forward pass dimensions
// -----------------------------------------------------------------------

#[test]
fn test_forward_output_dimensions() -> LoraResult<()> {
    let d = 16;
    let k = 32;
    let n = 5;
    let w = constant_weight(d, k, 0.01);
    let mut layer = LoraLayer::new(w, cfg(4))?;
    let input: Vec<Vec<f64>> = vec![vec![1.0; k]; n];
    let out = layer.forward(&input)?;
    assert_eq!(out.len(), n);
    assert_eq!(out[0].len(), d);
    Ok(())
}

// -----------------------------------------------------------------------
// 3. Merge/unmerge round-trip preserves base weight
// -----------------------------------------------------------------------

#[test]
fn test_merge_unmerge_roundtrip() -> LoraResult<()> {
    let d = 8;
    let k = 8;
    let w = identity_weight(d);
    let original: Vec<Vec<f64>> = w.clone();
    let mut layer = LoraLayer::new(w, cfg(2))?;

    layer.merge()?;
    assert!(layer.merged);
    layer.unmerge()?;
    assert!(!layer.merged);

    for (i, (row_got, row_orig)) in layer
        .base_weight
        .iter()
        .zip(original.iter())
        .enumerate()
        .take(d)
    {
        for (j, (got, orig)) in row_got.iter().zip(row_orig.iter()).enumerate().take(k) {
            let diff = (got - orig).abs();
            assert!(diff < 1e-10, "roundtrip drift at ({i},{j}): {diff}");
        }
    }
    Ok(())
}

// -----------------------------------------------------------------------
// 4. Compression ratio
// -----------------------------------------------------------------------

#[test]
fn test_compression_ratio() -> LoraResult<()> {
    let d = 64;
    let k = 128;
    let r = 4;
    let w = constant_weight(d, k, 0.0);
    let layer = LoraLayer::new(w, cfg(r))?;

    let trainable = r * (d + k); // 4 * 192 = 768
    let total = d * k + trainable; // 8192 + 768 = 8960
    assert_eq!(layer.trainable_params(), trainable);
    assert_eq!(layer.total_params(), total);
    let expected_ratio = trainable as f64 / total as f64;
    assert!((layer.compression_ratio() - expected_ratio).abs() < 1e-12);
    Ok(())
}

// -----------------------------------------------------------------------
// 5. Error on rank == 0
// -----------------------------------------------------------------------

#[test]
fn test_error_rank_zero() {
    let w = identity_weight(8);
    let err = LoraLayer::new(w, cfg(0));
    assert!(err.is_err());
    match err.err() {
        Some(LoraError::InvalidRank(0)) => {}
        other => panic!("expected InvalidRank(0), got {other:?}"),
    }
}

// -----------------------------------------------------------------------
// 6. Error on rank > min(d, k)
// -----------------------------------------------------------------------

#[test]
fn test_error_rank_too_large() {
    let w = constant_weight(4, 6, 1.0);
    let err = LoraLayer::new(w, cfg(5));
    assert!(err.is_err());
    match err.err() {
        Some(LoraError::InvalidRank(5)) => {}
        other => panic!("expected InvalidRank(5), got {other:?}"),
    }
}

// -----------------------------------------------------------------------
// 7. Error on dimension mismatch in forward
// -----------------------------------------------------------------------

#[test]
fn test_error_forward_dim_mismatch() -> LoraResult<()> {
    let d = 8;
    let k = 16;
    let w = constant_weight(d, k, 0.0);
    let mut layer = LoraLayer::new(w, cfg(2))?;
    let bad_input = vec![vec![1.0; 10]; 3];
    let err = layer.forward(&bad_input);
    assert!(err.is_err());
    assert!(matches!(
        err.err(),
        Some(LoraError::DimensionMismatch { .. })
    ));
    Ok(())
}

// -----------------------------------------------------------------------
// 8. Adapter multi-layer forward
// -----------------------------------------------------------------------

#[test]
fn test_adapter_multi_layer_forward() -> LoraResult<()> {
    let config = LoraConfig {
        rank: 2,
        alpha: 2.0,
        dropout: 0.0,
        target_modules: vec!["q".into(), "k".into(), "v".into()],
        seed: 99,
    };
    let mut adapter = LoraAdapter::new(config);
    let d = 8;
    let k = 16;

    adapter.add_layer("q", constant_weight(d, k, 0.1))?;
    adapter.add_layer("k", constant_weight(d, k, 0.2))?;
    adapter.add_layer("v", constant_weight(d, k, 0.3))?;

    let input = vec![vec![1.0; k]; 4];
    for name in &["q", "k", "v"] {
        let out = adapter.forward(name, &input)?;
        assert_eq!(out.len(), 4);
        assert_eq!(out[0].len(), d);
    }
    Ok(())
}

// -----------------------------------------------------------------------
// 9. Trainable params count matches r*(d+k)
// -----------------------------------------------------------------------

#[test]
fn test_trainable_params_count() -> LoraResult<()> {
    let d = 32;
    let k = 64;
    let r = 8;
    let w = constant_weight(d, k, 0.0);
    let layer = LoraLayer::new(w, cfg(r))?;
    assert_eq!(layer.trainable_params(), r * (d + k));
    Ok(())
}

// -----------------------------------------------------------------------
// 10. Effective weight matches base_weight after merge
// -----------------------------------------------------------------------

#[test]
fn test_effective_weight_after_merge() -> LoraResult<()> {
    let d = 8;
    let k = 8;
    let w = identity_weight(d);
    let mut layer = LoraLayer::new(w, cfg(2))?;

    let eff_before = layer.effective_weight()?;
    layer.merge()?;
    let eff_after = layer.effective_weight()?;

    for (i, (row_before, row_after)) in eff_before.iter().zip(eff_after.iter()).enumerate().take(d)
    {
        for (j, (before, after)) in row_before.iter().zip(row_after.iter()).enumerate().take(k) {
            let diff = (before - after).abs();
            assert!(diff < 1e-10, "effective_weight drift at ({i},{j}): {diff}");
        }
    }
    Ok(())
}

// -----------------------------------------------------------------------
// 11. Double merge returns error
// -----------------------------------------------------------------------

#[test]
fn test_double_merge_error() -> LoraResult<()> {
    let w = identity_weight(8);
    let mut layer = LoraLayer::new(w, cfg(2))?;
    layer.merge()?;
    let err = layer.merge();
    assert!(err.is_err());
    assert!(matches!(err.err(), Some(LoraError::MergeError(_))));
    Ok(())
}

// -----------------------------------------------------------------------
// 12. Adapter summary totals
// -----------------------------------------------------------------------

#[test]
fn test_adapter_summary() -> LoraResult<()> {
    let config = LoraConfig {
        rank: 4,
        alpha: 4.0,
        dropout: 0.0,
        target_modules: Vec::new(),
        seed: 7,
    };
    let mut adapter = LoraAdapter::new(config);
    adapter.add_layer("a", constant_weight(16, 32, 0.0))?;
    adapter.add_layer("b", constant_weight(8, 64, 0.0))?;

    let summary = adapter.summary();
    assert_eq!(summary.layers.len(), 2);
    // layer a: 4 * (16 + 32) = 192,  layer b: 4 * (8 + 64) = 288
    assert_eq!(summary.total_trainable, 192 + 288);
    // layer a total: 16*32 + 192 = 704,  layer b total: 8*64 + 288 = 800
    assert_eq!(summary.total_params, 704 + 800);
    Ok(())
}