entrenar 0.7.8

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
285
286
287
288
289
290
291
292
293
294
295
296
297
298
299
300
301
302
303
304
305
306
307
308
309
310
311
312
313
314
315
316
317
318
319
320
321
322
323
324
325
326
327
328
329
330
331
332
333
334
335
336
337
338
339
340
341
342
343
344
345
346
347
348
349
350
351
352
353
354
355
356
357
358
359
360
361
362
363
364
365
366
367
368
369
370
371
372
373
374
375
376
377
378
379
380
381
382

//! Tests for fake quantization.

use super::*;
use crate::Tensor;
use approx::assert_abs_diff_eq;
use proptest::prelude::*;

// ========================================================================
// PROPERTY TESTS - Fake quantization correctness
// ========================================================================

proptest! {
    #![proptest_config(proptest::test_runner::Config::with_cases(200))]

    /// STE backward should always pass gradients unchanged
    #[test]
    fn prop_ste_backward_identity(
        grad in prop::collection::vec(-10.0f32..10.0, 1..32),
    ) {
        let grad_tensor = Tensor::from_vec(grad.clone(), true);
        let fq = FakeQuantize::q8();

        let backward = fq.backward(&grad_tensor);

        // STE should pass through unchanged
        prop_assert_eq!(backward.len(), grad.len());
        for (i, &g) in grad.iter().enumerate() {
            prop_assert!(
                (backward.data()[i] - g).abs() < 1e-6,
                "STE should preserve gradient at index {}", i
            );
        }
    }

    /// Fake quantize should produce values that are multiples of scale
    #[test]
    fn prop_fake_quantize_produces_quantized_values(
        values in prop::collection::vec(-5.0f32..5.0, 4..32),
        bits in 4usize..9,
    ) {
        let input = Tensor::from_vec(values.clone(), false);
        let config = FakeQuantConfig::symmetric(bits);
        let mut fq = FakeQuantize::new(config);
        fq.calibrate(&values);

        let output = fq.forward(&input);

        // Each output value should be a valid quantized level
        let scale = fq.scale();
        for &val in output.data() {
            // Value should be approximately q * scale for some integer q
            let q = (val / scale).round();
            let reconstructed = q * scale;
            prop_assert!(
                (val - reconstructed).abs() < 1e-5,
                "Value {} should be quantized (q={}, scale={})",
                val, q, scale
            );
        }
    }

    /// Fake quantize output should be bounded by qmin*scale and qmax*scale
    #[test]
    fn prop_fake_quantize_bounded_output(
        values in prop::collection::vec(-100.0f32..100.0, 4..32),
        bits in 4usize..9,
    ) {
        let input = Tensor::from_vec(values.clone(), false);
        let config = FakeQuantConfig::symmetric(bits);
        let mut fq = FakeQuantize::new(config);
        fq.calibrate(&values);

        let output = fq.forward(&input);

        let qmin_float = fq.config.qmin as f32 * fq.scale();
        let qmax_float = fq.config.qmax as f32 * fq.scale();

        for &val in output.data() {
            prop_assert!(
                val >= qmin_float - 1e-5 && val <= qmax_float + 1e-5,
                "Output {} should be in [{}, {}]",
                val, qmin_float, qmax_float
            );
        }
    }

    /// Calibration should set scale based on data range
    #[test]
    fn prop_calibration_sets_appropriate_scale(
        values in prop::collection::vec(-10.0f32..10.0, 4..32),
        bits in 4usize..9,
    ) {
        let config = FakeQuantConfig::symmetric(bits);
        let mut fq = FakeQuantize::new(config);

        fq.calibrate(&values);

        prop_assert!(fq.is_initialized());
        prop_assert!(fq.scale() > 0.0);

        // For symmetric, scale should be max_abs / qmax
        let max_abs = values.iter().map(|v| v.abs()).fold(0.0f32, f32::max);
        let expected_scale = max_abs / fq.config.qmax as f32;

        // Allow small tolerance for numerical precision
        if max_abs > 1e-8 {
            prop_assert!(
                (fq.scale() - expected_scale).abs() < 1e-5,
                "Scale {} should be {} (max_abs={}, qmax={})",
                fq.scale(), expected_scale, max_abs, fq.config.qmax
            );
        }
    }

    /// Number of quantization levels should be correct
    #[test]
    fn prop_num_levels_correct(bits in 2usize..10) {
        let config = FakeQuantConfig::symmetric(bits);
        let fq = FakeQuantize::new(config);

        // Symmetric: qmin = -(2^(bits-1)-1), qmax = 2^(bits-1)-1
        // Levels = qmax - qmin + 1 = 2 * (2^(bits-1)-1) + 1 = 2^bits - 1
        let expected = (1 << bits) - 1;
        prop_assert_eq!(fq.num_levels(), expected);
    }
}

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

#[test]
fn test_fake_quantize_config_symmetric() {
    let config = FakeQuantConfig::symmetric(4);
    assert_eq!(config.bits, 4);
    assert!(config.symmetric);
    assert_eq!(config.qmin, -7);
    assert_eq!(config.qmax, 7);

    let config8 = FakeQuantConfig::symmetric(8);
    assert_eq!(config8.qmin, -127);
    assert_eq!(config8.qmax, 127);
}

#[test]
fn test_fake_quantize_config_asymmetric() {
    let config = FakeQuantConfig::asymmetric(4);
    assert_eq!(config.bits, 4);
    assert!(!config.symmetric);
    assert_eq!(config.qmin, 0);
    assert_eq!(config.qmax, 15);

    let config8 = FakeQuantConfig::asymmetric(8);
    assert_eq!(config8.qmin, 0);
    assert_eq!(config8.qmax, 255);
}

#[test]
fn test_fake_quantize_forward() {
    let input = Tensor::from_vec(vec![0.0, 1.0, -1.0, 0.5, -0.5], false);
    let mut fq = FakeQuantize::q8();
    fq.calibrate(input.data().as_slice().expect("operation should succeed"));

    let output = fq.forward(&input);

    assert_eq!(output.len(), 5);
    // 0 should stay 0
    assert_abs_diff_eq!(output.data()[0], 0.0, epsilon = 1e-5);
}

#[test]
fn test_fake_quantize_forward_with_calibration() {
    let input = Tensor::from_vec(vec![0.0, 1.0, -1.0, 0.5, -0.5], false);
    let mut fq = FakeQuantize::q8();

    assert!(!fq.is_initialized());

    let output = fq.forward_with_calibration(&input);

    assert!(fq.is_initialized());
    assert_eq!(output.len(), 5);
}

#[test]
fn test_ste_backward() {
    let grad = Tensor::from_vec(vec![1.0, 2.0, 3.0, 4.0], true);
    let fq = FakeQuantize::q8();

    let backward = fq.backward(&grad);

    // STE: gradient should pass through unchanged
    assert_eq!(backward.len(), 4);
    for i in 0..4 {
        assert_abs_diff_eq!(backward.data()[i], grad.data()[i], epsilon = 1e-6);
    }
}

#[test]
fn test_clamped_ste_backward() {
    let grad = Tensor::from_vec(vec![1.0, 1.0, 1.0], true);
    let input = Tensor::from_vec(vec![0.5, 10.0, -10.0], false); // 10, -10 outside range

    let mut fq = FakeQuantize::q4();
    fq.scale = 1.0; // Set scale so range is [-7, 7]
    fq.initialized = true;

    let backward = fq.backward_clamped(&grad, &input);

    // 0.5 is in range: gradient passes
    assert_abs_diff_eq!(backward.data()[0], 1.0, epsilon = 1e-6);
    // 10.0 is outside range: gradient clipped to 0
    assert_abs_diff_eq!(backward.data()[1], 0.0, epsilon = 1e-6);
    // -10.0 is outside range: gradient clipped to 0
    assert_abs_diff_eq!(backward.data()[2], 0.0, epsilon = 1e-6);
}

#[test]
fn test_calibration_symmetric() {
    let mut fq = FakeQuantize::q8();
    let data = vec![0.0, 1.0, -2.0, 1.5, -1.5];

    fq.calibrate(&data);

    // max_abs = 2.0, qmax = 127
    // scale = 2.0 / 127
    let expected_scale = 2.0 / 127.0;
    assert_abs_diff_eq!(fq.scale(), expected_scale, epsilon = 1e-6);
    assert_eq!(fq.zero_point(), 0);
}

#[test]
fn test_fake_quantize_convenience_function() {
    let input = Tensor::from_vec(vec![0.0, 1.0, -1.0], false);

    let output = fake_quantize(&input, 8, true);

    assert_eq!(output.len(), 3);
}

#[test]
fn test_ste_backward_convenience_function() {
    let grad = Tensor::from_vec(vec![1.0, 2.0, 3.0], true);

    let backward = ste_backward(&grad);

    for i in 0..3 {
        assert_abs_diff_eq!(backward.data()[i], grad.data()[i], epsilon = 1e-6);
    }
}

#[test]
fn test_num_levels() {
    let fq4 = FakeQuantize::q4();
    assert_eq!(fq4.num_levels(), 15); // -7 to 7 = 15 levels

    let fq8 = FakeQuantize::q8();
    assert_eq!(fq8.num_levels(), 255); // -127 to 127 = 255 levels
}

#[test]
fn test_quantize_dequantize_round_trip() {
    let input = Tensor::from_vec(vec![0.0, 0.5, 1.0, -0.5, -1.0], false);
    let mut fq = FakeQuantize::q8();
    fq.calibrate(input.data().as_slice().expect("operation should succeed"));

    let output = fq.forward(&input);

    // Output should be close to input (with quantization noise)
    for (i, (&orig, &out)) in input.data().iter().zip(output.data().iter()).enumerate() {
        let error = (orig - out).abs();
        assert!(error < 0.1, "Error {error} at index {i} too large: {orig} vs {out}");
    }
}

#[test]
fn test_fake_quant_config_default() {
    let config = FakeQuantConfig::default();
    // Default should be q8_symmetric
    assert_eq!(config.bits, 8);
    assert!(config.symmetric);
    assert_eq!(config.qmin, -127);
    assert_eq!(config.qmax, 127);
}

#[test]
fn test_calibrate_empty_data() {
    let mut fq = FakeQuantize::q8();
    fq.calibrate(&[]);
    // Should not crash, and scale should remain at default
    assert!(!fq.is_initialized());
    assert_eq!(fq.scale(), 1.0);
}

#[test]
fn test_calibration_asymmetric() {
    let config = FakeQuantConfig::asymmetric(8);
    let mut fq = FakeQuantize::new(config);

    // All positive data
    let data = vec![0.0, 1.0, 2.0, 3.0, 4.0];
    fq.calibrate(&data);

    assert!(fq.is_initialized());
    assert!(fq.scale() > 0.0);
    // For asymmetric, zero_point should be computed
    // scale = (4-0) / (255-0) ≈ 0.0157
    // zero_point = round(0 - 0/scale) = 0
    assert!(fq.zero_point() >= 0);
}

#[test]
fn test_calibration_asymmetric_negative_offset() {
    let config = FakeQuantConfig::asymmetric(8);
    let mut fq = FakeQuantize::new(config);

    // Shifted data
    let data = vec![10.0, 11.0, 12.0, 13.0, 14.0];
    fq.calibrate(&data);

    assert!(fq.is_initialized());
    assert!(fq.scale() > 0.0);
}

#[test]
fn test_asymmetric_forward() {
    let config = FakeQuantConfig::asymmetric(8);
    let mut fq = FakeQuantize::new(config);

    let data = vec![0.0, 0.5, 1.0, 1.5, 2.0];
    fq.calibrate(&data);

    let input = Tensor::from_vec(data.clone(), false);
    let output = fq.forward(&input);

    assert_eq!(output.len(), 5);
    // Output should be quantized values
    for &val in output.data() {
        assert!(val.is_finite());
    }
}

#[test]
fn test_fake_quantize_asymmetric_convenience() {
    let input = Tensor::from_vec(vec![0.0, 1.0, 2.0], false);
    let output = fake_quantize(&input, 8, false);
    assert_eq!(output.len(), 3);
}

#[test]
fn test_fake_quant_config_clone() {
    let config = FakeQuantConfig::symmetric(4);
    let cloned = config.clone();
    assert_eq!(config.bits, cloned.bits);
    assert_eq!(config.symmetric, cloned.symmetric);
    assert_eq!(config.qmin, cloned.qmin);
    assert_eq!(config.qmax, cloned.qmax);
}

#[test]
fn test_fake_quantize_clone() {
    let mut fq = FakeQuantize::q8();
    fq.scale = 0.5;
    fq.zero_point = 10;
    fq.initialized = true;

    let cloned = fq.clone();
    assert_eq!(fq.scale, cloned.scale);
    assert_eq!(fq.zero_point, cloned.zero_point);
    assert_eq!(fq.initialized, cloned.initialized);
}

#[test]
fn test_calibration_near_zero_scale() {
    let mut fq = FakeQuantize::q8();
    // All zeros - would result in zero scale without protection
    let data = vec![0.0, 0.0, 0.0];
    fq.calibrate(&data);

    // Scale should be clamped to minimum
    assert!(fq.scale() >= 1e-10);
    assert!(fq.is_initialized());
}