tensorlogic-trustformers 0.1.0

Transformer-as-rules: Self-attention and FFN layers as einsum expressions
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

//! Tests for the quantization module.

use ndarray::Array2;
use tensorlogic_scirs_backend::quantization::{
    QuantizationGranularity, QuantizationParams, QuantizationScheme, QuantizationType,
};

use crate::quantization::{calibrate_linear, QuantizedLinear};

// -----------------------------------------------------------------------
// Helper: 4×8 weight matrix with simple values.
// -----------------------------------------------------------------------

fn sample_weight_4x8() -> Array2<f64> {
    Array2::from_shape_fn((4, 8), |(i, j)| (i * 8 + j) as f64 * 0.5 - 14.0)
}

// -----------------------------------------------------------------------
// Test 1: round-trip per-tensor INT8 accuracy
// -----------------------------------------------------------------------

#[test]
fn test_roundtrip_per_tensor() {
    let weight = sample_weight_4x8();
    let params = calibrate_linear(
        &weight,
        QuantizationType::Int8,
        QuantizationGranularity::PerTensor,
    );
    let qlinear = QuantizedLinear::from_fp(&weight, &params).expect("from_fp per-tensor");

    let deq = qlinear.dequantize();
    let max_abs_err = weight
        .iter()
        .zip(deq.iter())
        .map(|(o, d)| (o - d).abs())
        .fold(0.0_f64, f64::max);

    // INT8 quantization noise: step = scale/2 ≤ (max_range/127)/2 ≈ 0.11
    // Allow up to 2.0 for safety across different calibrated ranges.
    assert!(
        max_abs_err < 2.0,
        "per-tensor round-trip max error={max_abs_err} >= 2.0"
    );
}

// -----------------------------------------------------------------------
// Test 2: round-trip per-channel INT8 accuracy
// -----------------------------------------------------------------------

#[test]
fn test_roundtrip_per_channel() {
    let weight = sample_weight_4x8();
    let params = calibrate_linear(
        &weight,
        QuantizationType::Int8,
        QuantizationGranularity::PerChannel,
    );
    let qlinear = QuantizedLinear::from_fp(&weight, &params).expect("from_fp per-channel");

    let deq = qlinear.dequantize();
    let max_abs_err = weight
        .iter()
        .zip(deq.iter())
        .map(|(o, d)| (o - d).abs())
        .fold(0.0_f64, f64::max);

    assert!(
        max_abs_err < 2.0,
        "per-channel round-trip max error={max_abs_err} >= 2.0"
    );
}

// -----------------------------------------------------------------------
// Test 3: forward pass matches f64 matmul within tolerance
// -----------------------------------------------------------------------

#[test]
fn test_forward_matches_fp() {
    let weight = sample_weight_4x8();
    let params = calibrate_linear(
        &weight,
        QuantizationType::Int8,
        QuantizationGranularity::PerChannel,
    );
    let qlinear = QuantizedLinear::from_fp(&weight, &params).expect("from_fp forward test");

    // x: [batch=2, in_features=8]
    let x = Array2::from_shape_fn((2, 8), |(i, j)| (i + j) as f64 * 0.1);

    let out_q = qlinear.forward(&x);

    // Reference: dequantize weights and matmul
    let weight_fp = qlinear.dequantize();
    let out_ref = x.dot(&weight_fp.t());

    assert_eq!(out_q.shape(), &[2, 4]);
    // forward() calls dequantize() internally, so they must match exactly.
    for (a, b) in out_q.iter().zip(out_ref.iter()) {
        assert!((a - b).abs() < 1e-12, "forward mismatch: {a} vs {b}");
    }
}

// -----------------------------------------------------------------------
// Test 4: calibration produces sensible scale and zero_point
// -----------------------------------------------------------------------

#[test]
fn test_calibration_sanity() {
    let weight = sample_weight_4x8();
    let params = calibrate_linear(
        &weight,
        QuantizationType::Int8,
        QuantizationGranularity::PerTensor,
    );

    // scale must be positive and finite
    assert!(params.scale[0] > 0.0, "scale[0]={}", params.scale[0]);
    assert!(params.scale[0].is_finite(), "scale not finite");
    // Symmetric calibration → zero_point == 0
    assert_eq!(
        params.zero_point[0], 0,
        "symmetric should have zero_point==0"
    );
    // sanity: scale < 1.0 for values in [-14, 14]  (14/127 ≈ 0.11)
    assert!(
        params.scale[0] < 1.0,
        "scale[0]={} unreasonably large",
        params.scale[0]
    );
}

// -----------------------------------------------------------------------
// Test 5: per-channel granularity uses different scales per output row
// -----------------------------------------------------------------------

#[test]
fn test_per_channel_uses_different_scales() {
    // Design a weight where rows have very different magnitudes:
    //   row 0 ~ [-100, 100]  → large dynamic range
    //   row 1 ~ [-1,    1]   → tiny dynamic range
    let data = vec![100.0_f64, -100.0, 50.0, -50.0, 1.0_f64, -1.0, 0.5, -0.5];
    let weight = Array2::from_shape_vec((2, 4), data).expect("build test weight");
    let params = calibrate_linear(
        &weight,
        QuantizationType::Int8,
        QuantizationGranularity::PerChannel,
    );

    assert_eq!(params.scale.len(), 2, "PerChannel needs 2 scales");
    // Row 0 scale ≈ 100/127 ≈ 0.787; Row 1 scale ≈ 1/127 ≈ 0.0079
    // They must differ significantly (at least 10x ratio).
    let ratio = params.scale[0] / params.scale[1];
    assert!(
        ratio > 10.0,
        "scale[0]={} scale[1]={} ratio={} (expected >10)",
        params.scale[0],
        params.scale[1],
        ratio
    );
}

// -----------------------------------------------------------------------
// Test 6: bias support
// -----------------------------------------------------------------------

#[test]
fn test_bias_shapes_checked() {
    let weight = sample_weight_4x8();
    let params = calibrate_linear(
        &weight,
        QuantizationType::Int8,
        QuantizationGranularity::PerTensor,
    );
    let qlinear = QuantizedLinear::from_fp(&weight, &params).expect("from_fp bias test");

    // Wrong bias length must fail
    let bad_bias = ndarray::Array1::zeros(3_usize); // out_features == 4
    assert!(qlinear.with_bias(bad_bias).is_err());
}

#[test]
fn test_bias_forward_correct() {
    let weight = sample_weight_4x8();
    let params = calibrate_linear(
        &weight,
        QuantizationType::Int8,
        QuantizationGranularity::PerTensor,
    );
    let qlinear = QuantizedLinear::from_fp(&weight, &params).expect("from_fp bias forward");

    let bias = ndarray::Array1::from_vec(vec![1.0, 2.0, 3.0, 4.0]);
    let qlinear_b = qlinear.with_bias(bias.clone()).expect("with_bias");

    let x = Array2::from_shape_fn((2, 8), |(i, j)| (i + j) as f64 * 0.1);
    let out_no_bias = QuantizedLinear::from_fp(&weight, &params)
        .expect("from_fp no-bias")
        .forward(&x);
    let out_with_bias = qlinear_b.forward(&x);

    // Each row of output_with_bias == output_no_bias + bias
    for batch in 0..2 {
        for ch in 0..4 {
            let expected = out_no_bias[[batch, ch]] + bias[ch];
            let got = out_with_bias[[batch, ch]];
            assert!(
                (expected - got).abs() < 1e-12,
                "bias mismatch at [{batch},{ch}]: expected={expected} got={got}"
            );
        }
    }
}

// -----------------------------------------------------------------------
// Test 7: invalid qtype returns error
// -----------------------------------------------------------------------

#[test]
fn test_invalid_qtype_returns_error() {
    let weight = sample_weight_4x8();
    let params = QuantizationParams {
        qtype: QuantizationType::Int4,
        scheme: QuantizationScheme::Symmetric,
        granularity: QuantizationGranularity::PerTensor,
        scale: vec![1.0],
        zero_point: vec![0],
        min_val: vec![-1.0],
        max_val: vec![1.0],
    };
    assert!(
        QuantizedLinear::from_fp(&weight, &params).is_err(),
        "Int4 should be rejected"
    );
}