boostr 0.1.0

ML framework built on numr - attention, quantization, model architectures
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

//! Tests for CPU QuantMatmulOps implementation.

use crate::quant::QuantTensor;
use crate::quant::format::QuantFormat;
use crate::quant::traits::{DequantOps, QuantMatmulOps};
use half::f16;
use numr::dtype::DType;
use numr::ops::MatmulOps;
use numr::runtime::cpu::{CpuClient, CpuDevice, CpuRuntime};
use numr::tensor::Tensor;

fn setup() -> (CpuClient, CpuDevice) {
    let device = CpuDevice::new();
    let client = CpuClient::new(device.clone());
    (client, device)
}

#[test]
fn test_quant_matmul_q4_0_basic() {
    let (client, device) = setup();

    // activation [1, 32], weight [1, 32] → output [1, 1]
    let act = Tensor::<CpuRuntime>::from_slice(&[1.0f32; 32], &[1, 32], &device);

    let mut block = [0u8; 18];
    block[0..2].copy_from_slice(&f16::from_f32(2.0).to_le_bytes());
    block[2..18].fill(0x99); // dequant value = 2.0

    let qt = QuantTensor::<CpuRuntime>::from_bytes(&block, QuantFormat::Q4_0, &[1, 32], &device)
        .unwrap();

    let result = client.quant_matmul(&act, &qt).unwrap();
    assert_eq!(result.shape(), &[1, 1]);

    let data = result.to_vec::<f32>();
    assert!(
        (data[0] - 64.0).abs() < 0.5,
        "expected ~64.0, got {}",
        data[0]
    );
}

#[test]
fn test_quant_matmul_matches_dequant_matmul() {
    let (client, device) = setup();

    // activation [2, 32]
    let act_data: Vec<f32> = (0..64).map(|i| (i as f32) * 0.1).collect();
    let act = Tensor::<CpuRuntime>::from_slice(&act_data, &[2, 32], &device);

    // weight [3, 32] as Q8_0 (3 rows, each 34 bytes)
    let mut weight_bytes = Vec::new();
    for row in 0..3 {
        let mut block = [0u8; 34];
        block[0..2].copy_from_slice(&f16::from_f32(0.5).to_le_bytes());
        // Different qs per row for variety
        block[2..34].fill((row + 1) as u8);
        weight_bytes.extend_from_slice(&block);
    }

    let qt =
        QuantTensor::<CpuRuntime>::from_bytes(&weight_bytes, QuantFormat::Q8_0, &[3, 32], &device)
            .unwrap();

    // Method 1: quant_matmul
    let result_qm = client.quant_matmul(&act, &qt).unwrap();

    // Method 2: dequant then matmul
    let dequant_w = client.dequantize(&qt, DType::F32).unwrap();
    // dequant gives [3, 32], matmul needs act [2, 32] × w^T [32, 3]
    // Our quant_matmul does act × w^T layout, so we need to transpose for standard matmul
    let dequant_w_t = dequant_w.transpose(0isize, 1isize).unwrap();
    let result_dm = MatmulOps::matmul(&client, &act, &dequant_w_t).unwrap();

    assert_eq!(result_qm.shape(), result_dm.shape());

    let qm_data = result_qm.to_vec::<f32>();
    let dm_data = result_dm.to_vec::<f32>();
    for (i, (&a, &b)) in qm_data.iter().zip(dm_data.iter()).enumerate() {
        assert!(
            (a - b).abs() < 1e-2,
            "mismatch at index {}: quant_matmul={}, dequant+matmul={}",
            i,
            a,
            b
        );
    }
}

#[test]
fn test_quant_matmul_dim_mismatch() {
    let (client, device) = setup();

    let act = Tensor::<CpuRuntime>::from_slice(&vec![1.0f32; 64], &[2, 32], &device);

    // Weight K=64 ≠ activation K=32
    let block = vec![0u8; 2 * 34]; // 2 blocks of Q8_0 = 64 elements
    let qt = QuantTensor::<CpuRuntime>::from_bytes(&block, QuantFormat::Q8_0, &[1, 64], &device)
        .unwrap();

    let result = client.quant_matmul(&act, &qt);
    assert!(result.is_err());
}

#[test]
fn test_quant_matmul_q2k_basic() {
    let (client, device) = setup();

    let act = Tensor::<CpuRuntime>::from_slice(&vec![1.0f32; 256], &[1, 256], &device);

    // Q2K: 256 elements, 84 bytes/block — all zeros dequantizes to zeros
    let block = vec![0u8; 84];
    let qt = QuantTensor::<CpuRuntime>::from_bytes(&block, QuantFormat::Q2K, &[1, 256], &device)
        .unwrap();

    let result = client.quant_matmul(&act, &qt).unwrap();
    assert_eq!(result.shape(), &[1, 1]);
    let data = result.to_vec::<f32>();
    assert!(data[0].abs() < 1e-5, "expected ~0.0, got {}", data[0]);
}

#[test]
#[allow(clippy::needless_range_loop)]
fn test_quant_matmul_q2k_matches_dequant_matmul() {
    let (client, device) = setup();

    // K=512 (2 Q2K blocks), N=2, M=2
    let k = 512;
    let n = 2;
    let m = 2;

    let act_data: Vec<f32> = (0..m * k).map(|i| ((i % 17) as f32 - 8.0) * 0.01).collect();
    let act = Tensor::<CpuRuntime>::from_slice(&act_data, &[m, k], &device);

    // Build Q2K blocks with non-trivial data
    let mut weight_bytes = Vec::new();
    for row in 0..n {
        for blk in 0..2 {
            let mut block = [0u8; 84];
            // scales: non-zero low nibble (sub-scale) and high nibble (sub-min)
            for i in 0..16 {
                let s = ((i + row + blk) % 15 + 1) as u8; // 1-15
                let m_val = ((i + row * 3 + blk) % 10) as u8; // 0-9
                block[i] = s | (m_val << 4);
            }
            // qs: non-trivial 2-bit values packed in bytes
            for i in 0..64 {
                block[16 + i] = ((i + row * 7 + blk * 3) % 256) as u8;
            }
            // d = 0.5, dmin = 0.1
            block[80..82].copy_from_slice(&f16::from_f32(0.5).to_le_bytes());
            block[82..84].copy_from_slice(&f16::from_f32(0.1).to_le_bytes());
            weight_bytes.extend_from_slice(&block);
        }
    }

    let qt =
        QuantTensor::<CpuRuntime>::from_bytes(&weight_bytes, QuantFormat::Q2K, &[n, k], &device)
            .unwrap();

    // Method 1: quant_matmul (generic path)
    let result_qm = client.quant_matmul(&act, &qt).unwrap();

    // Method 2: dequant then matmul
    let dequant_w = client.dequantize(&qt, DType::F32).unwrap();
    let dequant_w_t = dequant_w.transpose(0isize, 1isize).unwrap();
    let result_dm = MatmulOps::matmul(&client, &act, &dequant_w_t).unwrap();

    assert_eq!(result_qm.shape(), result_dm.shape());

    let qm_data = result_qm.to_vec::<f32>();
    let dm_data = result_dm.to_vec::<f32>();
    for (i, (&a, &b)) in qm_data.iter().zip(dm_data.iter()).enumerate() {
        let tol = 0.05 * b.abs().max(1.0);
        assert!(
            (a - b).abs() < tol,
            "Q2K mismatch at index {}: quant_matmul={}, dequant+matmul={}, tol={}",
            i,
            a,
            b,
            tol
        );
    }
}

#[test]
#[allow(clippy::needless_range_loop)]
fn test_quant_matmul_q3k_matches_dequant_matmul() {
    let (client, device) = setup();

    // K=512 (2 Q3K blocks), N=3, M=2
    let k = 512;
    let n = 3;
    let m = 2;

    let act_data: Vec<f32> = (0..m * k).map(|i| ((i % 13) as f32 - 6.0) * 0.01).collect();
    let act = Tensor::<CpuRuntime>::from_slice(&act_data, &[m, k], &device);

    // Build Q3K blocks with non-trivial data
    let mut weight_bytes = Vec::new();
    for row in 0..n {
        for blk in 0..2 {
            let mut block = [0u8; 110];
            // hmask[32]: non-trivial high bits
            for i in 0..32 {
                block[i] = ((i * 7 + row * 13 + blk * 5) % 256) as u8;
            }
            // qs[64]: non-trivial 2-bit values
            for i in 0..64 {
                block[32 + i] = ((i * 11 + row * 3 + blk * 7) % 256) as u8;
            }
            // scales[12]: non-trivial packed 6-bit scales
            for i in 0..12 {
                block[96 + i] = ((i * 5 + row * 9 + blk) % 256) as u8;
            }
            // d = 0.3
            block[108..110].copy_from_slice(&f16::from_f32(0.3).to_le_bytes());
            weight_bytes.extend_from_slice(&block);
        }
    }

    let qt =
        QuantTensor::<CpuRuntime>::from_bytes(&weight_bytes, QuantFormat::Q3K, &[n, k], &device)
            .unwrap();

    // Method 1: quant_matmul (generic path)
    let result_qm = client.quant_matmul(&act, &qt).unwrap();

    // Method 2: dequant then matmul
    let dequant_w = client.dequantize(&qt, DType::F32).unwrap();
    let dequant_w_t = dequant_w.transpose(0isize, 1isize).unwrap();
    let result_dm = MatmulOps::matmul(&client, &act, &dequant_w_t).unwrap();

    assert_eq!(result_qm.shape(), result_dm.shape());

    let qm_data = result_qm.to_vec::<f32>();
    let dm_data = result_dm.to_vec::<f32>();
    for (i, (&a, &b)) in qm_data.iter().zip(dm_data.iter()).enumerate() {
        let tol = 0.05 * b.abs().max(1.0);
        assert!(
            (a - b).abs() < tol,
            "Q3K mismatch at index {}: quant_matmul={}, dequant+matmul={}, tol={}",
            i,
            a,
            b,
            tol
        );
    }
}