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

//! Backend parity tests for NF4 operations

use super::helpers::*;
use boostr::DequantOps;
use numr::runtime::cpu::CpuRuntime;
use numr::tensor::Tensor;

#[test]
fn test_nf4_dequant_codebook_values() {
    let (cpu_client, cpu_device) = setup_cpu();

    // Create data with known codebook indices: all zeros (idx 0 = 0.0)
    let nf4_bytes = vec![0u8; 16]; // 16 bytes = 32 elements, all index 0
    let nf4_data = Tensor::<CpuRuntime>::from_slice(&nf4_bytes, &[16], &cpu_device);
    let absmax = Tensor::<CpuRuntime>::from_slice(&[1.0f32], &[1], &cpu_device);

    let result = cpu_client.nf4_dequant(&nf4_data, &absmax, 32).unwrap();
    assert_eq!(result.shape(), &[32]);

    let data = result.to_vec::<f32>();
    for &v in &data {
        assert!((v - 0.0).abs() < 1e-6, "expected 0.0, got {}", v);
    }
}

#[test]
fn test_nf4_dequant_scaling() {
    let (cpu_client, cpu_device) = setup_cpu();

    // Index 15 = codebook value 1.0, absmax = 2.0 → output = 2.0
    let nf4_bytes = vec![0xFFu8; 16]; // all index 15
    let nf4_data = Tensor::<CpuRuntime>::from_slice(&nf4_bytes, &[16], &cpu_device);
    let absmax = Tensor::<CpuRuntime>::from_slice(&[2.0f32], &[1], &cpu_device);

    let result = cpu_client.nf4_dequant(&nf4_data, &absmax, 32).unwrap();
    let data = result.to_vec::<f32>();
    for &v in &data {
        assert!((v - 2.0).abs() < 1e-6, "expected 2.0, got {}", v);
    }
}

#[test]
fn test_nf4_gemm_parity() {
    let (cpu_client, cpu_device) = setup_cpu();

    let m = 2;
    let k = 32;
    let n = 8;
    let blocksize = 32;

    let input = det_tensor(&[m, k], &cpu_device);

    // NF4 weight: [N*K/2] bytes
    let num_bytes = n * k / 2;
    let nf4_bytes: Vec<u8> = (0..num_bytes)
        .map(|i| {
            let lo = (i % 16) as u8;
            let hi = ((i + 5) % 16) as u8;
            (hi << 4) | lo
        })
        .collect();
    let nf4_weight = Tensor::<CpuRuntime>::from_slice(&nf4_bytes, &[num_bytes], &cpu_device);

    // Absmax: [N * K/blocksize]
    let num_absmax = n * k / blocksize;
    let absmax_data: Vec<f32> = (0..num_absmax).map(|i| 0.5 + i as f32 * 0.1).collect();
    let absmax = Tensor::<CpuRuntime>::from_slice(&absmax_data, &[num_absmax], &cpu_device);

    let cpu_result = cpu_client
        .nf4_gemm(&input, &nf4_weight, &absmax, n, k, blocksize)
        .unwrap();
    assert_eq!(cpu_result.shape(), &[m, n]);

    let cpu_vec = cpu_result.to_vec::<f32>();
    for &v in &cpu_vec {
        assert!(v.is_finite(), "non-finite value: {}", v);
    }

    #[cfg(feature = "cuda")]
    with_cuda_backend(|cuda_client, cuda_device| {
        use boostr::DequantOps as _;
        use numr::tensor::Tensor;

        let input_c = Tensor::from_slice(&input.to_vec::<f32>(), input.shape(), &cuda_device);
        let nf4_c =
            Tensor::from_slice(&nf4_weight.to_vec::<u8>(), nf4_weight.shape(), &cuda_device);
        let absmax_c = Tensor::from_slice(&absmax.to_vec::<f32>(), absmax.shape(), &cuda_device);

        let cuda_result = cuda_client
            .nf4_gemm(&input_c, &nf4_c, &absmax_c, n, k, blocksize)
            .unwrap();
        assert_parity_f32_relaxed(
            &cuda_result.to_vec::<f32>(),
            &cpu_vec,
            "nf4_gemm CUDA vs CPU",
        );
    });

    #[cfg(feature = "wgpu")]
    with_wgpu_backend(|wgpu_client, wgpu_device| {
        use boostr::DequantOps as _;
        use numr::tensor::Tensor;

        let input_w = Tensor::from_slice(&input.to_vec::<f32>(), input.shape(), &wgpu_device);
        let nf4_w =
            Tensor::from_slice(&nf4_weight.to_vec::<u8>(), nf4_weight.shape(), &wgpu_device);
        let absmax_w = Tensor::from_slice(&absmax.to_vec::<f32>(), absmax.shape(), &wgpu_device);

        let wgpu_result = wgpu_client
            .nf4_gemm(&input_w, &nf4_w, &absmax_w, n, k, blocksize)
            .unwrap();
        assert_parity_f32_relaxed(
            &wgpu_result.to_vec::<f32>(),
            &cpu_vec,
            "nf4_gemm WebGPU vs CPU",
        );
    });
}

#[test]
fn test_nf4_gemm_matches_dequant_matmul() {
    use numr::ops::MatmulOps;

    let (cpu_client, cpu_device) = setup_cpu();

    let m = 2;
    let k = 32;
    let n = 4;
    let blocksize = 32;

    let input = det_tensor(&[m, k], &cpu_device);

    // Create NF4 weight data
    let num_bytes = n * k / 2;
    let nf4_bytes: Vec<u8> = (0..num_bytes)
        .map(|i| {
            let lo = (i % 16) as u8;
            let hi = ((i + 7) % 16) as u8;
            (hi << 4) | lo
        })
        .collect();
    let nf4_weight = Tensor::<CpuRuntime>::from_slice(&nf4_bytes, &[num_bytes], &cpu_device);

    let num_absmax = n * k / blocksize;
    let absmax_data: Vec<f32> = (0..num_absmax).map(|i| 1.0 + i as f32 * 0.05).collect();
    let absmax = Tensor::<CpuRuntime>::from_slice(&absmax_data, &[num_absmax], &cpu_device);

    // Method 1: Fused NF4 GEMM
    let fused_result = cpu_client
        .nf4_gemm(&input, &nf4_weight, &absmax, n, k, blocksize)
        .unwrap();

    // Method 2: Dequant then matmul
    let dequant_weight = cpu_client
        .nf4_dequant(&nf4_weight, &absmax, blocksize)
        .unwrap();
    // Reshape to [N, K] and transpose to [K, N] for matmul
    let dequant_2d = dequant_weight.reshape(&[n, k]).unwrap();
    let dequant_t = dequant_2d.transpose(0isize, 1isize).unwrap().contiguous();
    let ref_result = MatmulOps::matmul(&cpu_client, &input, &dequant_t).unwrap();

    let cpu_fused_vec = fused_result.to_vec::<f32>();
    assert_parity_f32_relaxed(
        &cpu_fused_vec,
        &ref_result.to_vec::<f32>(),
        "nf4_gemm vs dequant+matmul",
    );

    #[cfg(feature = "cuda")]
    with_cuda_backend(|cuda_client, cuda_device| {
        use boostr::DequantOps as _;
        use numr::tensor::Tensor;

        let input_c = Tensor::from_slice(&input.to_vec::<f32>(), input.shape(), &cuda_device);
        let nf4_c =
            Tensor::from_slice(&nf4_weight.to_vec::<u8>(), nf4_weight.shape(), &cuda_device);
        let absmax_c = Tensor::from_slice(&absmax.to_vec::<f32>(), absmax.shape(), &cuda_device);

        let cuda_fused = cuda_client
            .nf4_gemm(&input_c, &nf4_c, &absmax_c, n, k, blocksize)
            .unwrap();
        assert_parity_f32_relaxed(
            &cuda_fused.to_vec::<f32>(),
            &cpu_fused_vec,
            "nf4_gemm fused CUDA vs CPU",
        );
    });

    #[cfg(feature = "wgpu")]
    with_wgpu_backend(|wgpu_client, wgpu_device| {
        use boostr::DequantOps as _;
        use numr::tensor::Tensor;

        let input_w = Tensor::from_slice(&input.to_vec::<f32>(), input.shape(), &wgpu_device);
        let nf4_w =
            Tensor::from_slice(&nf4_weight.to_vec::<u8>(), nf4_weight.shape(), &wgpu_device);
        let absmax_w = Tensor::from_slice(&absmax.to_vec::<f32>(), absmax.shape(), &wgpu_device);

        let wgpu_fused = wgpu_client
            .nf4_gemm(&input_w, &nf4_w, &absmax_w, n, k, blocksize)
            .unwrap();
        assert_parity_f32_relaxed(
            &wgpu_fused.to_vec::<f32>(),
            &cpu_fused_vec,
            "nf4_gemm fused WebGPU vs CPU",
        );
    });
}