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
250
251
252
253
254
255
256
257

//! CPU implementation of DequantOps

use crate::error::{Error, Result};
use crate::quant::traits::DequantOps;
use crate::quant::{QuantFormat, QuantTensor};
use numr::dtype::DType;
use numr::ops::TypeConversionOps;
use numr::runtime::cpu::{CpuClient, CpuRuntime};
use numr::tensor::Tensor;

use super::kernels::{dequant, nf4};

impl DequantOps<CpuRuntime> for CpuClient {
    fn nf4_dequant(
        &self,
        nf4_data: &Tensor<CpuRuntime>,
        absmax: &Tensor<CpuRuntime>,
        blocksize: usize,
    ) -> Result<Tensor<CpuRuntime>> {
        if nf4_data.dtype() != DType::U8 {
            return Err(Error::QuantError {
                reason: format!("nf4_dequant data must be U8, got {:?}", nf4_data.dtype()),
            });
        }
        let data = unsafe { nf4_data.storage().as_host_slice::<u8>() };
        let abs = unsafe { absmax.storage().as_host_slice::<f32>() };
        let n = data.len() * 2;
        let mut output = vec![0.0f32; n];
        nf4::nf4_dequant_f32(data, abs, blocksize, &mut output);
        Ok(Tensor::<CpuRuntime>::from_slice(
            &output,
            &[n],
            nf4_data.device(),
        ))
    }

    fn nf4_gemm(
        &self,
        input: &Tensor<CpuRuntime>,
        nf4_weight: &Tensor<CpuRuntime>,
        absmax: &Tensor<CpuRuntime>,
        n_out: usize,
        k: usize,
        blocksize: usize,
    ) -> Result<Tensor<CpuRuntime>> {
        if input.dtype() != DType::F32 {
            return Err(Error::QuantError {
                reason: format!("nf4_gemm input must be F32, got {:?}", input.dtype()),
            });
        }
        let in_shape = input.shape();
        let m: usize = in_shape.iter().product::<usize>() / k;
        let inp = unsafe { input.storage().as_host_slice::<f32>() };
        let wt = unsafe { nf4_weight.storage().as_host_slice::<u8>() };
        let abs = unsafe { absmax.storage().as_host_slice::<f32>() };
        let mut output = vec![0.0f32; m * n_out];
        nf4::nf4_gemm_f32(inp, wt, abs, &mut output, m, k, n_out, blocksize);
        let mut out_shape = in_shape[..in_shape.len() - 1].to_vec();
        out_shape.push(n_out);
        Ok(Tensor::<CpuRuntime>::from_slice(
            &output,
            &out_shape,
            input.device(),
        ))
    }

    fn dequantize(
        &self,
        qt: &QuantTensor<CpuRuntime>,
        target_dtype: DType,
    ) -> Result<Tensor<CpuRuntime>> {
        // Validate target is floating point
        if !matches!(
            target_dtype,
            DType::F32 | DType::F16 | DType::BF16 | DType::F64
        ) {
            return Err(Error::QuantError {
                reason: format!("dequantize target must be float, got {:?}", target_dtype),
            });
        }

        let numel = qt.numel();

        // Read raw block bytes from storage (zero-copy for CPU)
        // SAFETY: CpuRuntime stores data as host pointers
        let block_bytes = unsafe { qt.storage().as_host_slice::<u8>() };

        // Dequantize to f32 first
        let mut f32_output = vec![0.0f32; numel];

        match qt.format() {
            QuantFormat::Q4_0 => dequant::dequant_q4_0(block_bytes, &mut f32_output),
            QuantFormat::Q4_1 => dequant::dequant_q4_1(block_bytes, &mut f32_output),
            QuantFormat::Q5_0 => dequant::dequant_q5_0(block_bytes, &mut f32_output),
            QuantFormat::Q5_1 => dequant::dequant_q5_1(block_bytes, &mut f32_output),
            QuantFormat::Q8_0 => dequant::dequant_q8_0(block_bytes, &mut f32_output),
            QuantFormat::Q8_1 => dequant::dequant_q8_1(block_bytes, &mut f32_output),
            QuantFormat::Q2K => dequant::dequant_q2k(block_bytes, &mut f32_output),
            QuantFormat::Q3K => dequant::dequant_q3k(block_bytes, &mut f32_output),
            QuantFormat::Q4K => dequant::dequant_q4k(block_bytes, &mut f32_output),
            QuantFormat::Q5K => dequant::dequant_q5k(block_bytes, &mut f32_output),
            QuantFormat::Q6K => dequant::dequant_q6k(block_bytes, &mut f32_output),
            QuantFormat::Q8K => dequant::dequant_q8k(block_bytes, &mut f32_output),
            QuantFormat::IQ4NL => dequant::dequant_iq4_nl(block_bytes, &mut f32_output),
            QuantFormat::IQ4XS => dequant::dequant_iq4_xs(block_bytes, &mut f32_output),
            QuantFormat::IQ2XXS => dequant::dequant_iq2_xxs(block_bytes, &mut f32_output),
            QuantFormat::IQ2XS => dequant::dequant_iq2_xs(block_bytes, &mut f32_output),
            QuantFormat::IQ2S => dequant::dequant_iq2_s(block_bytes, &mut f32_output),
            QuantFormat::IQ3XXS => dequant::dequant_iq3_xxs(block_bytes, &mut f32_output),
            QuantFormat::IQ3S => dequant::dequant_iq3_s(block_bytes, &mut f32_output),
            QuantFormat::IQ1S => dequant::dequant_iq1_s(block_bytes, &mut f32_output),
            QuantFormat::IQ1M => dequant::dequant_iq1_m(block_bytes, &mut f32_output),
            QuantFormat::TQ1_0 => dequant::dequant_tq1_0(block_bytes, &mut f32_output),
            QuantFormat::TQ2_0 => dequant::dequant_tq2_0(block_bytes, &mut f32_output),
        }

        // Create f32 tensor
        let f32_tensor = Tensor::<CpuRuntime>::from_slice(&f32_output, qt.shape(), qt.device());

        // Cast to target dtype if needed
        if target_dtype == DType::F32 {
            Ok(f32_tensor)
        } else {
            self.cast(&f32_tensor, target_dtype).map_err(Error::Numr)
        }
    }
}

#[cfg(test)]
mod tests {
    use super::*;
    use half::f16;
    use numr::runtime::cpu::CpuDevice;

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

    #[test]
    fn test_dequant_q4_0_roundtrip() {
        let (client, device) = setup();

        // Create a Q4_0 block with known values
        // scale=2.0, all nibbles=9 → value = (9-8)*2.0 = 2.0
        let mut block = [0u8; 18];
        block[0..2].copy_from_slice(&f16::from_f32(2.0).to_le_bytes());
        block[2..18].fill(0x99);

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

        let result = client.dequantize(&qt, DType::F32).unwrap();
        assert_eq!(result.shape(), &[32]);
        assert_eq!(result.dtype(), DType::F32);

        let data = result.to_vec::<f32>();
        for &v in &data {
            assert!((v - 2.0).abs() < 0.01, "expected 2.0, got {}", v);
        }
    }

    #[test]
    fn test_dequant_q8_0_roundtrip() {
        let (client, device) = setup();

        let mut block = [0u8; 34];
        block[0..2].copy_from_slice(&f16::from_f32(0.5).to_le_bytes());
        // qs = 6 as i8 → value = 6 * 0.5 = 3.0
        block[2..34].fill(6);

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

        let result = client.dequantize(&qt, DType::F32).unwrap();
        let data = result.to_vec::<f32>();
        for &v in &data {
            assert!((v - 3.0).abs() < 0.01, "expected 3.0, got {}", v);
        }
    }

    #[test]
    fn test_dequant_q4k_basic() {
        let (client, device) = setup();

        // Minimal test: all zeros → output should be all zeros
        let block = vec![0u8; 144];
        let qt = QuantTensor::<CpuRuntime>::from_bytes(&block, QuantFormat::Q4K, &[256], &device)
            .unwrap();

        let result = client.dequantize(&qt, DType::F32).unwrap();
        assert_eq!(result.shape(), &[256]);
        let data = result.to_vec::<f32>();
        for &v in &data {
            assert!(v.abs() < 1e-5);
        }
    }

    #[test]
    fn test_dequant_q6k_basic() {
        let (client, device) = setup();

        let block = vec![0u8; 210];
        let qt = QuantTensor::<CpuRuntime>::from_bytes(&block, QuantFormat::Q6K, &[256], &device)
            .unwrap();

        let result = client.dequantize(&qt, DType::F32).unwrap();
        assert_eq!(result.shape(), &[256]);
    }

    #[test]
    fn test_dequant_to_f64() {
        let (client, device) = setup();

        let mut block = [0u8; 18];
        block[0..2].copy_from_slice(&f16::from_f32(1.0).to_le_bytes());
        block[2..18].fill(0x99); // value = 1.0

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

        let result = client.dequantize(&qt, DType::F64).unwrap();
        assert_eq!(result.dtype(), DType::F64);
    }

    #[test]
    fn test_dequant_iq1s_basic() {
        let (client, device) = setup();

        // IQ1S is now supported — test with all zeros
        let block = vec![0u8; 50];
        let qt = QuantTensor::<CpuRuntime>::from_bytes(&block, QuantFormat::IQ1S, &[256], &device)
            .unwrap();

        let result = client.dequantize(&qt, DType::F32);
        assert!(result.is_ok());
        let data = result.unwrap().to_vec::<f32>();
        assert_eq!(data.len(), 256);
        for &v in &data {
            assert!(v.abs() < 1e-5, "expected 0, got {}", v);
        }
    }

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

        let block = vec![0u8; 18];
        let qt = QuantTensor::<CpuRuntime>::from_bytes(&block, QuantFormat::Q4_0, &[32], &device)
            .unwrap();

        // I32 is not a valid dequant target
        let result = client.dequantize(&qt, DType::I32);
        assert!(result.is_err());
    }
}