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

//! Tests for the GGUF reader. Included only under #[cfg(test)] from reader.rs.

use super::super::io::{GGUF_MAGIC, align_offset};
use super::super::types::{GgmlType, GgufValueType};
use super::*;
use crate::test_utils::cpu_setup;
use numr::runtime::cpu::CpuRuntime;
use std::io::Write;
use tempfile::NamedTempFile;

/// Helper: write a GGUF string (u64 length + bytes)
fn write_str(buf: &mut Vec<u8>, s: &str) {
    buf.extend_from_slice(&(s.len() as u64).to_le_bytes());
    buf.extend_from_slice(s.as_bytes());
}

/// Build a minimal GGUF v3 byte buffer with one F32 tensor and one Q4_0 tensor.
fn create_test_gguf_bytes() -> Vec<u8> {
    let mut buf = Vec::new();

    // Magic
    buf.extend_from_slice(&GGUF_MAGIC.to_le_bytes());
    // Version 3
    buf.extend_from_slice(&3u32.to_le_bytes());
    // 2 tensors
    buf.extend_from_slice(&2u64.to_le_bytes());
    // 2 KV pairs
    buf.extend_from_slice(&2u64.to_le_bytes());

    // KV 1: general.architecture = "test"
    write_str(&mut buf, "general.architecture");
    buf.extend_from_slice(&(GgufValueType::String as u32).to_le_bytes());
    write_str(&mut buf, "test");

    // KV 2: test.block_count = 4
    write_str(&mut buf, "test.block_count");
    buf.extend_from_slice(&(GgufValueType::Uint32 as u32).to_le_bytes());
    buf.extend_from_slice(&4u32.to_le_bytes());

    // Tensor 1: "weight_f32" F32 [4]
    write_str(&mut buf, "weight_f32");
    buf.extend_from_slice(&1u32.to_le_bytes()); // n_dims
    buf.extend_from_slice(&4u64.to_le_bytes()); // dim[0]
    buf.extend_from_slice(&(GgmlType::F32 as u32).to_le_bytes());
    buf.extend_from_slice(&0u64.to_le_bytes()); // offset

    // Tensor 2: "weight_q4" Q4_0 [32]
    write_str(&mut buf, "weight_q4");
    buf.extend_from_slice(&1u32.to_le_bytes());
    buf.extend_from_slice(&32u64.to_le_bytes());
    buf.extend_from_slice(&(GgmlType::Q4_0 as u32).to_le_bytes());
    buf.extend_from_slice(&16u64.to_le_bytes()); // offset after 4 floats

    // Align to 32 bytes
    let aligned = buf.len().div_ceil(32) * 32;
    buf.resize(aligned, 0);

    // Data: weight_f32 = [1.0, 2.0, 3.0, 4.0]
    for f in [1.0f32, 2.0, 3.0, 4.0] {
        buf.extend_from_slice(&f.to_le_bytes());
    }

    // Data: weight_q4 - Q4_0 block (scale=1.0, all nibbles=8 -> dequant to 0)
    let scale_bits = half::f16::from_f32(1.0).to_bits();
    buf.push((scale_bits & 0xFF) as u8);
    buf.push(((scale_bits >> 8) & 0xFF) as u8);
    buf.extend(std::iter::repeat_n(0x88u8, 16));

    buf
}

/// Create a minimal GGUF v3 file with one F32 tensor and one Q4_0 tensor
fn create_test_gguf() -> NamedTempFile {
    let buf = create_test_gguf_bytes();
    let mut file = NamedTempFile::new().unwrap();
    file.write_all(&buf).unwrap();
    file.flush().unwrap();
    file
}

#[test]
fn test_open_gguf() {
    let f = create_test_gguf();
    let gguf = Gguf::open(f.path()).unwrap();
    assert_eq!(gguf.version(), 3);
    assert_eq!(gguf.len(), 2);
    assert_eq!(gguf.metadata().architecture(), Some("test"));
    assert_eq!(gguf.metadata().block_count(), Some(4));
}

#[test]
fn test_tensor_info_gguf() {
    let f = create_test_gguf();
    let gguf = Gguf::open(f.path()).unwrap();

    let f32_info = gguf.tensor_info("weight_f32").unwrap();
    assert_eq!(f32_info.shape, vec![4]);
    assert_eq!(f32_info.ggml_type, GgmlType::F32);
    assert_eq!(f32_info.size_bytes(), 16);

    let q4_info = gguf.tensor_info("weight_q4").unwrap();
    assert_eq!(q4_info.shape, vec![32]);
    assert_eq!(q4_info.ggml_type, GgmlType::Q4_0);
    assert_eq!(q4_info.size_bytes(), 18);
}

#[test]
fn test_load_f32_tensor() {
    let (_, device) = cpu_setup();
    let f = create_test_gguf();
    let mut gguf = Gguf::open(f.path()).unwrap();

    let tensor = gguf
        .load_tensor_f32::<CpuRuntime>("weight_f32", &device)
        .unwrap();
    // GGUF 1D tensor: shape reversed is still [4]
    assert_eq!(tensor.shape(), &[4]);
    let data = tensor.to_vec::<f32>();
    assert!((data[0] - 1.0).abs() < 1e-6);
    assert!((data[3] - 4.0).abs() < 1e-6);
}

#[test]
fn test_load_quantized_tensor() {
    let f = create_test_gguf();
    let mut gguf = Gguf::open(f.path()).unwrap();

    let device = numr::runtime::cpu::CpuDevice::new();
    let qt = gguf
        .load_tensor_quantized::<numr::runtime::cpu::CpuRuntime>("weight_q4", &device)
        .unwrap();
    assert_eq!(qt.shape(), &[32]);
    assert_eq!(qt.format(), crate::quant::QuantFormat::Q4_0);
}

#[test]
fn test_align_offset() {
    assert_eq!(align_offset(0, 32), 0);
    assert_eq!(align_offset(1, 32), 32);
    assert_eq!(align_offset(32, 32), 32);
    assert_eq!(align_offset(33, 32), 64);
}

#[test]
fn test_tensor_not_found() {
    let f = create_test_gguf();
    let gguf = Gguf::open(f.path()).unwrap();
    assert!(gguf.tensor_info("nonexistent").is_err());
}

#[test]
fn test_open_with_mmap() {
    let (_, device) = cpu_setup();
    let f = create_test_gguf();
    let mut gguf = Gguf::open_with_mmap(f.path(), true).unwrap();
    assert_eq!(gguf.version(), 3);
    let tensor = gguf
        .load_tensor_f32::<CpuRuntime>("weight_f32", &device)
        .unwrap();
    assert_eq!(tensor.shape(), &[4]);
    let data = tensor.to_vec::<f32>();
    assert!((data[0] - 1.0).abs() < 1e-6);
    assert!((data[3] - 4.0).abs() < 1e-6);
}

#[test]
fn test_open_without_mmap() {
    let (_, device) = cpu_setup();
    let f = create_test_gguf();
    // use_mmap=false should behave identically to open()
    let mut gguf = Gguf::open_with_mmap(f.path(), false).unwrap();
    let tensor = gguf
        .load_tensor_f32::<CpuRuntime>("weight_f32", &device)
        .unwrap();
    let data = tensor.to_vec::<f32>();
    assert!((data[0] - 1.0).abs() < 1e-6);
}

#[test]
fn test_from_bytes() {
    let buf = create_test_gguf_bytes();
    let gguf = Gguf::from_bytes(buf).unwrap();
    assert_eq!(gguf.version(), 3);
    assert_eq!(gguf.len(), 2);
    assert_eq!(gguf.metadata().architecture(), Some("test"));
}

#[test]
fn test_from_bytes_load_f32() {
    let (_, device) = cpu_setup();
    let buf = create_test_gguf_bytes();
    let mut gguf = Gguf::from_bytes(buf).unwrap();

    let tensor = gguf
        .load_tensor_f32::<CpuRuntime>("weight_f32", &device)
        .unwrap();
    assert_eq!(tensor.shape(), &[4]);
    let data = tensor.to_vec::<f32>();
    assert!((data[0] - 1.0).abs() < 1e-6);
    assert!((data[3] - 4.0).abs() < 1e-6);
}

#[test]
fn test_from_bytes_load_quantized() {
    let buf = create_test_gguf_bytes();
    let mut gguf = Gguf::from_bytes(buf).unwrap();

    let device = numr::runtime::cpu::CpuDevice::new();
    let qt = gguf
        .load_tensor_quantized::<CpuRuntime>("weight_q4", &device)
        .unwrap();
    assert_eq!(qt.shape(), &[32]);
    assert_eq!(qt.format(), crate::quant::QuantFormat::Q4_0);
}