modelc 0.1.1

Compile model weight files to standalone executable binaries
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

//! Shared fixtures for integration tests; not every helper is referenced by each test binary.

#![allow(dead_code)]

use std::collections::HashMap;
use std::io::Write;
use std::path::Path;

use modelc::model::{DataType, Model, TensorData};

pub fn create_test_model() -> Model {
    let mut tensors = HashMap::new();

    let weight_data: Vec<u8> = [1.0f32, 2.0, 3.0, 4.0, 5.0, 6.0]
        .iter()
        .flat_map(|f| f.to_le_bytes())
        .collect();
    tensors.insert(
        "weight".to_string(),
        TensorData {
            shape: vec![2, 3],
            dtype: DataType::F32,
            data: weight_data,
        },
    );

    let bias_data: Vec<u8> = [0.1f32, 0.2].iter().flat_map(|f| f.to_le_bytes()).collect();
    tensors.insert(
        "bias".to_string(),
        TensorData {
            shape: vec![2],
            dtype: DataType::F32,
            data: bias_data,
        },
    );

    let mut metadata = HashMap::new();
    metadata.insert("source".to_string(), "test".to_string());

    Model {
        name: "test_model".to_string(),
        architecture: "mlp".to_string(),
        tensors,
        metadata,
    }
}

pub fn create_large_test_model() -> Model {
    let mut tensors = HashMap::new();

    let hidden_dim = 64usize;
    let weight_data: Vec<u8> = (0..hidden_dim * hidden_dim)
        .flat_map(|i| (i as f32 / 100.0).to_le_bytes())
        .collect();
    tensors.insert(
        "layer0.weight".to_string(),
        TensorData {
            shape: vec![hidden_dim, hidden_dim],
            dtype: DataType::F32,
            data: weight_data,
        },
    );

    let bias_data: Vec<u8> = (0..hidden_dim)
        .flat_map(|i| (i as f32 / 100.0).to_le_bytes())
        .collect();
    tensors.insert(
        "layer0.bias".to_string(),
        TensorData {
            shape: vec![hidden_dim],
            dtype: DataType::F32,
            data: bias_data,
        },
    );

    let ln_weight: Vec<u8> = (0..hidden_dim).flat_map(|_| 1.0f32.to_le_bytes()).collect();
    tensors.insert(
        "layer0.ln_weight".to_string(),
        TensorData {
            shape: vec![hidden_dim],
            dtype: DataType::F32,
            data: ln_weight,
        },
    );

    let ln_bias: Vec<u8> = (0..hidden_dim).flat_map(|_| 0.0f32.to_le_bytes()).collect();
    tensors.insert(
        "layer0.ln_bias".to_string(),
        TensorData {
            shape: vec![hidden_dim],
            dtype: DataType::F32,
            data: ln_bias,
        },
    );

    tensors.insert(
        "layer1.weight".to_string(),
        TensorData {
            shape: vec![hidden_dim, hidden_dim],
            dtype: DataType::F32,
            data: vec![0u8; hidden_dim * hidden_dim * 4],
        },
    );

    tensors.insert(
        "layer1.bias".to_string(),
        TensorData {
            shape: vec![hidden_dim],
            dtype: DataType::F32,
            data: vec![0u8; hidden_dim * 4],
        },
    );

    Model {
        name: "large_test_model".to_string(),
        architecture: "mlp".to_string(),
        tensors,
        metadata: HashMap::new(),
    }
}

pub fn create_safetensors_file(path: &Path, tensors: Vec<(&str, &str, Vec<usize>, Vec<u8>)>) {
    let mut sorted: Vec<_> = tensors.into_iter().collect();
    sorted.sort_by(|a, b| a.0.cmp(b.0));

    let mut header = serde_json::Map::new();
    header.insert(
        "__metadata__".to_string(),
        serde_json::Value::Object(serde_json::Map::new()),
    );

    let mut offset = 0usize;
    let mut entries = Vec::new();
    for (name, dtype, shape, data) in sorted {
        let end = offset + data.len();
        entries.push((
            name.to_string(),
            dtype.to_string(),
            shape,
            data,
            offset,
            end,
        ));
        offset = end;
    }

    for (name, dtype, shape, _, start, end) in &entries {
        let mut obj = serde_json::Map::new();
        obj.insert(
            "dtype".to_string(),
            serde_json::Value::String(dtype.clone()),
        );
        obj.insert("shape".to_string(), serde_json::json!(shape));
        obj.insert(
            "data_offsets".to_string(),
            serde_json::json!([*start, *end]),
        );
        header.insert(name.clone(), serde_json::Value::Object(obj));
    }

    let header_json = serde_json::to_string(&header).unwrap();
    let header_bytes = header_json.as_bytes();

    let mut file = std::fs::File::create(path).unwrap();
    file.write_all(&(header_bytes.len() as u64).to_le_bytes())
        .unwrap();
    file.write_all(header_bytes).unwrap();
    for (_, _, _, data, _, _) in &entries {
        file.write_all(data).unwrap();
    }
}

pub fn f32_to_bytes(values: &[f32]) -> Vec<u8> {
    values.iter().flat_map(|f| f.to_le_bytes()).collect()
}

pub fn bytes_to_f32(bytes: &[u8]) -> Vec<f32> {
    bytes
        .chunks_exact(4)
        .map(|c| f32::from_le_bytes([c[0], c[1], c[2], c[3]]))
        .collect()
}