ferrum-models 0.7.7

Model architectures (LLaMA, Qwen, BERT) for Ferrum inference
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

use ferrum_kernels::backend::cpu::CpuBackend;
use ferrum_kernels::backend::Backend;
use ferrum_models::{load_runtime_lora_adapter, ActiveLoraAdapter, LoraAdapterConfig};
use safetensors::tensor::{serialize_to_file, Dtype, TensorView};
use std::collections::HashMap;
use tempfile::TempDir;

fn f32_bytes(values: &[f32]) -> Vec<u8> {
    let mut out = Vec::with_capacity(values.len() * 4);
    for value in values {
        out.extend_from_slice(&value.to_le_bytes());
    }
    out
}

fn write_adapter(
    dir: &std::path::Path,
    config: &LoraAdapterConfig,
    tensors: Vec<(&str, Vec<usize>, Vec<f32>)>,
) {
    std::fs::write(
        dir.join("adapter_config.json"),
        serde_json::to_vec_pretty(config).unwrap(),
    )
    .unwrap();
    let encoded: Vec<(String, TensorView<'_>)> = tensors
        .iter()
        .map(|(name, shape, values)| {
            let bytes = f32_bytes(values);
            let leaked: &'static [u8] = Box::leak(bytes.into_boxed_slice());
            (
                (*name).to_string(),
                TensorView::new(Dtype::F32, shape.clone(), leaked).unwrap(),
            )
        })
        .collect();
    serialize_to_file(
        encoded,
        &None::<HashMap<String, String>>,
        &dir.join("adapter_model.safetensors"),
    )
    .unwrap();
}

#[test]
fn runtime_lora_loader_materializes_projection_metadata() {
    let tmp = TempDir::new().unwrap();
    let config = LoraAdapterConfig {
        r: 2,
        lora_alpha: 6,
        target_modules: vec!["qkv_proj".to_string()],
        base_model_name_or_path: "qwen3:0.6b".to_string(),
    };
    write_adapter(
        tmp.path(),
        &config,
        vec![
            (
                "base_model.model.layers.7.self_attn.qkv_proj.lora_A.weight",
                vec![2, 3],
                vec![0.1; 6],
            ),
            (
                "base_model.model.layers.7.self_attn.qkv_proj.lora_B.weight",
                vec![4, 2],
                vec![0.2; 8],
            ),
        ],
    );

    let adapter = load_runtime_lora_adapter::<CpuBackend>(&ActiveLoraAdapter {
        name: "sql".to_string(),
        path: tmp.path().to_path_buf(),
    })
    .unwrap();

    assert_eq!(adapter.name, "sql");
    assert_eq!(adapter.config.r, 2);
    assert_eq!(adapter.linears.len(), 1);
    let linear = &adapter.linears[0];
    assert_eq!(linear.layer_index, Some(7));
    assert_eq!(linear.target_module, "qkv_proj");
    assert_eq!(linear.in_features, 3);
    assert_eq!(linear.out_features, 4);
    assert_eq!(linear.rank, 2);
    assert_eq!(linear.scaling, 3.0);
}

#[test]
fn runtime_lora_apply_projection_adds_scaled_delta() {
    let tmp = TempDir::new().unwrap();
    let config = LoraAdapterConfig {
        r: 2,
        lora_alpha: 6,
        target_modules: vec!["qkv_proj".to_string()],
        base_model_name_or_path: "qwen3:0.6b".to_string(),
    };
    write_adapter(
        tmp.path(),
        &config,
        vec![
            (
                "base_model.model.layers.7.self_attn.qkv_proj.lora_A.weight",
                vec![2, 3],
                vec![1.0, 0.0, 0.0, 0.0, 1.0, 0.0],
            ),
            (
                "base_model.model.layers.7.self_attn.qkv_proj.lora_B.weight",
                vec![4, 2],
                vec![1.0, 0.0, 0.0, 1.0, 1.0, 1.0, 0.0, 0.0],
            ),
        ],
    );
    let adapter = load_runtime_lora_adapter::<CpuBackend>(&ActiveLoraAdapter {
        name: "sql".to_string(),
        path: tmp.path().to_path_buf(),
    })
    .unwrap();

    let mut ctx = CpuBackend::new_context();
    let input = CpuBackend::from_slice(&[1.0, 2.0, 3.0]);
    let mut out = CpuBackend::from_slice(&[10.0, 20.0, 30.0, 40.0]);
    adapter
        .apply_projection(&mut ctx, 7, "qkv_proj", &input, &mut out, 1)
        .unwrap();

    assert_eq!(CpuBackend::to_vec(&out, 4), vec![13.0, 26.0, 39.0, 40.0]);
}