ferrum-models 0.7.6

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
129
130
131
132
133
134
135

use ferrum_models::{
    load_startup_lora_adapter, load_startup_lora_adapters, LoraAdapterConfig, StartupLoraSpec,
};
use safetensors::tensor::{serialize_to_file, Dtype, TensorView};
use std::collections::HashMap;
use std::path::Path;
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: &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();
}

fn valid_config() -> LoraAdapterConfig {
    LoraAdapterConfig {
        r: 2,
        lora_alpha: 4,
        target_modules: vec!["linear".to_string()],
        base_model_name_or_path: "qwen3:0.6b".to_string(),
    }
}

#[test]
fn lora_loader_reads_config_and_safetensors() {
    let tmp = TempDir::new().unwrap();
    write_adapter(
        tmp.path(),
        &valid_config(),
        vec![
            (
                "base_model.model.layers.0.linear.lora_A.weight",
                vec![2, 3],
                vec![0.0; 6],
            ),
            (
                "base_model.model.layers.0.linear.lora_B.weight",
                vec![4, 2],
                vec![0.0; 8],
            ),
        ],
    );

    let adapter = load_startup_lora_adapter("sql", tmp.path(), "qwen3:0.6b:sql").unwrap();
    assert_eq!(adapter.name, "sql");
    assert_eq!(adapter.public_model_id, "qwen3:0.6b:sql");
    assert_eq!(adapter.config.r, 2);
    assert_eq!(adapter.tensors.len(), 1);
    assert_eq!(adapter.tensors[0].in_features, 3);
    assert_eq!(adapter.tensors[0].out_features, 4);
}

#[test]
fn lora_loader_shape_mismatch_fails() {
    let tmp = TempDir::new().unwrap();
    write_adapter(
        tmp.path(),
        &valid_config(),
        vec![
            ("linear.lora_A.weight", vec![1, 3], vec![0.0; 3]),
            ("linear.lora_B.weight", vec![4, 2], vec![0.0; 8]),
        ],
    );

    let err = load_startup_lora_adapter("bad", tmp.path(), "base:bad").unwrap_err();
    assert!(err.to_string().contains("LoRA rank mismatch"));
}

#[test]
fn lora_loader_unknown_target_fails() {
    let tmp = TempDir::new().unwrap();
    let mut config = valid_config();
    config.target_modules = vec!["not_a_projection".to_string()];
    write_adapter(tmp.path(), &config, vec![]);

    let err = load_startup_lora_adapter("bad", tmp.path(), "base:bad").unwrap_err();
    assert!(err.to_string().contains("unsupported LoRA target module"));
}

#[test]
fn lora_loader_duplicate_names_fail() {
    let tmp = TempDir::new().unwrap();
    write_adapter(
        tmp.path(),
        &valid_config(),
        vec![
            ("linear.lora_A.weight", vec![2, 3], vec![0.0; 6]),
            ("linear.lora_B.weight", vec![4, 2], vec![0.0; 8]),
        ],
    );
    let specs = vec![
        StartupLoraSpec {
            name: "sql".to_string(),
            path: tmp.path().to_path_buf(),
        },
        StartupLoraSpec {
            name: "sql".to_string(),
            path: tmp.path().to_path_buf(),
        },
    ];

    let err = load_startup_lora_adapters("qwen3:0.6b", None, &specs).unwrap_err();
    assert!(err.to_string().contains("duplicate LoRA adapter name"));
}