aprender-core 0.29.2

Next-generation machine learning library in pure Rust
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

#![allow(clippy::disallowed_methods)]
//! Generate test APR v2 file for CLI testing
use aprender::format::v2::{AprV2Metadata, AprV2Writer, TensorDType};

fn main() {
    let mut metadata = AprV2Metadata::new("whisper");
    metadata.name = Some("whisper-tiny-test".to_string());
    metadata.description = Some("Test Whisper model for APR v2 serving".to_string());

    let mut writer = AprV2Writer::new(metadata);

    // Helper to create f32 tensor bytes
    fn f32_bytes(vals: &[f32]) -> Vec<u8> {
        vals.iter().flat_map(|f| f.to_le_bytes()).collect()
    }

    // Encoder layers
    for i in 0..2 {
        let prefix = format!("encoder.layers.{}", i);
        writer.add_tensor(
            &format!("{}.self_attn.q_proj.weight", prefix),
            TensorDType::F32,
            vec![384, 384],
            f32_bytes(&vec![0.01; 384 * 384]),
        );
        writer.add_tensor(
            &format!("{}.self_attn.k_proj.weight", prefix),
            TensorDType::F32,
            vec![384, 384],
            f32_bytes(&vec![0.01; 384 * 384]),
        );
        writer.add_tensor(
            &format!("{}.self_attn.v_proj.weight", prefix),
            TensorDType::F32,
            vec![384, 384],
            f32_bytes(&vec![0.01; 384 * 384]),
        );
        writer.add_tensor(
            &format!("{}.self_attn.out_proj.weight", prefix),
            TensorDType::F32,
            vec![384, 384],
            f32_bytes(&vec![0.01; 384 * 384]),
        );
        writer.add_tensor(
            &format!("{}.mlp.fc1.weight", prefix),
            TensorDType::F32,
            vec![1536, 384],
            f32_bytes(&vec![0.01; 1536 * 384]),
        );
        writer.add_tensor(
            &format!("{}.mlp.fc2.weight", prefix),
            TensorDType::F32,
            vec![384, 1536],
            f32_bytes(&vec![0.01; 384 * 1536]),
        );
    }

    // Decoder layers
    for i in 0..2 {
        let prefix = format!("decoder.layers.{}", i);
        // Self attention
        writer.add_tensor(
            &format!("{}.self_attn.q_proj.weight", prefix),
            TensorDType::F32,
            vec![384, 384],
            f32_bytes(&vec![0.02; 384 * 384]),
        );
        writer.add_tensor(
            &format!("{}.self_attn.k_proj.weight", prefix),
            TensorDType::F32,
            vec![384, 384],
            f32_bytes(&vec![0.02; 384 * 384]),
        );
        writer.add_tensor(
            &format!("{}.self_attn.v_proj.weight", prefix),
            TensorDType::F32,
            vec![384, 384],
            f32_bytes(&vec![0.02; 384 * 384]),
        );
        writer.add_tensor(
            &format!("{}.self_attn.out_proj.weight", prefix),
            TensorDType::F32,
            vec![384, 384],
            f32_bytes(&vec![0.02; 384 * 384]),
        );
        // Cross attention
        writer.add_tensor(
            &format!("{}.cross_attn.q_proj.weight", prefix),
            TensorDType::F32,
            vec![384, 384],
            f32_bytes(&vec![0.03; 384 * 384]),
        );
        writer.add_tensor(
            &format!("{}.cross_attn.k_proj.weight", prefix),
            TensorDType::F32,
            vec![384, 384],
            f32_bytes(&vec![0.03; 384 * 384]),
        );
        writer.add_tensor(
            &format!("{}.cross_attn.v_proj.weight", prefix),
            TensorDType::F32,
            vec![384, 384],
            f32_bytes(&vec![0.03; 384 * 384]),
        );
        writer.add_tensor(
            &format!("{}.cross_attn.out_proj.weight", prefix),
            TensorDType::F32,
            vec![384, 384],
            f32_bytes(&vec![0.03; 384 * 384]),
        );
        // FFN
        writer.add_tensor(
            &format!("{}.mlp.fc1.weight", prefix),
            TensorDType::F32,
            vec![1536, 384],
            f32_bytes(&vec![0.02; 1536 * 384]),
        );
        writer.add_tensor(
            &format!("{}.mlp.fc2.weight", prefix),
            TensorDType::F32,
            vec![384, 1536],
            f32_bytes(&vec![0.02; 384 * 1536]),
        );
    }

    let apr_bytes = writer.write().expect("Failed to write APR");
    std::fs::write("/tmp/test-whisper-v2.apr", &apr_bytes).expect("Failed to save");
    println!(
        "Created /tmp/test-whisper-v2.apr ({} bytes)",
        apr_bytes.len()
    );
}