ztensor 1.2.3

Unified, zero-copy, and safe I/O for deep learning formats
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

use std::collections::BTreeMap;
use std::io::Write;
use tempfile::NamedTempFile;

pub struct PtTensorSpec {
    pub name: String,
    pub storage_type: String,
    pub storage_key: String,
    pub shape: Vec<usize>,
    pub stride: Vec<usize>,
    pub storage_offset: usize,
    pub numel: usize,
}

pub fn compute_strides(shape: &[usize]) -> Vec<usize> {
    if shape.is_empty() {
        return vec![];
    }
    let mut strides = vec![1usize; shape.len()];
    for i in (0..shape.len().saturating_sub(1)).rev() {
        strides[i] = strides[i + 1] * shape[i + 1];
    }
    strides
}

fn push_int(p: &mut Vec<u8>, val: usize) {
    if val <= 255 {
        p.push(0x4b); // BININT1
        p.push(val as u8);
    } else if val <= 65535 {
        p.push(0x4d); // BININT2
        p.extend_from_slice(&(val as u16).to_le_bytes());
    } else {
        p.push(0x4a); // BININT
        p.extend_from_slice(&(val as i32).to_le_bytes());
    }
}

fn push_int_tuple(p: &mut Vec<u8>, vals: &[usize]) {
    match vals.len() {
        0 => p.push(0x29), // EMPTY_TUPLE
        1 => {
            push_int(p, vals[0]);
            p.push(0x85); // TUPLE1
        }
        2 => {
            push_int(p, vals[0]);
            push_int(p, vals[1]);
            p.push(0x86); // TUPLE2
        }
        3 => {
            push_int(p, vals[0]);
            push_int(p, vals[1]);
            push_int(p, vals[2]);
            p.push(0x87); // TUPLE3
        }
        _ => {
            p.push(0x28); // MARK
            for &v in vals {
                push_int(p, v);
            }
            p.push(0x74); // TUPLE
        }
    }
}

fn push_global(p: &mut Vec<u8>, module: &str, name: &str) {
    p.push(0x63); // GLOBAL
    p.extend_from_slice(module.as_bytes());
    p.push(b'\n');
    p.extend_from_slice(name.as_bytes());
    p.push(b'\n');
}

fn push_short_binunicode(p: &mut Vec<u8>, s: &str) {
    assert!(s.len() <= 255);
    p.push(0x8c); // SHORT_BINUNICODE
    p.push(s.len() as u8);
    p.extend_from_slice(s.as_bytes());
}

fn build_pickle_state_dict(specs: &[PtTensorSpec]) -> Vec<u8> {
    let mut p = Vec::new();

    // PROTO 2
    p.push(0x80);
    p.push(2);

    // EMPTY_DICT
    p.push(0x7d);

    // MARK for SETITEMS
    p.push(0x28);

    for spec in specs {
        // Key: tensor name
        push_short_binunicode(&mut p, &spec.name);

        // Value: _rebuild_tensor_v2(storage, offset, shape, stride, False, OrderedDict())
        push_global(&mut p, "torch._utils", "_rebuild_tensor_v2");

        // MARK for args tuple
        p.push(0x28);

        // arg0: storage via BINPERSID
        p.push(0x28); // MARK for persistent_id tuple
        push_short_binunicode(&mut p, "storage");
        push_global(&mut p, "torch", &spec.storage_type);
        push_short_binunicode(&mut p, &spec.storage_key);
        push_short_binunicode(&mut p, "cpu");
        push_int(&mut p, spec.numel);
        p.push(0x74); // TUPLE
        p.push(0x51); // BINPERSID

        // arg1: storage_offset
        push_int(&mut p, spec.storage_offset);

        // arg2: shape tuple
        push_int_tuple(&mut p, &spec.shape);

        // arg3: stride tuple
        push_int_tuple(&mut p, &spec.stride);

        // arg4: requires_grad = False
        p.push(0x89); // NEWFALSE

        // arg5: OrderedDict()
        push_global(&mut p, "collections", "OrderedDict");
        p.push(0x29); // EMPTY_TUPLE
        p.push(0x52); // REDUCE

        p.push(0x74); // TUPLE (closes args)
        p.push(0x52); // REDUCE
    }

    // SETITEMS
    p.push(0x75);

    // STOP
    p.push(0x2e);

    p
}

pub fn build_pytorch_zip(
    specs: &[PtTensorSpec],
    storage_data: &BTreeMap<String, Vec<u8>>,
) -> NamedTempFile {
    let mut file = NamedTempFile::new().unwrap();
    {
        let mut zip = zip::ZipWriter::new(&mut file);
        let options = zip::write::SimpleFileOptions::default()
            .compression_method(zip::CompressionMethod::Stored);

        let pickle = build_pickle_state_dict(specs);
        zip.start_file("archive/data.pkl", options).unwrap();
        zip.write_all(&pickle).unwrap();

        for (key, data) in storage_data {
            zip.start_file(format!("archive/data/{}", key), options)
                .unwrap();
            zip.write_all(data).unwrap();
        }

        zip.finish().unwrap();
    }
    file
}