pytorch

1
# mypy: allow-untyped-defs
2
import io
3

4
import torch
5
from torch.package import Importer, OrderedImporter, PackageImporter, sys_importer
6
from torch.package._package_pickler import create_pickler
7
from torch.package._package_unpickler import PackageUnpickler
8
from torch.serialization import _maybe_decode_ascii
9

10

11
def _save_storages(importer, obj):
12
    serialized_storages = []
13
    serialized_dtypes = []
14

15
    importer = importer if isinstance(importer, torch.package.PackageImporter) else None
16
    importers: Importer
17
    if importer is not None:
18
        importers = OrderedImporter(importer, sys_importer)
19
    else:
20
        importers = sys_importer
21

22
    def persistent_id(obj):
23
        if torch.is_storage(obj) or isinstance(obj, torch.storage.TypedStorage):
24
            if isinstance(obj, torch.storage.TypedStorage):
25
                # TODO: Once we decide to break serialization FC, we can
26
                # remove this case
27
                dtype = obj.dtype
28
            else:
29
                dtype = torch.uint8
30

31
            serialized_storages.append(obj)
32
            serialized_dtypes.append(dtype)
33
            return ("storage", len(serialized_storages) - 1)
34

35
        if hasattr(obj, "__reduce_deploy__"):
36
            if _serialized_reduces.get(id(obj)) is None:
37
                _serialized_reduces[id(obj)] = (
38
                    "reduce_deploy",
39
                    id(obj),
40
                    *obj.__reduce_deploy__(importers),
41
                )
42
            return _serialized_reduces[id(obj)]
43

44
        return None
45

46
    # Write the pickle data for `obj`
47
    data_buf = io.BytesIO()
48
    pickler = create_pickler(data_buf, importers)
49
    pickler.persistent_id = persistent_id
50
    pickler.dump(obj)
51
    data_value = data_buf.getvalue()
52
    return (
53
        data_value,
54
        serialized_storages,
55
        serialized_dtypes,
56
        importer.zip_reader if importer else None,
57
    )
58

59

60
def _load_storages(id, zip_reader, obj_bytes, serialized_storages, serialized_dtypes):
61
    def persistent_load(saved_id):
62
        assert isinstance(saved_id, tuple)
63
        typename = _maybe_decode_ascii(saved_id[0])
64
        data = saved_id[1:]
65

66
        if typename == "storage":
67
            # TODO: Once we decide to break serialization FC, we can
68
            # stop wrapping with TypedStorage
69
            storage = serialized_storages[data[0]]
70
            dtype = serialized_dtypes[data[0]]
71
            return torch.storage.TypedStorage(
72
                wrap_storage=storage.untyped(), dtype=dtype
73
            )
74

75
        if typename == "reduce_deploy":
76
            reduce_id, func, args = data
77
            if reduce_id not in _loaded_reduces:
78
                _loaded_reduces[reduce_id] = func(_raw_packages[zip_reader], *args)
79
            return _loaded_reduces[reduce_id]
80

81
        return None
82

83
    importer: Importer
84
    if zip_reader is not None:
85
        importer = OrderedImporter(_get_package(zip_reader), sys_importer)
86
    else:
87
        importer = sys_importer
88

89
    unpickler = PackageUnpickler(importer, io.BytesIO(obj_bytes))
90
    unpickler.persistent_load = persistent_load  # type: ignore[method-assign]
91
    result = _deploy_objects[id] = unpickler.load()
92
    return result
93

94

95
def _get_package(zip_reader):
96
    if zip_reader not in _raw_packages:
97
        _raw_packages[zip_reader] = PackageImporter(zip_reader)
98
    return _raw_packages[zip_reader]
99

100

101
_raw_packages: dict = {}
102
_deploy_objects: dict = {}
103
_serialized_reduces: dict = {}
104
_loaded_reduces: dict = {}
105