Source code for multistorageclient.contrib.torch.core

 1# SPDX-FileCopyrightText: Copyright (c) 2025 NVIDIA CORPORATION & AFFILIATES. All rights reserved.
 2# SPDX-License-Identifier: Apache-2.0
 3#
 4# Licensed under the Apache License, Version 2.0 (the "License");
 5# you may not use this file except in compliance with the License.
 6# You may obtain a copy of the License at
 7#
 8# http://www.apache.org/licenses/LICENSE-2.0
 9#
10# Unless required by applicable law or agreed to in writing, software
11# distributed under the License is distributed on an "AS IS" BASIS,
12# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
13# See the License for the specific language governing permissions and
14# limitations under the License.
15
16import os
17from typing import IO, Any, Union
18
19import torch as _torch
20
21from ...pathlib import MultiStoragePath
22from ...shortcuts import open as msc_open
23
24

[docs]
25def load(f: Union[str, os.PathLike[str], IO[bytes]], *args: Any, **kwargs: Any) -> Any:
26    """
27    Adapt ``torch.load``.
28    """
29    if isinstance(f, str):
30        with msc_open(f, "rb", prefetch_file=True) as fp:
31            return _torch.load(fp, *args, **kwargs)
32    elif isinstance(f, MultiStoragePath):
33        with f.open("rb") as fp:
34            return _torch.load(fp, *args, **kwargs)
35    else:
36        return _torch.load(f, *args, **kwargs)

37
38

[docs]
39def save(obj: object, f: Union[str, os.PathLike[str], IO[bytes]], *args: Any, **kwargs: Any) -> Any:
40    """
41    Adapt ``torch.save``.
42    """
43    attributes_dict = {}
44    if "attributes" in kwargs:
45        attributes_dict = kwargs.pop("attributes")
46    if isinstance(f, str):
47        with msc_open(f, "wb", attributes=attributes_dict) as fp:
48            return _torch.save(obj, fp, *args, **kwargs)
49    elif isinstance(f, MultiStoragePath):
50        with f.open("wb", attributes=attributes_dict) as fp:
51            return _torch.save(obj, fp, *args, **kwargs)
52    else:
53        return _torch.save(obj, f, *args, **kwargs)