torch-atomic-save 0.1.0__py3-none-any.whl
This diff represents the content of publicly available package versions that have been released to one of the supported registries. The information contained in this diff is provided for informational purposes only and reflects changes between package versions as they appear in their respective public registries.
- torch_atomic_save/__init__.py +3 -0
- torch_atomic_save/manager.py +172 -0
- torch_atomic_save-0.1.0.dist-info/METADATA +52 -0
- torch_atomic_save-0.1.0.dist-info/RECORD +7 -0
- torch_atomic_save-0.1.0.dist-info/WHEEL +5 -0
- torch_atomic_save-0.1.0.dist-info/licenses/LICENSE +21 -0
- torch_atomic_save-0.1.0.dist-info/top_level.txt +1 -0
|
@@ -0,0 +1,172 @@
|
|
|
1
|
+
import os
|
|
2
|
+
import shutil
|
|
3
|
+
import tempfile
|
|
4
|
+
import threading
|
|
5
|
+
import concurrent.futures
|
|
6
|
+
import logging
|
|
7
|
+
import signal
|
|
8
|
+
import torch
|
|
9
|
+
from typing import Any, Optional, Dict, Tuple
|
|
10
|
+
|
|
11
|
+
logger = logging.getLogger(__name__)
|
|
12
|
+
|
|
13
|
+
|
|
14
|
+
class SlurmAtomicManager:
|
|
15
|
+
"""
|
|
16
|
+
Manages asynchronous, atomic checkpoint saving across different filesystems.
|
|
17
|
+
Optimized for Slurm/Lustre environments.
|
|
18
|
+
"""
|
|
19
|
+
|
|
20
|
+
def __init__(self, max_workers: int = 4):
|
|
21
|
+
self._executor = concurrent.futures.ThreadPoolExecutor(
|
|
22
|
+
max_workers=max_workers,
|
|
23
|
+
thread_name_prefix="SlurmSaveWorker"
|
|
24
|
+
)
|
|
25
|
+
# path -> (future, cancellation_event)
|
|
26
|
+
self._registry: Dict[str, Tuple[concurrent.futures.Future, threading.Event]] = {}
|
|
27
|
+
self._lock = threading.Lock()
|
|
28
|
+
self._shutdown_event = threading.Event()
|
|
29
|
+
|
|
30
|
+
def __enter__(self):
|
|
31
|
+
return self
|
|
32
|
+
|
|
33
|
+
def __exit__(self, exc_type, exc_val, exc_tb):
|
|
34
|
+
self.shutdown()
|
|
35
|
+
|
|
36
|
+
def wait_for_all(self, timeout: Optional[float] = None):
|
|
37
|
+
"""Blocks until all currently pending saves are complete."""
|
|
38
|
+
with self._lock:
|
|
39
|
+
futures = [f for f, _ in self._registry.values()]
|
|
40
|
+
concurrent.futures.wait(futures, timeout=timeout)
|
|
41
|
+
|
|
42
|
+
def save(self, model: torch.nn.Module, path: str, tmp_dir: Optional[str] = None, half_prec: bool = False):
|
|
43
|
+
"""
|
|
44
|
+
Main entry point for saving models.
|
|
45
|
+
Moves tensors to CPU and clones them to prevent race conditions.
|
|
46
|
+
"""
|
|
47
|
+
path = os.path.abspath(path)
|
|
48
|
+
|
|
49
|
+
# Snapshot weights: Move to CPU and Clone
|
|
50
|
+
if half_prec:
|
|
51
|
+
state_dict = {k: v.half().cpu().clone() for k, v in model.state_dict().items()}
|
|
52
|
+
else:
|
|
53
|
+
state_dict = {k: v.cpu().clone() for k, v in model.state_dict().items()}
|
|
54
|
+
|
|
55
|
+
if tmp_dir is None:
|
|
56
|
+
# Synchronous fallback
|
|
57
|
+
os.makedirs(os.path.dirname(path), exist_ok=True)
|
|
58
|
+
torch.save(state_dict, path)
|
|
59
|
+
else:
|
|
60
|
+
self._atomic_save(state_dict, path, tmp_dir)
|
|
61
|
+
|
|
62
|
+
def _atomic_save(self, obj: Any, path: str, tmp_dir: str) -> None:
|
|
63
|
+
"""
|
|
64
|
+
Saves an object to a staging area, then schedules an atomic cross-FS move.
|
|
65
|
+
"""
|
|
66
|
+
if self._shutdown_event.is_set():
|
|
67
|
+
logger.warning(f"Manager is shutting down. Ignoring save request for {path}")
|
|
68
|
+
return
|
|
69
|
+
|
|
70
|
+
path = os.path.abspath(path)
|
|
71
|
+
os.makedirs(os.path.dirname(path), exist_ok=True)
|
|
72
|
+
|
|
73
|
+
if not os.path.isdir(tmp_dir):
|
|
74
|
+
raise FileNotFoundError(f"Staging directory missing: {tmp_dir}")
|
|
75
|
+
|
|
76
|
+
# 1. Setup staging dir
|
|
77
|
+
fd, tmp_src = tempfile.mkstemp(dir=tmp_dir, suffix=".pt.staging")
|
|
78
|
+
os.close(fd)
|
|
79
|
+
|
|
80
|
+
# 2. Registry Update & Cancellation
|
|
81
|
+
cancelled = threading.Event()
|
|
82
|
+
with self._lock:
|
|
83
|
+
if path in self._registry:
|
|
84
|
+
prev_f, prev_ev = self._registry[path]
|
|
85
|
+
prev_ev.set() # Signal background thread to skip the move
|
|
86
|
+
prev_f.cancel()
|
|
87
|
+
|
|
88
|
+
future = self._executor.submit(
|
|
89
|
+
self._save_and_copy, obj, tmp_src, path, cancelled
|
|
90
|
+
)
|
|
91
|
+
self._registry[path] = (future, cancelled)
|
|
92
|
+
future.add_done_callback(lambda f: self._cleanup_registry(path, f))
|
|
93
|
+
|
|
94
|
+
def _cleanup_registry(self, path: str, future: concurrent.futures.Future) -> None:
|
|
95
|
+
"""Removes the future from registry and logs errors."""
|
|
96
|
+
with self._lock:
|
|
97
|
+
current = self._registry.get(path)
|
|
98
|
+
if current and current[0] is future:
|
|
99
|
+
self._registry.pop(path, None)
|
|
100
|
+
|
|
101
|
+
if not future.cancelled():
|
|
102
|
+
exc = future.exception()
|
|
103
|
+
if exc:
|
|
104
|
+
logger.error(f"Async save to {path} failed: {exc}", exc_info=exc)
|
|
105
|
+
|
|
106
|
+
def _save_and_copy(self, obj, tmp_src, dst, cancelled):
|
|
107
|
+
"""Handles both the initial write and the atomic move."""
|
|
108
|
+
try:
|
|
109
|
+
if not cancelled.is_set():
|
|
110
|
+
torch.save(obj, tmp_src)
|
|
111
|
+
|
|
112
|
+
self._atomic_copy_and_cleanup(tmp_src, dst, cancelled)
|
|
113
|
+
except Exception as e:
|
|
114
|
+
# Cleanup tmp_src if torch.save failed
|
|
115
|
+
if os.path.exists(tmp_src):
|
|
116
|
+
os.unlink(tmp_src)
|
|
117
|
+
raise
|
|
118
|
+
|
|
119
|
+
def _atomic_copy_and_cleanup(self, src: str, dst: str, cancelled: threading.Event, max_retries: int = 5) -> None:
|
|
120
|
+
"""The background worker logic."""
|
|
121
|
+
dst_dir = os.path.dirname(os.path.abspath(dst))
|
|
122
|
+
try:
|
|
123
|
+
# Check if same device (ST_DEV)
|
|
124
|
+
if _same_filesystem(src, dst):
|
|
125
|
+
if not cancelled.is_set():
|
|
126
|
+
os.replace(src, dst)
|
|
127
|
+
else:
|
|
128
|
+
# Cross-FS: Create sibling tmp in destination's parent
|
|
129
|
+
fd, sibling_tmp = tempfile.mkstemp(dir=dst_dir, suffix=".atomic_tmp")
|
|
130
|
+
os.close(fd)
|
|
131
|
+
try:
|
|
132
|
+
shutil.copy2(src, sibling_tmp)
|
|
133
|
+
if not cancelled.is_set():
|
|
134
|
+
os.replace(sibling_tmp, dst)
|
|
135
|
+
else:
|
|
136
|
+
os.unlink(sibling_tmp)
|
|
137
|
+
except Exception:
|
|
138
|
+
try:
|
|
139
|
+
os.unlink(sibling_tmp)
|
|
140
|
+
except OSError:
|
|
141
|
+
pass
|
|
142
|
+
raise
|
|
143
|
+
finally:
|
|
144
|
+
try:
|
|
145
|
+
os.unlink(src)
|
|
146
|
+
except OSError:
|
|
147
|
+
pass
|
|
148
|
+
|
|
149
|
+
def shutdown(self, wait: bool = True):
|
|
150
|
+
"""Cleanly shut down the executor."""
|
|
151
|
+
self._shutdown_event.set()
|
|
152
|
+
self._executor.shutdown(wait=wait)
|
|
153
|
+
|
|
154
|
+
|
|
155
|
+
def install_slurm_handler(manager: SlurmAtomicManager):
|
|
156
|
+
"""
|
|
157
|
+
Installs a handler for SIGTERM (Slurm preemption) to ensure
|
|
158
|
+
the manager flushes pending IO before the process exits.
|
|
159
|
+
"""
|
|
160
|
+
|
|
161
|
+
def handle_sigterm(signum, frame):
|
|
162
|
+
logger.info("Received SIGTERM/Preemption signal. Flushing IO...")
|
|
163
|
+
manager.shutdown(wait=True)
|
|
164
|
+
# Optional: Re-raise or exit
|
|
165
|
+
# os._exit(0)
|
|
166
|
+
|
|
167
|
+
signal.signal(signal.SIGTERM, handle_sigterm)
|
|
168
|
+
|
|
169
|
+
|
|
170
|
+
def _same_filesystem(path_a: str, path_b: str) -> bool:
|
|
171
|
+
stat_b_target = path_b if os.path.exists(path_b) else os.path.dirname(os.path.abspath(path_b))
|
|
172
|
+
return os.stat(path_a).st_dev == os.stat(stat_b_target).st_dev
|
|
@@ -0,0 +1,52 @@
|
|
|
1
|
+
Metadata-Version: 2.4
|
|
2
|
+
Name: torch_atomic_save
|
|
3
|
+
Version: 0.1.0
|
|
4
|
+
Summary: Atomic, asynchronous checkpoint saving for Pytorch for Slurm/Lustre environments
|
|
5
|
+
Author-email: Patrick Carnahan <pcarnah@uwo.ca>
|
|
6
|
+
License: MIT
|
|
7
|
+
Requires-Python: >=3.9
|
|
8
|
+
Description-Content-Type: text/markdown
|
|
9
|
+
License-File: LICENSE
|
|
10
|
+
Requires-Dist: torch>=2.0.0
|
|
11
|
+
Requires-Dist: numpy
|
|
12
|
+
Provides-Extra: dev
|
|
13
|
+
Requires-Dist: pytest; extra == "dev"
|
|
14
|
+
Requires-Dist: pytest-cov; extra == "dev"
|
|
15
|
+
Dynamic: license-file
|
|
16
|
+
|
|
17
|
+
# torch-atomic-save
|
|
18
|
+
|
|
19
|
+
[](https://github.com/pcarnah/torch-atomic-save/actions)
|
|
20
|
+
[](https://opensource.org/licenses/MIT)
|
|
21
|
+
|
|
22
|
+
An asynchronous, atomic checkpointing utility for PyTorch, optimized for Slurm and Lustre/NFS environments.
|
|
23
|
+
|
|
24
|
+
## Key Features
|
|
25
|
+
* **Atomic Moves:** Prevents corrupted checkpoints during Slurm preemption.
|
|
26
|
+
* **Non-Blocking:** Offloads I/O to a background thread pool.
|
|
27
|
+
* **Race Condition Protection:** Automatically clones tensors to CPU before background saving.
|
|
28
|
+
* **Cross-FS Support:** Handles moves between local SSD and network storage safely.
|
|
29
|
+
|
|
30
|
+
## Installation
|
|
31
|
+
```bash
|
|
32
|
+
pip install torch-atomic-save
|
|
33
|
+
```
|
|
34
|
+
|
|
35
|
+
|
|
36
|
+
## Usage
|
|
37
|
+
```python
|
|
38
|
+
from torch_atomic_save import SlurmAtomicManager
|
|
39
|
+
|
|
40
|
+
# Initialize the manager
|
|
41
|
+
manager = SlurmAtomicManager(max_workers=4)
|
|
42
|
+
|
|
43
|
+
# In your training loop:
|
|
44
|
+
if epoch % save_interval == 0:
|
|
45
|
+
manager.save(model, "path/to/checkpoints/model.pt", tmp_dir="/scratch/user/tmp")
|
|
46
|
+
|
|
47
|
+
# Ensure all I/O is finished before exiting
|
|
48
|
+
manager.wait_for_all()
|
|
49
|
+
```
|
|
50
|
+
|
|
51
|
+
## Attribution
|
|
52
|
+
If you use this software in your research, please cite it using the "Cite this repository" button or the provided CITATION.cff.
|
|
@@ -0,0 +1,7 @@
|
|
|
1
|
+
torch_atomic_save/__init__.py,sha256=NhsLuYY4EffBDJgeShgLEsNrEnl7rXiVlDqBlHN_S8g,123
|
|
2
|
+
torch_atomic_save/manager.py,sha256=XXpNCHpRP8hmxcHIYvHMQkzTnC15wPrKhPl2K9GUxbE,6383
|
|
3
|
+
torch_atomic_save-0.1.0.dist-info/licenses/LICENSE,sha256=WvXpsuVMCjnZIRoaaalqisATp86XcY0_PQSxx7HRkkA,1092
|
|
4
|
+
torch_atomic_save-0.1.0.dist-info/METADATA,sha256=BVcChI39xwIsEdsBsZ1rnlDEdAoPCHe1PVSbAHccZzA,1799
|
|
5
|
+
torch_atomic_save-0.1.0.dist-info/WHEEL,sha256=aeYiig01lYGDzBgS8HxWXOg3uV61G9ijOsup-k9o1sk,91
|
|
6
|
+
torch_atomic_save-0.1.0.dist-info/top_level.txt,sha256=h0Omk6Y6-mjulu1PfoOZ7TblXy4rtRT6KRgd5s8gpTA,18
|
|
7
|
+
torch_atomic_save-0.1.0.dist-info/RECORD,,
|
|
@@ -0,0 +1,21 @@
|
|
|
1
|
+
MIT License
|
|
2
|
+
|
|
3
|
+
Copyright (c) 2026 Patrick Carnahan
|
|
4
|
+
|
|
5
|
+
Permission is hereby granted, free of charge, to any person obtaining a copy
|
|
6
|
+
of this software and associated documentation files (the "Software"), to deal
|
|
7
|
+
in the Software without restriction, including without limitation the rights
|
|
8
|
+
to use, copy, modify, merge, publish, distribute, sublicense, and/or sell
|
|
9
|
+
copies of the Software, and to permit persons to whom the Software is
|
|
10
|
+
furnished to do so, subject to the following conditions:
|
|
11
|
+
|
|
12
|
+
The above copyright notice and this permission notice shall be included in all
|
|
13
|
+
copies or substantial portions of the Software.
|
|
14
|
+
|
|
15
|
+
THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR
|
|
16
|
+
IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY,
|
|
17
|
+
FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT. IN NO EVENT SHALL THE
|
|
18
|
+
AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER
|
|
19
|
+
LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM,
|
|
20
|
+
OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN THE
|
|
21
|
+
SOFTWARE.
|
|
@@ -0,0 +1 @@
|
|
|
1
|
+
torch_atomic_save
|