1# SPDX-FileCopyrightText: NVIDIA CORPORATION & AFFILIATES
2# Copyright (c) 2024 NVIDIA CORPORATION & AFFILIATES. All rights reserved.
3# SPDX-License-Identifier: Apache-2.0
4#
5# Licensed under the Apache License, Version 2.0 (the "License");
6# you may not use this file except in compliance with the License.
7# You may obtain a copy of the License at
8#
9# http://www.apache.org/licenses/LICENSE-2.0
10#
11# Unless required by applicable law or agreed to in writing, software
12# distributed under the License is distributed on an "AS IS" BASIS,
13# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
14# See the License for the specific language governing permissions and
15# limitations under the License.
16
17
18# This example demonstrates how to integrate ``inprocess.Wrapper()`` into an
19# existing PyTorch training codebase.
20#
21# In this case, the entire ``main()`` function is wrapped. While all features
22# of ``inprocess.Wrapper()`` are available and active, the Wrapper is
23# configured to restart the entire application upon any failure. Consequently,
24# the application state is not preserved between restarts and the entire
25# ``main()`` is relaunched, leading to less efficient recovery from failures.
26#
27# NOTE: inprocess.Wrapper is not fully compatible with modern
28# ``torch.distributed.run``, because it automatically terminates all local
29# workers upon any local worker process failure; in this case inprocess.Wrapper
30# can only recover from transient faults that don't terminate any of the
31# training processes
32
33import argparse
34import datetime
35import logging
36import os
37import pathlib
38import random
39import time
40from typing import Optional
41
42os.environ['TORCH_CPP_LOG_LEVEL'] = 'error'
43import torch
44
45import nvidia_resiliency_ext.inprocess as inprocess
46
47raise_timestamp = None
48
49
50def parse_args():
51 parser = argparse.ArgumentParser(
52 description='Inprocess Restart Basic Example',
53 formatter_class=argparse.ArgumentDefaultsHelpFormatter,
54 )
55
56 parser.add_argument(
57 '--size',
58 default=64,
59 type=int,
60 help='model hidden size',
61 )
62 parser.add_argument(
63 '--layers',
64 default=4,
65 type=int,
66 help='number of layers',
67 )
68 parser.add_argument(
69 '--log-interval',
70 default=100,
71 type=int,
72 help='logging interval',
73 )
74 parser.add_argument(
75 '--chkpt-interval',
76 default=100,
77 type=int,
78 help='checkpointing interval',
79 )
80 parser.add_argument(
81 '--total-iterations',
82 default=1000000,
83 type=int,
84 help='total training iterations',
85 )
86 parser.add_argument(
87 '--seed',
88 default=None,
89 type=int,
90 help='random seed, time-based if None',
91 )
92 parser.add_argument(
93 '--path',
94 default='/tmp/',
95 type=str,
96 help='directory for the checkpoint file',
97 )
98 parser.add_argument(
99 '--fault-prob',
100 default=0.001,
101 type=float,
102 help='fault injection probability',
103 )
104 parser.add_argument(
105 '--device',
106 default='cpu',
107 choices=['cpu', 'cuda'],
108 help='device',
109 )
110 parser.add_argument(
111 '--log-level',
112 type=lambda s: logging._nameToLevel[s.upper()],
113 default=logging.INFO,
114 help='logging level',
115 )
116
117 return parser.parse_args()
118
119
120# TCPStore created by the Wrapper uses ``(MASTER_PORT + 1)`` port for the
121# internal Wrapper's TCPStore to avoid conflicts with application's TCPStore
122# listening on ``(MASTER_PORT + 2 + iteration)``, and with TCPStore created by
123# ``torch.distributed.run`` listening on ``MASTER_PORT``.
124#
125# An instance of ``inprocess.CallWrapper` is automatically injected into
126# wrapped function arguments when Wrapper is invoked.
127@inprocess.Wrapper(
128 store_kwargs={'port': int(os.getenv('MASTER_PORT', 29500)) + 1},
129 health_check=inprocess.health_check.CudaHealthCheck(),
130)
131def main(call_wrapper: Optional[inprocess.CallWrapper] = None):
132 global raise_timestamp
133 if raise_timestamp is not None:
134 restart_latency = time.perf_counter() - raise_timestamp
135 logging.info(f'restart latency: {restart_latency:.3f}s')
136 raise_timestamp = None
137
138 args = parse_args()
139 logging.info(f'{args}')
140
141 log_interval = args.log_interval
142 chkpt_interval = args.chkpt_interval
143
144 rank = int(os.environ['RANK'])
145 local_rank = int(os.environ['LOCAL_RANK'])
146 world_size = int(os.environ['WORLD_SIZE'])
147
148 if args.device == 'cuda':
149 torch.cuda.set_device(local_rank)
150 device = torch.device('cuda')
151 backend = 'nccl'
152 timeout = datetime.timedelta(seconds=150)
153 elif args.device == 'cpu':
154 device = torch.device('cpu')
155 backend = 'gloo'
156 timeout = datetime.timedelta(seconds=10)
157 else:
158 raise RuntimeError
159
160 if args.seed is not None:
161 torch.manual_seed(args.seed)
162 model = torch.nn.Sequential(
163 *[torch.nn.Linear(args.size, args.size) for _ in range(args.layers)]
164 ).to(device)
165 opt = torch.optim.Adam(model.parameters(), lr=1e-5)
166
167 # Application's TCPStore uses ``(MASTER_PORT + 2 + iteration)`` to avoid
168 # conflicts with a TCPStore created by ``torch.distributed.run``,
169 # inprocess.Wrapper and application's TCPStores created in previous restart
170 # iterations.
171 store = torch.distributed.TCPStore(
172 host_name=os.environ['MASTER_ADDR'],
173 port=int(os.environ['MASTER_PORT']) + 2 + call_wrapper.iteration,
174 world_size=int(os.environ['WORLD_SIZE']),
175 is_master=int(os.environ['RANK']) == 0,
176 multi_tenant=True,
177 wait_for_workers=True,
178 use_libuv=True,
179 )
180
181 torch.distributed.init_process_group(
182 backend=backend,
183 store=store,
184 rank=int(os.environ['RANK']),
185 world_size=int(os.environ['WORLD_SIZE']),
186 timeout=timeout,
187 )
188 model_ddp = torch.nn.parallel.DistributedDataParallel(model)
189
190 iteration = 0
191 loss = torch.tensor(float('nan'))
192 checkpoint_path = pathlib.Path(args.path) / 'checkpoint.pt'
193
194 # Application loads state from the latest checkpoint on every restart
195 # iteration of the wrapped function.
196 if checkpoint_path.exists():
197 checkpoint = torch.load(checkpoint_path)
198 model.load_state_dict(checkpoint['model'])
199 opt.load_state_dict(checkpoint['opt'])
200 torch.set_rng_state(checkpoint['rng'])
201 iteration = checkpoint['iteration']
202
203 if args.seed is not None:
204 random.seed(args.seed + iteration * world_size + rank)
205 else:
206 random.seed(time.perf_counter_ns())
207
208 for iteration in range(iteration, args.total_iterations):
209
210 # Application periodically saves a checkpoint. The checkpoint allows
211 # the application to continue from previous state after a restart.
212 if iteration % chkpt_interval == chkpt_interval - 1:
213 torch.distributed.barrier()
214 if rank == 0:
215 checkpoint = {
216 'model': model.state_dict(),
217 'opt': opt.state_dict(),
218 'rng': torch.get_rng_state(),
219 'iteration': iteration,
220 }
221 # Saving the checkpoint is performed within atomic() context
222 # manager to ensure that the main thread won't execute
223 # torch.save while a restart procedure is in progress.
224 with call_wrapper.atomic():
225 torch.save(checkpoint, checkpoint_path)
226
227 # Randomly trigger an example fault
228 if random.random() < args.fault_prob:
229 raise_timestamp = time.perf_counter()
230 raise RuntimeError(f'example fault at {iteration=} from {rank=}')
231
232 inp = torch.rand(args.size, args.size).to(device)
233 model.zero_grad()
234 out = model_ddp(inp)
235 loss = out.square().mean()
236 loss.backward()
237 opt.step()
238 loss.item()
239
240 if rank == 0 and iteration % log_interval == log_interval - 1:
241 logging.info(f'{rank=} {iteration=} {loss.item()=}')
242
243
244if __name__ == '__main__':
245 # ``inprocess.Wrapper`` uses logging library to output messages. In this
246 # example the Wrapper is applied to ``main()``, therefore logging needs to
247 # be initialized and configured before the Wrapper is launched.
248 args = parse_args()
249 logging.basicConfig(
250 format="%(asctime)s | %(levelname)s | %(name)s | %(message)s",
251 level=args.log_level,
252 )
253 main()