[torchelastic] ensure grandchild processes are restarted correctly #113231
Conversation
This file contains hidden or bidirectional Unicode text that may be interpreted or compiled differently than what appears below. To review, open the file in an editor that reveals hidden Unicode characters.
Learn more about bidirectional Unicode characters
When torchelastic notices that one rank has failed, it will sent a SIGTERM signal to other trainer ranks to tear them down before restarting. However, if the trainer itself launches subprocesses, or is launched by a non-python wrapper script, then the SIGTERM will be delived only to the direct child of torch eleastic and not all descendants. This opens subprocesses in a new linux 'session' which starts a new process group with the pgid the same as the trainers pid. Then when we send signals, we deliver them to the process group rather than just the direct child. [ghstack-poisoned]
🔗 Helpful Links🧪 See artifacts and rendered test results at hud.pytorch.org/pr/113231
Note: Links to docs will display an error until the docs builds have been completed. ✅ No FailuresAs of commit f5c3318 with merge base dbb96ef ( This comment was automatically generated by Dr. CI and updates every 15 minutes. |
zdevito
added a commit
that referenced
this pull request
Nov 8, 2023
When torchelastic notices that one rank has failed, it will sent a SIGTERM signal to other trainer ranks to tear them down before restarting. However, if the trainer itself launches subprocesses, or is launched by a non-python wrapper script, then the SIGTERM will be delived only to the direct child of torch eleastic and not all descendants. This opens subprocesses in a new linux 'session' which starts a new process group with the pgid the same as the trainers pid. Then when we send signals, we deliver them to the process group rather than just the direct child. ghstack-source-id: 5672a1d Pull Request resolved: #113231
H-Huang
reviewed
Nov 9, 2023
| def _popen(self, args: Tuple, env: Dict[str, str]) -> subprocess.Popen: | ||
| kwargs = {} | ||
| if not IS_WINDOWS: | ||
| kwargs['start_new_session'] = True |
There was a problem hiding this comment.
if new processes are started this way, will they be detached from the parent? For example, if I run something like torchrun my_script.py --nnodes=1 --nproc-per-node=2 and then I Ctrl+C, will the child processes still be able to exit?
There was a problem hiding this comment.
The will be in a different session than the parent, but torchrun is a kind of like bash in that it is a process manager, so when it gets a SIGINT, it will propagate that to its children before existing. So when you Ctrl-C torchrun you get:
[2023-11-09 10:53:46,286] torch.distributed.elastic.agent.server.api: [WARNING] Received Signals.SIGINT death signal, shutting down workers
And the workers will stop too.
There was a problem hiding this comment.
Sounds good, thanks for confirming!
H-Huang
approved these changes
Nov 10, 2023
…orrectly" When torchelastic notices that one rank has failed, it will sent a SIGTERM signal to other trainer ranks to tear them down before restarting. However, if the trainer itself launches subprocesses, or is launched by a non-python wrapper script, then the SIGTERM will be delivered only to the direct child of torch eleastic and not all descendants. This opens subprocesses in a new linux 'session' which starts a new process group with the pgid the same as the trainers pid. Then when we send signals, we deliver them to the process group rather than just the direct child. [ghstack-poisoned]
…orrectly" When torchelastic notices that one rank has failed, it will sent a SIGTERM signal to other trainer ranks to tear them down before restarting. However, if the trainer itself launches subprocesses, or is launched by a non-python wrapper script, then the SIGTERM will be delivered only to the direct child of torch eleastic and not all descendants. This opens subprocesses in a new linux 'session' which starts a new process group with the pgid the same as the trainers pid. Then when we send signals, we deliver them to the process group rather than just the direct child. [ghstack-poisoned]
pytorchmergebot
pushed a commit
that referenced
this pull request
Nov 19, 2023
I've run into a couple of cases now where max_split_size_mb has been set in projects as a workaround for fragmentation but it ends up causing problems later, such as degraded performance from freeing empty segments. While it is a useful setting to have, expandable_segments is probably a better first resort for fixing fragmentation since when it works it is less likely to need synchronous GPU operations to continue running. Pull Request resolved: #113481 Approved by: https://github.com/msaroufim, https://github.com/albanD ghstack dependencies: #113231
Sign up for free
to join this conversation on GitHub.
Already have an account?
Sign in to comment
3 participants
Add this suggestion to a batch that can be applied as a single commit.
This suggestion is invalid because no changes were made to the code.
Suggestions cannot be applied while the pull request is closed.
Suggestions cannot be applied while viewing a subset of changes.
Only one suggestion per line can be applied in a batch.
Add this suggestion to a batch that can be applied as a single commit.
Applying suggestions on deleted lines is not supported.
You must change the existing code in this line in order to create a valid suggestion.
Outdated suggestions cannot be applied.
This suggestion has been applied or marked resolved.
Suggestions cannot be applied from pending reviews.
Suggestions cannot be applied on multi-line comments.
Suggestions cannot be applied while the pull request is queued to merge.
Suggestion cannot be applied right now. Please check back later.
Stack from ghstack (oldest at bottom):
When torchelastic notices that one rank has failed, it will sent a SIGTERM
signal to other trainer ranks to tear them down before restarting. However,
if the trainer itself launches subprocesses, or is launched by a non-python
wrapper script, then the SIGTERM will be delivered only to the direct child of
torch eleastic and not all descendants. This opens subprocesses in a new
linux 'session' which starts a new process group with the pgid the same
as the trainers pid. Then when we send signals, we deliver them to the
process group rather than just the direct child.