[WIP] Support multiprocessing training #141
Closed
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
Vikas-kum
suggested changes
Jan 10, 2020
| @pytest.mark.slow | ||
| def test_run_net_distributed_multiproc_save_all_workers(): | ||
| size = 2 | ||
| os.environ["SMDEBUG_NUM_WORKERS"] = "2" |
There was a problem hiding this comment.
monkeypatch, see here -
Vikas-kum
suggested changes
Jan 17, 2020
Comment on lines
+93
to
+94
| if hasattr(dist, "is_initialized") and dist.is_initialized(): | ||
| average_gradients(model) |
There was a problem hiding this comment.
reason for these changes?
There was a problem hiding this comment.
when using the multiprocessing approach, torch.distributed is not used (init_process_group is not called). so, any reference to dist.get_rank or dist.get_world_size() will error out.
Comment on lines
+98
to
+99
| if hasattr(dist, "is_initialized") and dist.is_initialized(): | ||
| assert hook._get_worker_name() == f"worker_{dist.get_rank()}" |
Comment on lines
+93
to
+94
| if hasattr(dist, "is_initialized") and dist.is_initialized(): | ||
| average_gradients(model) |
|
@vandanavk updates? |
|
In last sprint meeting, we decided to modify the example to use torch.distributed instead of multiprocessing directly. The solution in this PR doesn't fix all scenarios - example, with include_worker="one" |
leleamol
added a commit
that referenced
this pull request
Dec 8, 2020
* Bugfix: Invalid Worker (#139) * smdistributed.dataparallel environment check * addressed comments * Modified check_smdataparallel_env logic Co-authored-by: Nihal Harish <nihal42harish@gmail.com> Co-authored-by: Karan Jariwala <karankjariwala@gmail.com>
Sign up for free
to join this conversation on GitHub.
Already have an account?
Sign in to comment
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.
Description of changes:
Training a model by splitting the dataset across multiple processes on the same machine is considered distributed training. While using SM Debugger along with a distributed training script, the user can provide the option to save data from all workers or just 1 worker (include_workers in the hook).
Currently, this option is used in the following scenarios:
The example in https://github.com/awslabs/amazon-sagemaker-examples/tree/master/sagemaker-python-sdk/dgl_kge uses Python multiprocessing for MXNet distributed training and torch.multiprocessing for PyTorch distributed training. Performing distributed training using multiprocessing is not yet handled by smdebug.
To handle this scenario, introducing env variables SMDEBUG_NUM_WORKERS and SMDEBUG_WORKER_NAME. User must specify these if they are using multiprocessing library in the training script.
The other alternatives considered were:
sagemaker-debugger/tests/pytorch/test_distributed_training.py
Line 98 in 2433316
Style and formatting:
I have run
pre-commit installto ensure that auto-formatting happens with every commit.Issue number, if available
By submitting this pull request, I confirm that you can use, modify, copy, and redistribute this contribution, under the terms of your choice.