Multiple fixes in SageMakerTrainer #10687
Merged
Conversation
This file contains 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
sgugger
commented
Mar 12, 2021
Comment on lines
+149
to
+150
| if self.is_model_parallel_enabled: | ||
| self._save_smp(output_dir) |
There was a problem hiding this comment.
We need to use a special save because model parallelism requires to:
- gather the state dict on all processes of d_rank 0 (it triggers a sync across those processes)
- save it only on processes 0
sgugger
commented
Mar 12, 2021
Comment on lines
+199
to
+200
| # Consolidate the state dict on all processed of dp_rank 0 | ||
| opt_state_dict = self.optimizer.state_dict() |
There was a problem hiding this comment.
The method is overloaded for this particular line/behavior. As for the model, the state dict of the optimizer needs to be gathered from all dp rank 0 processes.
| os.path.join(checkpoint, "scheduler.pt") | ||
| ): | ||
| self.optimizer.load_state_dict( | ||
| torch.load(os.path.join(checkpoint, "optimizer.pt"), map_location="cpu") |
There was a problem hiding this comment.
The method is overloaded for this particular line (needs to be loaded on the CPU and not the device).
| @@ -927,6 +924,9 @@ def train( | |||
| if delay_optimizer_creation: | |||
| self.create_optimizer_and_scheduler(num_training_steps=max_steps) | |||
|
|
|||
| # Check if saved optimizer or scheduler states exist | |||
| self._load_optimizer_and_scheduler(resume_from_checkpoint) | |||
There was a problem hiding this comment.
This move should be harmless.
| elif self.args.local_rank != -1: | ||
| world_size = dist.get_world_size() | ||
| world_size = max(1, world_size) | ||
| world_size = max(1, self.args.world_size) |
There was a problem hiding this comment.
Refactor + allows world_size to be overloaded in SageMakerTrainingArguments
philschmid
approved these changes
Mar 12, 2021
anirudh2290
approved these changes
Mar 14, 2021
Iwontbecreative
pushed a commit
to Iwontbecreative/transformers
that referenced
this pull request
Jul 15, 2021
* Handle save differently * Missing imports * Fix typo * Adapt to recent changes in save_pretrained * Forgotten brackets * Optimizer load * Fix world size * Deal wth None * Remove needless self
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.
What does this PR do?
This PR adds quite a few fixes to the
SageMakerTrainerto make sure example scripts run fully. In particular it fixes:drop_last=Truewhich is not something anyone wants.The goal is now to test a little bit more that functionality before merging the
SageMakerTrainerinto the mainTrainer(otherwise one can't use model parallelism in seq2seq examples or QA example). The plan is to have them merged in the v4.5.0.