-
Notifications
You must be signed in to change notification settings - Fork 3.4k
New issue
Have a question about this project? Sign up for a free GitHub account to open an issue and contact its maintainers and the community.
By clicking “Sign up for GitHub”, you agree to our terms of service and privacy statement. We’ll occasionally send you account related emails.
Already on GitHub? Sign in to your account
[fix] Better support for rank_zero_only setting for SLURM and torchelastic #6802
Conversation
Hello @ananthsub! Thanks for updating this PR. There are currently no PEP 8 issues detected in this Pull Request. Cheers! 🍻 Comment last updated at 2021-04-07 08:54:44 UTC |
Codecov Report
@@ Coverage Diff @@
## master #6802 +/- ##
========================================
- Coverage 92% 44% -48%
========================================
Files 193 192 -1
Lines 12271 12191 -80
========================================
- Hits 11234 5346 -5888
- Misses 1037 6845 +5808 |
There was a problem hiding this comment.
Choose a reason for hiding this comment
The reason will be displayed to describe this comment to others. Learn more.
note, the tests won't run because the file distributed.py
is not prefixed with test_
Co-authored-by: Adrian Wälchli <aedu.waelchli@gmail.com>
@@ -44,8 +44,18 @@ def wrapped_fn(*args, **kwargs): | |||
return wrapped_fn | |||
|
|||
|
|||
# TODO: this should be part of the cluster environment | |||
def _get_rank() -> int: |
There was a problem hiding this comment.
Choose a reason for hiding this comment
The reason will be displayed to describe this comment to others. Learn more.
Can you move this directly to SLURM cluster environment ?
What if RANK
, SLURM_PROCID
or LOCAL_RANK
are different ? Should we take the latest or did you order rank_keys
based on priority ?
Best,
T.C
There was a problem hiding this comment.
Choose a reason for hiding this comment
The reason will be displayed to describe this comment to others. Learn more.
RANK = torchelastic
SLURM_PROCID = slurm
LOCAL_RANK = parity with existing setup though I think it's not right
I set local rank last as RANK and SLURM_PROCID correspond to global rank already. The linked issue has more discussion, but I think we should make global rank and world size properties of the cluster environment. So the cluster environment becomes the source of truth propagating from Cluster environment => training type plugin => accelerator => trainer.
The main issue now is the global rank isn't set on trainer initialization. If the cluster environment is marked as creating children, then we can leave the initialization of these fields for later, but both torchelastic and slurm have this data already available in the environment variables, and we should expose that as soon as possible (on Trainer init) for users to read this state.
Currently, this waits for trainer.fit()
to be called, going through the accelerator setup flow for these properties to be initialized on the training type plugin.
Co-authored-by: Adrian Wälchli <aedu.waelchli@gmail.com>
…ng into fix-slurm-rank
…ter) to github/third-party/PyTorchLightning/pytorch-lightning Summary: ### New commit log messages ## [UnReleased] - 2021-MM-DD ### Added - Added more explicit exception message when trying to execute `trainer.test()` or `trainer.validate()` with `fast_dev_run=True` ([#6667](Lightning-AI/pytorch-lightning#6667)) - Added `LightningCLI` class to provide simple reproducibility with minimum boilerplate training cli. ([#4492](Lightning-AI/pytorch-lightning#4492)) - Trigger warning when non-metric logged value with multi processes hasn't been reduced ([#6417](Lightning-AI/pytorch-lightning#6417)) - Added `gradient_clip_algorithm` argument to Trainer for gradient clipping by value ([#6123](Lightning-AI/pytorch-lightning#6123)). - Added a way to print to terminal without breaking up the progress bar ([#5470](Lightning-AI/pytorch-lightning#5470)) - Added support to checkpoint after training steps in `ModelCheckpoint` callback ([#6146](Lightning-AI/pytorch-lightning#6146)) - Added `checkpoint` parameter to callback's `on_save_checkpoint` hook ([#6072](Lightning-AI/pytorch-lightning#6072)) - Added `RunningStage.SANITY_CHECKING` ([#4945](Lightning-AI/pytorch-lightning#4945)) - Added `TrainerState.{FITTING,VALIDATING,TESTING,PREDICTING,TUNING}` ([#4945](Lightning-AI/pytorch-lightning#4945)) - Added `Trainer.validate()` method to perform one evaluation epoch over the validation set ([#4948](Lightning-AI/pytorch-lightning#4948)) - Added `LightningEnvironment` for Lightning-specific DDP ([#5915](Lightning-AI/pytorch-lightning#5915)) - Added `teardown()` hook to LightningDataModule ([#4673](Lightning-AI/pytorch-lightning#4673)) - Added `auto_insert_metric_name` parameter to `ModelCheckpoint` ([#6277](Lightning-AI/pytorch-lightning#6277)) - Added arg to `self.log` that enables users to give custom names when dealing with multiple dataloaders ([#6274](Lightning-AI/pytorch-lightning#6274)) - Added `teardown` method to `BaseProfiler` to enable subclasses defining post-profiling steps outside of `__del__` ([#6370](Lightning-AI/pytorch-lightning#6370)) - Added `setup` method to `BaseProfiler` to enable subclasses defining pre-profiling steps for every process ([#6633](Lightning-AI/pytorch-lightning#6633)) - Added no return warning to predict ([#6139](Lightning-AI/pytorch-lightning#6139)) - Added `Trainer.predict` config validation ([#6543](Lightning-AI/pytorch-lightning#6543)) - Added `AbstractProfiler` interface ([#6621](Lightning-AI/pytorch-lightning#6621)) - Added support for including module names for forward in the autograd trace of `PyTorchProfiler` ([#6349](Lightning-AI/pytorch-lightning#6349)) - Added support for the PyTorch 1.8.1 autograd profiler ([#6618](Lightning-AI/pytorch-lightning#6618)) - Added `outputs` parameter to callback's `on_validation_epoch_end` & `on_test_epoch_end` hooks ([#6120](Lightning-AI/pytorch-lightning#6120)) - Added `configure_sharded_model` hook ([#6679](Lightning-AI/pytorch-lightning#6679)) - Added support for `precision=64`, enabling training with double precision ([#6595](Lightning-AI/pytorch-lightning#6595)) - Added support for DDP communication hooks ([#6736](Lightning-AI/pytorch-lightning#6736)) - Added `artifact_location` argument to `MLFlowLogger` which will be passed to the `MlflowClient.create_experiment` call ([#6677](Lightning-AI/pytorch-lightning#6677)) - Added `model` parameter to precision plugins' `clip_gradients` signature ([#6764](Lightning-AI/pytorch-lightning#6764)) ### Changed - Renamed `pytorch_lightning.callbacks.swa` to `pytorch_lightning.callbacks.stochastic_weight_avg` ([#6259](Lightning-AI/pytorch-lightning#6259)) - Refactor `RunningStage` and `TrainerState` usage ([#4945](Lightning-AI/pytorch-lightning#4945)) - Changed `trainer.evaluating` to return `True` if validating or testing ([#4945](Lightning-AI/pytorch-lightning#4945)) - Changed `setup()` and `teardown()` stage argument to take any of `{fit,validate,test,predict}` ([#6386](Lightning-AI/pytorch-lightning#6386)) - Changed profilers to save separate report files per state and rank ([#6621](Lightning-AI/pytorch-lightning#6621)) - Changed `PyTorchProfiler` to use `torch.autograd.profiler.record_function` to record functions ([#6349](Lightning-AI/pytorch-lightning#6349)) ### Deprecated - `period` has been deprecated in favor of `every_n_val_epochs` in the `ModelCheckpoint` callback ([#6146](Lightning-AI/pytorch-lightning#6146)) - Deprecated `trainer.running_sanity_check` in favor of `trainer.sanity_checking` ([#4945](Lightning-AI/pytorch-lightning#4945)) - Deprecated `Profiler(output_filename)` in favor of `dirpath` and `filename` ([#6621](Lightning-AI/pytorch-lightning#6621)) - Deprecated `PytorchProfiler(profiled_functions)` in favor of `record_functions` ([#6349](Lightning-AI/pytorch-lightning#6349)) - Deprecated metrics in favor of `torchmetrics` ([#6505](Lightning-AI/pytorch-lightning#6505), [#6530](Lightning-AI/pytorch-lightning#6530), [#6540](Lightning-AI/pytorch-lightning#6540), [#6547](Lightning-AI/pytorch-lightning#6547), [#6515](Lightning-AI/pytorch-lightning#6515), [#6572](Lightning-AI/pytorch-lightning#6572), [#6573](Lightning-AI/pytorch-lightning#6573), [#6584](Lightning-AI/pytorch-lightning#6584), [#6636](Lightning-AI/pytorch-lightning#6636), [#6637](Lightning-AI/pytorch-lightning#6637), [#6649](Lightning-AI/pytorch-lightning#6649), [#6659](Lightning-AI/pytorch-lightning#6659), ) ### Removed - Removed support for passing a bool value to `profiler` argument of Trainer ([#6164](Lightning-AI/pytorch-lightning#6164)) - Removed no return warning from val/test step ([#6139](Lightning-AI/pytorch-lightning#6139)) - Removed passing a `ModelCheckpoint` instance to `Trainer(checkpoint_callback)` ([#6166](Lightning-AI/pytorch-lightning#6166)) - Removed deprecated Trainer argument `enable_pl_optimizer` and `automatic_optimization` ([#6163](Lightning-AI/pytorch-lightning#6163)) - Removed deprecated metrics ([#6161](Lightning-AI/pytorch-lightning#6161)) * from `pytorch_lightning.metrics.functional.classification` removed `to_onehot`, `to_categorical`, `get_num_classes`, `roc`, `multiclass_roc`, `average_precision`, `precision_recall_curve`, `multiclass_precision_recall_curve` * from `pytorch_lightning.metrics.functional.reduction` removed `reduce`, `class_reduce` - Removed deprecated `ModelCheckpoint` arguments `prefix`, `mode="auto"` ([#6162](Lightning-AI/pytorch-lightning#6162)) - Removed `mode='auto'` from `EarlyStopping` ([#6167](Lightning-AI/pytorch-lightning#6167)) - Removed legacy references for magic keys in the `Result` object ([#6016](Lightning-AI/pytorch-lightning#6016)) - Removed deprecated `LightningModule` `hparams` setter ([#6207](Lightning-AI/pytorch-lightning#6207)) - Removed legacy code to log or include metrics in the progress bar by returning them in a dict with the `"log"/"progress_bar"` magic keys. Use `self.log` instead ([#6734](Lightning-AI/pytorch-lightning#6734)) - Removed `optimizer_idx` argument from `training_step` in manual optimization ([#6093](Lightning-AI/pytorch-lightning#6093)) ### Fixed - Set better defaults for `rank_zero_only.rank` when training is launched with SLURM and torchelastic ([#6802](Lightning-AI/pytorch-lightning#6802)) - Made the `Plugin.reduce` method more consistent across all Plugins to reflect a mean-reduction by default ([#6011](Lightning-AI/pytorch-lightning#6011)) - Move lightning module to correct device type when using LightningDistributedWrapper ([#6070](Lightning-AI/pytorch-lightning#6070)) - Do not print top-k verbose log with `ModelCheckpoint(monitor=None)` ([#6109](Lightning-AI/pytorch-lightning#6109)) - Fixed csv extension check ([#6436](Lightning-AI/pytorch-lightning#6436)) - Fixed `ModelCheckpoint(monitor=None, save_last=True)` not saving checkpoints ([#6136](Lightning-AI/pytorch-lightning#6136)) - Fixed `ModelCheckpoint(save_top_k=0, save_last=True)` not saving the `last` checkpoint ([#6136](Lightning-AI/pytorch-lightning#6136)) - Fixed `.teardown(stage='fit')` getting called during `trainer.test` ([#6386](Lightning-AI/pytorch-lightning#6386)) - Fixed `.on_fit_{start,end}()` getting called during `trainer.test` ([#6386](Lightning-AI/pytorch-lightning#6386)) - Fixed LightningModule `all_gather` on cpu tensors ([#6416](Lightning-AI/pytorch-lightning#6416)) - Fixed torch distributed not available in setup hook for DDP ([#6506](Lightning-AI/pytorch-lightning#6506)) - Fixed `EarlyStopping` logic when `min_epochs` or `min_steps` requirement is not met ([#6705](Lightning-AI/pytorch-lightning#6705)) ## [1.2.7] - 2021-04-06 ### Fixed - Fixed resolve a bug with omegaconf and xm.save ([#6741](Lightning-AI/pytorch-lightning#6741)) - Fixed an issue with IterableDataset when __len__ is not defined ([#6828](Lightning-AI/pytorch-lightning#6828)) - Sanitize None params during pruning ([#6836](Lightning-AI/pytorch-lightning#6836)) - Enforce an epoch scheduler interval when using SWA ([#6588](Lightning-AI/pytorch-lightning#6588)) - Fixed TPU Colab hang issue, post training ([#6816](Lightning-AI/pytorch-lightning#6816)) - Fixed a bug where `TensorBoardLogger` would give a warning and not log correctly to a symbolic link `save_dir` ([#6730](Lightning-AI/pytorch-lightning#6730)) ## [1.2.6] - 2021-03-30 ### Changed - Changed the behavior of `on_epoch_start` to run at the beginning of validation & test epoch ([#6498](Lightning-AI/pytorch-lightning#6498)) ### Removed - Removed legacy code to include `step` dictionary returns in `callback_metrics`. Use `self.log_dict` instead. ([#6682](Lightning-AI/pytorch-lightning#6682)) ### Fixed - Fixed `DummyLogger.log_hyperparams` raising a `TypeError` when running with `fast_dev_run=True` ([#6398](Lightning-AI/pytorch-lightning#6398)) - Fixed error on TPUs when there was no `ModelCheckpoint` ([#6654](Lightning-AI/pytorch-lightning#6654)) - Fixed `trainer.test` freeze on TPUs ([#6654](Lightning-AI/pytorch-lightning#6654)) - Fixed a bug where gradients were disabled after calling `Trainer.predict` ([#6657](Lightning-AI/pytorch-lightning#6657)) - Fixed bug where no TPUs were detected in a TPU pod env ([#6719](Lightning-AI/pytorch-lightning#6719)) ## [1.2.5] - 2021-03-23 ### Changed - Update Gradient Clipping for the TPU Accelerator ([#6576](Lightning-AI/pytorch-lightning#6576)) - Refactored setup for typing friendly ([#6590](Lightning-AI/pytorch-lightning#6590)) ### Fixed - Fixed a bug where `all_gather` would not work correctly with `tpu_cores=8` ([#6587](Lightning-AI/pytorch-lightning#6587)) - Fixed comparing required versions ([#6434](Lightning-AI/pytorch-lightning#6434)) - Fixed duplicate logs appearing in console when using the python logging module ([#6275](Lightning-AI/pytorch-lightning#6275)) - Added Autocast in validation, test and predict modes for Native AMP ([#6565](Lightning-AI/pytorch-lightning#6565)) Reviewed By: shuyingsunshine21 Differential Revision: D27528929 fbshipit-source-id: 311c88f71461c2c79bbf185e28d7a6d683ccc26f
After Sasha's update of pytorch lightning on MMF master, it broke MMF codebase for multinode job. The root problem to PR Lightning-AI/pytorch-lightning#6802. The assumption that SLURM_PROCID points to worker rank is wrong as some frameworks launch their own processes later using multiprocessing spawn and have ntasks_per_node=1 set. This means that first node will have procid = 0, second node will have procid = 1 set and so on. Now, since this is used in prepare_data masking in LightningDataModule, this leads to it running on all workers on first node and thus causing inconsistencies. Now, this leads to prepare_data being called on all workers on first node instead of rank zero. Specifically, the barrier call in prepare_data, is called on first node workers but not on others leading to block later on. This PR fixes this by ensuring on our side that we only call prepare_data on rank zero. Furthermore, this can cause further confusion, we remove sync barrier calls from download as well. Users are now supposed to handle is_master checks on their own. Test Plan: Tested in multinode settings.
After Sasha's update of pytorch lightning on MMF master, it broke MMF codebase for multinode job. The root problem to PR Lightning-AI/pytorch-lightning#6802. The assumption that SLURM_PROCID points to worker rank is wrong as some frameworks launch their own processes later using multiprocessing spawn and have ntasks_per_node=1 set. This means that first node will have procid = 0, second node will have procid = 1 set and so on. Now, since this is used in prepare_data masking in LightningDataModule, this leads to it running on all workers on first node and thus causing inconsistencies. Now, this leads to prepare_data being called on all workers on first node instead of rank zero. Specifically, the barrier call in prepare_data, is called on first node workers but not on others leading to block later on. This PR fixes this by ensuring on our side that we only call prepare_data on rank zero. Furthermore, this can cause further confusion, we remove sync barrier calls from download as well. Users are now supposed to handle is_master checks on their own. Test Plan: Tested in multinode settings.
Summary: After Sasha's update of pytorch lightning on MMF master, it broke MMF codebase for multinode job. The root problem to PR Lightning-AI/pytorch-lightning#6802. The assumption that SLURM_PROCID points to worker rank is wrong as some frameworks launch their own processes later using multiprocessing spawn and have ntasks_per_node=1 set. This means that first node will have procid = 0, second node will have procid = 1 set and so on. Now, since this is used in prepare_data masking in LightningDataModule, this leads to it running on all workers on first node and thus causing inconsistencies. Now, this leads to prepare_data being called on all workers on first node instead of rank zero. Specifically, the barrier call in prepare_data, is called on first node workers but not on others leading to block later on. This PR fixes this by ensuring on our side that we only call prepare_data on rank zero. Furthermore, this can cause further confusion, we remove sync barrier calls from download as well. Users are now supposed to handle is_master checks on their own. Pull Request resolved: #921 Test Plan: Tested in multinode settings. Reviewed By: vedanuj Differential Revision: D28156855 Pulled By: apsdehal fbshipit-source-id: 4e0dd5317e15153f558d34c6951a89299602454f
What does this PR do?
Fixes #6797
This is a mitigation for the issue. The environment variable handling is currently split in a few different places:
This needs to be consolidated to better support more custom cluster environments in the future. For now this is a quickfix for SLURM users until we can come up with a better design.
Before submitting
PR review
Anyone in the community is free to review the PR once the tests have passed.
Before you start reviewing make sure you have read Review guidelines. In short, see the following bullet-list:
Did you have fun?
Make sure you had fun coding 🙃