* Add single checkpoint capability
* Fix checkpointing in test, few cleanups
* Add comment
* Change restore logic
* Move vars around, add better explanation, make todo align with DeepSpeed team
* Fix checkpointing
* Remove deepspeed from extra, install in Dockerfile
* push
* pull
* Split to two tests to see if it fixes Deepspeed error
* Add comment
* Fix some test errors
Summary:
Test Plan:
Reviewers:
Subscribers:
Tasks:
Tags:
* checkpoint consolidation
* Update ddp_spawn.py
* Update test_metric_result_integration.py
* Update test_results.py
* Update utils.py
* Update utils.py
* Update test_all_gather_grad.py
* Update test_all_gather_grad.py
* Update test_results.py
* Revert "Update test_results.py"
This reverts commit 9d4a2b891d.
* Revert "Merge pull request #1 from shuyingsunshine21/shuyingsunshine21-checkpoint_consolidate"
This reverts commit c5053da789, reversing
changes made to 0d23d75bc9.
* Revert "Update test_all_gather_grad.py"
This reverts commit 0d23d75bc9.
* Revert "Update utils.py"
This reverts commit 70fe5da9c6.
* Revert "Update utils.py"
This reverts commit a9aae99f6e.
* Revert "Update test_results.py"
This reverts commit ea74906878.
* Revert "Update test_metric_result_integration.py"
This reverts commit bf70e431b3.
* Revert "Update ddp_spawn.py"
This reverts commit f17210183b.
* Revert "checkpoint consolidation"
This reverts commit 536c1323b0.
* Revert "Revert "checkpoint consolidation""
This reverts commit 3a9fde915a.
* Revert "Revert "Revert "checkpoint consolidation"""
This reverts commit 7a369f47e1.
* Revert "Revert "Update ddp_spawn.py""
This reverts commit 8222dc98ea.
* Revert "Revert "Update test_metric_result_integration.py""
This reverts commit 6c095b2370.
* Revert "Revert "Update test_results.py""
This reverts commit 250d0aaaa2.
* Revert "Revert "Update utils.py""
This reverts commit 8651d54d79.
* Revert "Revert "Update test_all_gather_grad.py""
This reverts commit dcdcd29731.
* modify distributed environment to make test pass
* add DDP communication hook
* remove test related setting
* remove more test related setting
* fix ddp comm hook util import issue
* comments
* one more fix for test_custom_plugin
* fix ddp spwan
* fix sgd
* address comments and add tests
* 1. add is gpu checking 2. modify test a bit 3. formatting
* formatting nit
* fix conda 3.7 1.7 issue for no torch.distributed.algorithms module
* need at least 1.8.0
* minor fix
* modify changelog
* changelog should link to PR number instead of issue number
* refine a bit on doc for register_ddp_comm_hook function, like ddp_comm_wrapper explanation and add hyperparameter for power sgd states in example usge
* move single device checking before call register_ddp_comm_hook
* formatting
* comments
* typo
* pre-commit formatting
* Add context to call hook to handle all modules defined within the hook
* Expose some additional parameters
* Added docs, exposed parameters
* Make sure we only configure if necessary
* Setup activation checkpointing regardless, saves the user having to do it manually
* Add some tests that fail currently
* update
* update
* update
* add tests
* change docstring
* resolve accumulate_grad_batches
* resolve flake8
* Update DeepSpeed to use latest version, add some comments
* add metrics
* update
* Small formatting fixes, clean up some code
* Few cleanups
* No need for default state
* Fix tests, add some boilerplate that should move eventually
* Add hook removal
* Add a context manager to handle hook
* Small naming cleanup
* wip
* move save_checkpoint responsability to accelerator
* resolve flake8
* add BC
* Change recommended scale to 16
* resolve flake8
* update test
* update install
* update
* update test
* update
* update
* update test
* resolve flake8
* update
* update
* update on comments
* Push
* pull
* Update pytorch_lightning/plugins/training_type/deepspeed.py
Co-authored-by: Adrian Wälchli <aedu.waelchli@gmail.com>
* Update pytorch_lightning/plugins/training_type/deepspeed.py
Co-authored-by: Adrian Wälchli <aedu.waelchli@gmail.com>
* update
* Apply suggestions from code review
* Swap to using world size defined by plugin
* update
* update todo
* Remove deepspeed from extra, keep it in the base cuda docker install
* Push
* pull
* update
* update
* update
* update
* Minor changes
* duplicate
* format
* format2
Co-authored-by: SeanNaren <sean@grid.ai>
Co-authored-by: Adrian Wälchli <aedu.waelchli@gmail.com>
Co-authored-by: Sean Naren <sean.narenthiran@gmail.com>
Co-authored-by: Carlos Mocholi <carlossmocholi@gmail.com>
Co-authored-by: Jirka Borovec <Borda@users.noreply.github.com>
Co-authored-by: Jirka Borovec <jirka.borovec@seznam.cz>
* Add base hook for model parallel
* fix callback signature
* Simplify hook
* Add hook logic
* add tests
* add property setter
* add logic for being called once
* Update changelog
* Fix
* fix return type
* fix lambda callback test
* Fix tests
* Apply code suggestions
* add logic for setup_optimizers_predispatch
* add common dummy model
* Swap call order
* Remove test that isn't needed anymore
* Update tests
* Add a bit more doc
* Few code review fixes
* Update pytorch_lightning/accelerators/accelerator.py
Co-authored-by: Carlos Mocholí <carlossmocholi@gmail.com>
* Change hook name
* Fix test
* Test setup hook, refactor names
* Swap call order of callbacks and model initialization
* Change name of context manager
Co-authored-by: SeanNaren <sean@grid.ai>
Co-authored-by: Sean Naren <sean.narenthiran@gmail.com>
Co-authored-by: Carlos Mocholí <carlossmocholi@gmail.com>
* Move connection setup into the setup function. Call setup hook after we set up the accelerator
* Added CHANGELOG.md
* fix setup order in callback test
* fix input arguments in test
* Mock distributed function, remove protection to turn into training type hook
* Remove import
* Add missing mock, ensure custom plugin does not create children process
* Skip test on windows
* Update deepspeed to init connection in setup
* Do not initialize distributed module
* Move DeepSpeed tests to special tests since dist communication is being set up
* Special the test to see if this fixes CI
* Delete accelerator connector test to see if its causing build to fail
* Delete deepspeed test
* Revert "Delete accelerator connector test to see if its causing build to fail"
This reverts commit edde60b8
* Revert "Delete deepspeed test"
This reverts commit 9d317429
* Reverse hook
* Reverse setup hooks to debug again
* Add todo so i know where i left off
* For single device move in pre_dispatch after setup function
* Add additional model to device hook if any additional parameters have been set
* See if we can enable deepspeed tests
* Revert "See if we can enable deepspeed tests"
This reverts commit b5450def
* See if this hook approach works
* Introduce new granular hooks
* Remove import, fix tpu spawn by moving the function to setup
* Added missing special test
Co-authored-by: Adrian Wälchli <aedu.waelchli@gmail.com>
* Ensure we set the default device before initializing deepspeed
* Add CHANGELOG.md
* Update pytorch_lightning/plugins/training_type/deepspeed.py
Co-authored-by: Kaushik B <45285388+kaushikb11@users.noreply.github.com>
Co-authored-by: Kaushik B <45285388+kaushikb11@users.noreply.github.com>
* Update code
Co-authored-by: EliaCereda
* More property updates
* Move properties. Introduce trainer._fitting
* Use trainer.fitting
* Fix reset dataloaders
* Unused code
* RunningStage.SANITY_CHECKING
* Use setters
* Fix bugs
* Fix bugs
* TrainerState.{FITTING,VALIDATING,TESTING,PREDICTING,TUNING}
* Fix bugs
* Fix bugs
* Fix tests
* Update CHANGELOG. Add deprecation warning. Fix tests
* Unused imports
* Optional trainer
* More deprecation. More refactoring
* Correct version
* Use properties
* Address comments
* flake8
* Missed renamings
* Typo
* is -> ==
It is recommended to use for Enums since they are singletons, however, since the LightningEnum subclasses str, it's not a good idea in case a user sets the state/stage with a str
* Also for tests
* Typo
* Address @tchaton's comments
* PEP8
* Correct property
* Update CHANGELOG
* Apply suggestions from code review
Co-authored-by: Adrian Wälchli <aedu.waelchli@gmail.com>
* Update pytorch_lightning/trainer/trainer.py
Co-authored-by: Adrian Wälchli <aedu.waelchli@gmail.com>
* Remove called sanity check
Co-authored-by: Carlos Mocholi <carlossmocholi@gmail.com>
Co-authored-by: Adrian Wälchli <aedu.waelchli@gmail.com>
* Change default for CPU offload to false for best throughput/memory efficiency
* Add changelog
* default
Co-authored-by: Jirka Borovec <Borda@users.noreply.github.com>
Hinrich B. Winther