Skip to content

Reduce number of times optimizers are instantiated with FSDP #12267

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.

Sign up for GitHub

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

Merged
merged 4 commits into from
Mar 21, 2022
Merged

Reduce number of times optimizers are instantiated with FSDP #12267

Show file tree
Hide file tree
Changes from all commits
Commits
Show all changes
4 commits
Select commit Hold shift + click to select a range
b0ed863
Update fully_sharded.py
ananthsub Mar 8, 2022
10bda43
Update CHANGELOG.md
ananthsub Mar 8, 2022
c81a477
Update fully_sharded.py
ananthsub Mar 10, 2022
92c3ed4
Merge branch 'master' into fix/improve-fsdp-opt
awaelchli Mar 20, 2022
File filter

Filter by extension

Filter by extension
Conversations
Failed to load comments.
Loading
Jump to
Jump to file
Failed to load files.
Loading
Diff view
Diff view
5 changes: 4 additions & 1 deletion CHANGELOG.md
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -806,7 +806,10 @@ The format is based on [Keep a Changelog](http://keepachangelog.com/en/1.0.0/).
- Fixed to avoid common hook warning if no hook is overridden ([#12131](https://github.com/PyTorchLightning/pytorch-lightning/pull/12131))


- Fixed the case where logger=None is passed to the Trainer ([#12249](https://github.com/PyTorchLightning/pytorch-lightning/pull/12249))
- Fixed the case where `logger=None` is passed to the Trainer ([#12249](https://github.com/PyTorchLightning/pytorch-lightning/pull/12249))


- Fixed initializing optimizers unnecessarily in `DDPFullyShardedStrategy` ([#12267](https://github.com/PyTorchLightning/pytorch-lightning/pull/12267))


## [1.5.10] - 2022-02-08
Expand Down
10 changes: 3 additions & 7 deletions pytorch_lightning/strategies/fully_sharded.py
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -138,16 +138,15 @@ def setup_distributed(self) -> None:

def setup(self, trainer: "pl.Trainer") -> None:
self.accelerator.setup(trainer)
self.setup_optimizers(trainer)
self.setup_precision_plugin()
optimizers_to_device(self.optimizers, self.root_device)

if trainer.state.fn == TrainerFn.FITTING and self._layer_sync:
self.model = self._layer_sync.apply(self.model)

self.configure_ddp()
self.barrier()
self.setup_optimizers(trainer)
optimizers_to_device(self.optimizers, self.root_device)
self.setup_precision_plugin()

@contextlib.contextmanager
def model_sharded_context(self) -> Generator:
Expand Down Expand Up @@ -176,17 +175,14 @@ def wrap_policy(*args, **kwargs):
log.detail(f"{self.__class__.__name__}: exiting model_sharded_context.")

def configure_ddp(self) -> None:
log.detail(f"{self.__class__.__name__}: configuring DDP... (cpu_offload: [{self.cpu_offload}])")
log.detail(f"{self.__class__.__name__}: configuring FSDP... (cpu_offload: [{self.cpu_offload}])")
if not self.cpu_offload:
# When using CPU Offload, FSDP will manage the CUDA movement for us.
# Note: this would be problematic for large model (which could not fit in one GPU)
# as FSDP module.to(device) would first summon all parameters
# (TODO: need to figure out solution)
self.model_to_device()

# setup optimizers after fully sharded has wrapped the lightning module
self.setup_optimizers(self.lightning_module.trainer)

def model_to_device(self) -> None:
log.detail(f"{self.__class__.__name__}: moving model to device [{self.root_device}]...")
# ensure we update the device type in the lightning module
Expand Down