Skip to content
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

Jump to bottom

Fix AC order for FLAVA example #434

Open
wants to merge 2 commits into
base: main
Choose a base branch
from
Open

Fix AC order for FLAVA example #434

Changes from all commits
Commits
Show all changes
2 commits
Select commit Hold shift + click to select a range
a7c8acc
Dont use scaler for bf16
rohan-varma Jul 7, 2023
aefeaea
Fix AC order for FLAVA example
rohan-varma Jul 7, 2023
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
39 changes: 19 additions & 20 deletions examples/flava/native/train.py
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -117,7 +117,11 @@ def __init__(self, config: DictConfig):
else torch.float16
)

self.scaler = ShardedGradScaler() if config.training.enable_amp else None
self.scaler = (
ShardedGradScaler()
if config.training.enable_amp and self.half_dtype == torch.float16
else None
)

def log(
self,
Expand All @@ -144,25 +148,6 @@ def create_model(self) -> torch.nn.Module:
f"size: {get_model_size_gb(model):.3} GB"
)

if self.config.training.activation_checkpointing:
check_fn = lambda submodule: isinstance(submodule, TransformerEncoderLayer)
checkpoint_impl = CheckpointImpl.REENTRANT

# DDP gradient hooks have compatibility issues with REENTRANT autograd
if strategy == "ddp":
checkpoint_impl = CheckpointImpl.NO_REENTRANT

checkpoint_wrapper_fn = partial(
checkpoint_wrapper,
offload_to_cpu=False,
checkpoint_impl=checkpoint_impl,
)
apply_activation_checkpointing(
model,
checkpoint_wrapper_fn=checkpoint_wrapper_fn,
check_fn=check_fn,
)

if strategy == "ddp":
# TODO do we have to do this in FSDP too? see https://github.com/pytorch/pytorch/issues/75478
model = torch.nn.SyncBatchNorm.convert_sync_batchnorm(model)
Expand Down Expand Up @@ -206,6 +191,20 @@ def create_model(self) -> torch.nn.Module:

print0(f"after FSDP {torch.cuda.memory_allocated()/1024**3:.3} GB")

if self.config.training.activation_checkpointing:
check_fn = lambda submodule: isinstance(submodule, TransformerEncoderLayer)
checkpoint_impl = CheckpointImpl.NO_REENTRANT

checkpoint_wrapper_fn = partial(
checkpoint_wrapper,
checkpoint_impl=checkpoint_impl,
)
apply_activation_checkpointing(
model,
checkpoint_wrapper_fn=checkpoint_wrapper_fn,
check_fn=check_fn,
)

else:
raise ValueError(f"unknown strategy: {strategy}")

Expand Down
Loading