Skip to content

opt reader and gather #11016

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

Open
wants to merge 1 commit into
base: dsv3_dev
Choose a base branch
from
Open

opt reader and gather #11016

Show file tree
Hide file tree
Changes from all commits
Commits
Show all changes
1 commit
Select commit
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
7 changes: 6 additions & 1 deletion paddlenlp/data/dist_dataloader.py
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -197,7 +197,12 @@ def __next__(self):
data = nested_copy_place(data, place=paddle.framework._current_expected_place())
except Exception as e:
logger.debug(e)
data = self._broadcast_data(data)
# data = self._broadcast_data(data)
if data is None:
data = {
"input_ids": paddle.empty([1, 4097], paddle.int64),
"labels": paddle.empty([1, 4097], paddle.int64),
}
return data


Expand Down
23 changes: 12 additions & 11 deletions paddlenlp/trainer/trainer.py
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -156,7 +156,8 @@
)
from .training_args import TrainingArguments
from .unified_checkpoint import UnifiedCheckpointHandler
from .utils import reshard as reshard_util

# from .utils import reshard as reshard_util
from .utils.async_save import AsyncSaver
from .utils.helper import ( # nested_truncate,
broadcast_dataset_rank0_model,
Expand Down Expand Up @@ -1185,10 +1186,10 @@ def fused_allreduce_gradients_no_sync(paramlist, hcg):
fused_allreduce_gradients_no_sync(list(model.parameters()), None)

# Pipeline parallel mode, handle gradient reduce here to overlap
enable_dp_comm_overlap = (
self.args.pipeline_parallel_degree > 1
and "enable_dp_comm_overlap" in args.pipeline_parallel_config
)
# enable_dp_comm_overlap = (
# self.args.pipeline_parallel_degree > 1
# and "enable_dp_comm_overlap" in args.pipeline_parallel_config
# )

enable_release_grads = False
if args.sharding_parallel_degree > 1:
Expand All @@ -1200,13 +1201,13 @@ def fused_allreduce_gradients_no_sync(paramlist, hcg):
if isinstance(self.optimizer, HybridParallelOptimizer) and not self.do_grad_scaling:
parameters_list = _obtain_optimizer_parameters_list(self.optimizer._inner_opt)

if not enable_dp_comm_overlap:
if self.optimizer._sharding_enable:
assert reshard_util.is_sharding_opt(self.optimizer)
self.optimizer._inner_opt.reduce_gradients(list(parameters_list), self.optimizer._hcg)
# if not enable_dp_comm_overlap:
# if self.optimizer._sharding_enable:
# assert reshard_util.is_sharding_opt(self.optimizer)
# self.optimizer._inner_opt.reduce_gradients(list(parameters_list), self.optimizer._hcg)

if self.optimizer._dp_enable or getattr(self.optimizer, "_sep_enable", False):
fused_allreduce_gradients_no_sync(list(parameters_list), self.optimizer._hcg)
# if self.optimizer._dp_enable or getattr(self.optimizer, "_sep_enable", False):
# fused_allreduce_gradients_no_sync(list(parameters_list), self.optimizer._hcg)
self.timers and self.timers("all-reduce").stop()
self.timers and self.timers("optimizer-step").start()

Expand Down
Loading