Skip to content

Revert workaround for T5 that sets number of workers to 0 & sync_batch_comm=False #5420

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
okuchaiev merged 6 commits into from
Nov 16, 2022
Merged

Revert workaround for T5 that sets number of workers to 0 & sync_batch_comm=False #5420

Changes from all commits
Commits
Show all changes
6 commits
Select commit Hold shift + click to select a range
c67bcbb
Revert workers workaround
MaximumEntropy Nov 15, 2022
c045c8b
Merge branch 'r1.13.0' into undo_t5_workers_workaround
okuchaiev Nov 15, 2022
ef10d53
Fix in config
MaximumEntropy Nov 16, 2022
f23a1a6
Merge branch 'r1.13.0' into undo_t5_workers_workaround
MaximumEntropy Nov 16, 2022
b1971b6
Fix
MaximumEntropy Nov 16, 2022
4464c93
Merge branch 'undo_t5_workers_workaround' of github.com:NVIDIA/NeMo i…
MaximumEntropy Nov 16, 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
10 changes: 9 additions & 1 deletion nemo/collections/nlp/models/language_modeling/megatron_lm_encoder_decoder_model.py
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -337,6 +337,7 @@ def training_step(self, batch, batch_idx):
tensor_shape=tensor_shape,
decoder_sequence_length=decoder_seq_length,
dtype=self.autocast_dtype,
sync_batch_comm=self.cfg.get('sync_batch_comm', False),
grad_scaler=self.trainer.precision_plugin.scaler if self.cfg.precision == 16 else None,
custom_sync_context_handler=custom_sync_context_handler,
)
Expand All @@ -349,6 +350,7 @@ def training_step(self, batch, batch_idx):
tensor_shape=tensor_shape,
decoder_sequence_length=decoder_seq_length,
dtype=self.autocast_dtype,
sync_batch_comm=self.cfg.get('sync_batch_comm', False),
grad_scaler=self.trainer.precision_plugin.scaler if self.cfg.precision == 16 else None,
custom_sync_context_handler=custom_sync_context_handler,
)
Expand Down Expand Up @@ -657,6 +659,7 @@ def validation_step_logits(self, batch, batch_idx):
tensor_shape=tensor_shape,
decoder_sequence_length=decoder_seq_length,
dtype=self.autocast_dtype,
sync_batch_comm=self.cfg.get('sync_batch_comm', False),
grad_scaler=self.trainer.precision_plugin.scaler if self.cfg.precision == 16 else None,
)
else:
Expand All @@ -668,6 +671,7 @@ def validation_step_logits(self, batch, batch_idx):
tensor_shape=tensor_shape,
decoder_sequence_length=decoder_seq_length,
dtype=self.autocast_dtype,
sync_batch_comm=self.cfg.get('sync_batch_comm', False),
grad_scaler=self.trainer.precision_plugin.scaler if self.cfg.precision == 16 else None,
)

Expand Down Expand Up @@ -700,6 +704,7 @@ def validation_step(self, batch, batch_idx):
tensor_shape=tensor_shape,
decoder_sequence_length=decoder_seq_length,
dtype=self.autocast_dtype,
sync_batch_comm=self.cfg.get('sync_batch_comm', False),
grad_scaler=self.trainer.precision_plugin.scaler if self.cfg.precision == 16 else None,
)
else:
Expand All @@ -711,6 +716,7 @@ def validation_step(self, batch, batch_idx):
tensor_shape=tensor_shape,
decoder_sequence_length=decoder_seq_length,
dtype=self.autocast_dtype,
sync_batch_comm=self.cfg.get('sync_batch_comm', False),
grad_scaler=self.trainer.precision_plugin.scaler if self.cfg.precision == 16 else None,
)

Expand Down Expand Up @@ -951,7 +957,7 @@ def setup_validation_data(self, cfg):
if hasattr(self, '_validation_ds'):
consumed_samples = 0
self._validation_dl = self.build_pretraining_data_loader(
self._validation_ds, consumed_samples, num_workers=0
self._validation_ds, consumed_samples, num_workers=self._cfg.data.num_workers
)

def setup_test_data(self, cfg):
Expand Down Expand Up @@ -1042,6 +1048,7 @@ def dummy():
tensor_shape=tensor_shape,
decoder_sequence_length=encoder_seq_length,
dtype=self.autocast_dtype,
sync_batch_comm=self.cfg.get('sync_batch_comm', False),
)
else:
output_tensor = forward_backward_no_pipelining(
Expand All @@ -1052,6 +1059,7 @@ def dummy():
tensor_shape=tensor_shape,
decoder_sequence_length=encoder_seq_length,
dtype=self.autocast_dtype,
sync_batch_comm=self.cfg.get('sync_batch_comm', False),
)

if output_tensor:
Expand Down