Skip to content

Debug support for interleaved pipeline parallelism with the distributed Adam optimizer #5236

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
MaximumEntropy merged 6 commits into from
Nov 16, 2022
Merged

Debug support for interleaved pipeline parallelism with the distributed Adam optimizer #5236

Show file tree
Hide file tree
Changes from all commits
Commits
Show all changes
6 commits
Select commit Hold shift + click to select a range
dee4492
Use updated API for overlapping grad sync with pipeline parallelism
timmoon10 Oct 24, 2022
c6b50ef
Merge branch 'main' into dist-adam-interleaved-pipeline-parallel
timmoon10 Oct 25, 2022
b8d18cd
Merge branch 'main' into dist-adam-interleaved-pipeline-parallel
timmoon10 Oct 26, 2022
84bb22e
Merge branch 'main' into dist-adam-interleaved-pipeline-parallel
timmoon10 Oct 27, 2022
18cc0b0
Merge branch 'main' into dist-adam-interleaved-pipeline-parallel
timmoon10 Oct 28, 2022
cf6c6e8
Merge branch 'main' into dist-adam-interleaved-pipeline-parallel
timmoon10 Nov 2, 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
11 changes: 6 additions & 5 deletions nemo/collections/nlp/models/language_modeling/megatron_gpt_model.py
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -282,13 +282,16 @@ def training_step(self, batch, batch_idx):
tensor_shape = [self.cfg.encoder_seq_length, self.cfg.micro_batch_size, self.cfg.hidden_size]

# handle asynchronous grad reduction
custom_sync_context_handler = None
custom_grad_sync_func = None
if self.with_distributed_adam:
if self.megatron_amp_o2:
# copy grads to main grad
custom_sync_context_handler = lambda: self._optimizer.no_sync(greedy_grad_copy=True)
else:
# keep grad tensors around
custom_sync_context_handler = lambda: self._optimizer.no_sync(greedy_grad_copy=False)
custom_grad_sync_func = self.reduce_overlap_gradients
else:
if self.megatron_amp_o2 and not self.cfg.get('sequence_parallel', False):
custom_sync_context_handler = self._optimizer.no_sync
Expand All @@ -309,6 +312,7 @@ def training_step(self, batch, batch_idx):
dtype=self.autocast_dtype,
grad_scaler=self.trainer.precision_plugin.scaler if self.cfg.precision == 16 else None,
custom_sync_context_handler=custom_sync_context_handler,
custom_grad_sync_func=custom_grad_sync_func,
sequence_parallel_enabled=self.cfg.get('sequence_parallel', False),
sync_batch_comm=self.cfg.get('sync_batch_comm', True),
num_micro_batches_with_partial_activation_checkpoints=self.cfg.get(
Expand All @@ -330,11 +334,8 @@ def training_step(self, batch, batch_idx):
self.allreduce_sequence_parallel_gradients()

if self.with_distributed_adam:
# launch grad reductions
# Note: grads in first pipeline stage have already been
# reduced
if not parallel_state.is_pipeline_first_stage():
self.reduce_overlap_gradients()
# gradients are reduced internally in distributed optimizer
pass
elif self.megatron_amp_o2:
# when using pipeline parallelism grads must be all-reduced after the pipeline (not asynchronously)
if self.cfg.get('pipeline_model_parallel_size', 1) > 1 or self.cfg.get('sequence_parallel', False):
Expand Down
11 changes: 6 additions & 5 deletions 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 @@ -308,13 +308,16 @@ def training_step(self, batch, batch_idx):
tensor_shape = [encoder_seq_length, get_micro_batch_size(), self.cfg.encoder.hidden_size]

# handle asynchronous grad reduction
custom_sync_context_handler = None
custom_grad_sync_func = None
if self.with_distributed_adam:
if self.megatron_amp_o2:
# copy grads to main grad
custom_sync_context_handler = lambda: self._optimizer.no_sync(greedy_grad_copy=True)
else:
# keep grad tensors around
custom_sync_context_handler = lambda: self._optimizer.no_sync(greedy_grad_copy=False)
custom_grad_sync_func = self.reduce_overlap_gradients
else:
if (
self.megatron_amp_o2
Expand All @@ -339,6 +342,7 @@ def training_step(self, batch, batch_idx):
dtype=self.autocast_dtype,
grad_scaler=self.trainer.precision_plugin.scaler if self.cfg.precision == 16 else None,
custom_sync_context_handler=custom_sync_context_handler,
custom_grad_sync_func=custom_grad_sync_func,
)
else:
losses_reduced_per_micro_batch = forward_backward_no_pipelining(
Expand All @@ -363,11 +367,8 @@ def training_step(self, batch, batch_idx):
loss_mean = torch.tensor(0.0).cuda()

if self.with_distributed_adam:
# launch grad reductions
# Note: grads in first pipeline stage have already been
# reduced
if not parallel_state.is_pipeline_first_stage():
self.reduce_overlap_gradients()
# gradients are reduced internally in distributed optimizer
pass
elif self.megatron_amp_o2:
# when using pipeline parallelism grads must be reduced after the pipeline (not asynchronously)
if self.cfg.get('pipeline_model_parallel_size', 1) > 1:
Expand Down