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

Add support for overlapped gradient and parameter synchronization for GPT SFT model #10041

Merged
merged 3 commits into from
Aug 6, 2024
Merged

Add support for overlapped gradient and parameter synchronization for GPT SFT model #10041

Changes from all commits
Commits
Show all changes
3 commits
Select commit Hold shift + click to select a range
f5a7259
Add support for overlapped gradient and parameter synchronization for…
michal2409 Aug 5, 2024
4f79c60
Add finalize_model_grads
michal2409 Aug 5, 2024
97282ff
Apply isort and black reformatting
michal2409 Aug 5, 2024
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
17 changes: 17 additions & 0 deletions nemo/collections/nlp/models/language_modeling/megatron_gpt_sft_model.py
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -40,6 +40,7 @@

try:
from megatron.core import parallel_state
from megatron.core.distributed import finalize_model_grads
from megatron.core.pipeline_parallel.schedules import get_forward_backward_func

HAVE_MEGATRON_CORE = True
Expand Down Expand Up @@ -378,11 +379,27 @@ def fwd_bwd_step(self, dataloader_iter, forward_only, first_val_step=None):
)
grad_sync_func = self.reduce_overlap_gradients
param_sync_func = self.sync_overlap_parameters
elif not forward_only and self.use_mcore_dist_optim:
if self.cfg.optim.get("overlap_grad_sync", False):
no_sync_func = [model_chunk.no_sync for model_chunk in self.model]
no_sync_func = no_sync_func[0] if len(self.model) == 1 else no_sync_func

if self.cfg.optim.get("delay_grad_reduce", True):
grad_sync_func = [model_chunk.start_grad_sync for model_chunk in self.model]
grad_sync_func = grad_sync_func[0] if len(self.model) == 1 else grad_sync_func
if self.cfg.optim.get("overlap_param_sync", False) and self.cfg.optim.get("delay_param_gather", False):
param_sync_func = [
lambda x, model_index=model_index: self._optimizer.finish_param_sync(model_index, x)
for model_index in range(len(self.model))
]
param_sync_func = param_sync_func[0] if len(self.model) == 1 else param_sync_func

for module in self.get_model_module_list():
module.config.no_sync_func = no_sync_func
module.config.grad_sync_func = grad_sync_func
module.config.param_sync_func = param_sync_func
if self.use_mcore_dist_optim:
module.config.finalize_model_grads_func = finalize_model_grads

fwd_bwd_function = get_forward_backward_func()

Expand Down
Loading