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

disable moe logging to avoid deepseek hang #12168

Merged
merged 4 commits into from
Feb 14, 2025
Merged

disable moe logging to avoid deepseek hang #12168

Changes from 3 commits
Commits
Show all changes
4 commits
Select commit Hold shift + click to select a range
c4d7200
disable moe logging to avoid deepseek hang
akoumpa Feb 13, 2025
172d67c
disable moe logging to avoid deepseek hang
akoumpa Feb 13, 2025
0fe69bd
Apply isort and black reformatting
akoumpa Feb 13, 2025
31a13f9
Update megatron_strategy.py
akoumpa Feb 14, 2025
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: 4 additions & 3 deletions nemo/lightning/pytorch/strategies/megatron_strategy.py
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -60,7 +60,7 @@

from nemo.core.optim.mcore_optim import McoreDistributedOptimizer
from nemo.lightning import _strategy_lib, io
from nemo.lightning.megatron_parallel import CallbackConnector, MegatronParallel, aggregate_moe_loss_stats
from nemo.lightning.megatron_parallel import CallbackConnector, MegatronParallel
from nemo.lightning.pytorch.callbacks import ModelTransform
from nemo.lightning.pytorch.strategies.utils import (
RestoreConfig,
Expand Down Expand Up @@ -628,9 +628,10 @@ def training_step(self, dataloader_iter, *args: Any, **kwargs: Any) -> STEP_OUTP
"reduced_train_loss", reduced_train_loss, prog_bar=True, batch_size=1, sync_dist=False
)
# Log any MoE losses.
# @akoumparouli: disabling this as it hangs with deepseek.
# TODO(@akoumparouli): loss_scale depends on the GBS.
for loss_name, loss_value in aggregate_moe_loss_stats(loss_scale=1.0).items():
self.lightning_module.log(loss_name, loss_value, prog_bar=True, rank_zero_only=True, batch_size=1)
# for loss_name, loss_value in aggregate_moe_loss_stats(loss_scale=1.0).items():
# self.lightning_module.log(loss_name, loss_value, prog_bar=True, rank_zero_only=True, batch_size=1)

return out

Expand Down
Loading