Skip to content

[bugfix] fix logps with PP #89

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
tastelikefeet merged 1 commit into from
Mar 1, 2026
Merged

[bugfix] fix logps with PP #89

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
3 changes: 0 additions & 3 deletions cookbook/megatron/tp.py
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -8,9 +8,6 @@
from twinkle.dataset import Dataset, DatasetMeta
from twinkle.model import MegatronModel
from twinkle.preprocessor import SelfCognitionProcessor
from twinkle.server.tinker.common import input_feature_to_datum
from twinkle.server.tinker.common.compat_base import TwinkleCompatModelBase

# Construct a device_mesh, tp=pp=cp=2, dp=1
device_mesh = DeviceMesh.from_sizes(dp_size=1, tp_size=2, pp_size=2, cp_size=2)
# use torchrun mode
Expand Down
2 changes: 1 addition & 1 deletion src/twinkle/model/megatron/megatron.py
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -481,7 +481,7 @@ def forward_step_func(data_iterator, model):
output_tensor = model(**batch)
batch['labels'] = labels
logps = None
if labels is not None:
if labels is not None and mpu.is_pipeline_last_stage():
loss_mask = (labels != -100).bool()
masked_labels = labels.clone()
masked_labels[~loss_mask] = 0
Expand Down
Loading