transformers
fix(trainer): Correct loss scaling for incomplete gradient accumulation steps
#39659
Merged
Merged
fix(trainer): Correct loss scaling for incomplete gradient accumulation steps #39659
SunMarc
merged
5 commits into
huggingface:main
from
fix-issue-38837
Fix issue[#38837]: wrong loss scaled in last step of epoch
701110c9
chore: trigger CI
ae0f42a3
Update src/transformers/trainer.py
b6b4b59d
Update src/transformers/modeling_flash_attention_utils.py
d16dc5b1
Merge branch 'main' into fix-issue-38837
da250646
qgallouedec
approved these changes
on 2025-07-25
SunMarc
approved these changes
on 2025-07-29
Login to write a write a comment.
Login via GitHub