Skip to content

fix: use shift_labels for metrics when using CP or SP #4579

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
kashif merged 2 commits into from
Nov 26, 2025
+3 −2
Merged

fix: use shift_labels for metrics when using CP or SP #4579

Changes from all commits
Commits
Show all changes
2 commits
Select commit Hold shift + click to select a range
44ab7ac
fix: use shift_labels for metrics when using CP or SP
jue-jue-zi Nov 26, 2025
f3b17c0
Merge branch 'main' into main
kashif Nov 26, 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
5 changes: 3 additions & 2 deletions trl/trainer/sft_trainer.py
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -1118,7 +1118,8 @@ def compute_loss(

# Set aside labels as it will be dropped by super().compute_loss() if a custom `compute_loss_func` is used.
# This can be removed when this issue is fixed.
labels = inputs["labels"]
# When using CP or SP, labels are pre-shifted, we must use shift_labels instead.
labels = inputs["labels"] if "shift_labels" not in inputs else None

# If not set, defaults from model config and may warn since cache isn't compatible with gradient checkpointing
inputs["use_cache"] = False
Expand Down Expand Up @@ -1172,7 +1173,7 @@ def compute_loss(
# Compute accuracy from logits using argmax (traditional method)
with torch.no_grad():
if "shift_labels" in inputs:
# When using CP, labels are pre-shifted. We must use these (and cannot manually shift) because:
# When using CP or SP, labels are pre-shifted. We must use these (and cannot manually shift) because:
# - The first discarded token from inputs["labels"] actually belongs to process n-1
# - The last logits require the label from process n+1
shift_logits = outputs.logits.contiguous()
Expand Down
Loading