Skip to content

Fix SFTDataset truncation bug #3158

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
asolergi-nv merged 3 commits into from
Jan 30, 2026
Merged

Fix SFTDataset truncation bug #3158

Changes from all commits
Commits
Show all changes
3 commits
Select commit Hold shift + click to select a range
87050a9
Fix SFTDataset truncation bug
duncanriach Jan 30, 2026
711efcf
Small tweak
duncanriach Jan 30, 2026
618d817
Tweak back
duncanriach Jan 30, 2026
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
19 changes: 6 additions & 13 deletions megatron/training/datasets/sft_dataset.py
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -142,19 +142,12 @@ def extend_with_padding(tokens, targets, positions, pad_len):

# Handle any necessary truncation
if len(pack_tokens) >= pack_length + 1: # +1 here to account for later alignment
truncate_left_not_right = True # TODO(duncan): plumb this switch in
if truncate_left_not_right: # Retain existing eod
max_body = pack_length
pack_tokens = pack_tokens[-max_body:]
pack_targets = pack_targets[-max_body:]
pack_tokens.append(pad)
pack_targets.append(pad)
else: # Truncate right (need to add eod)
max_body = pack_length - 1
pack_tokens = pack_tokens[:max_body]
pack_targets = pack_targets[:max_body]
pack_tokens.extend([eod, pad])
pack_targets.extend([eod, pad])
# Truncate on the right
max_body = pack_length
pack_tokens = pack_tokens[:max_body]
pack_targets = pack_targets[:max_body]
pack_tokens.extend(pad)
pack_targets.extend(pad)

@hsiehjackson hsiehjackson Jan 31, 2026 •
edited
Loading

Copy link
Copy Markdown

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

pad is an integer and should be a list as follows

pack_tokens.extend([pad])
pack_targets.extend([pad])

@duncanriach duncanriach Jan 31, 2026

Copy link
Copy Markdown
Contributor Author

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

Thank you!

I'll change it to .append(pad) in another PR

@duncanriach duncanriach Jan 31, 2026

Copy link
Copy Markdown
Contributor Author

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

@hsiehjackson, I opened 3185. GitHub is not allowing me to select you as a reviewer. I don't know why. Please review that.

pack_positions = pack_positions[:pack_length+1]
# Note len({pack_tokens, pack_targets, pack_positions}) should be pack_length + 1
cu_seqlens[-1] = len(pack_tokens) - 1
Expand Down
Loading