Skip to content

[BugFix] Fix mixed penalties batch with async scheduling #27910

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
njhill merged 1 commit into from
Nov 1, 2025
Merged

[BugFix] Fix mixed penalties batch with async scheduling #27910

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
8 changes: 8 additions & 0 deletions vllm/v1/sample/ops/penalties.py
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -21,6 +21,14 @@ def apply_all_penalties(
"""
_, vocab_size = logits.shape
output_tokens_t = _convert_to_tensors(output_token_ids, vocab_size, logits.device)

# In the async scheduling case, rows that won't have penalties applied may contain
# -1 placeholder token ids. We must replace these with valid token ids so that the
# scatter done in apply_penalties is valid.
# NOTE(nick): The penalties implementation is currently quite inefficient and
# will be reworked anyhow.
output_tokens_t.masked_fill_(output_tokens_t == -1, vocab_size)

@hidva hidva Dec 6, 2025 •
edited
Loading

Copy link
Copy Markdown
Contributor

Choose a reason for hiding this comment

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

Hi, I’d like to ask why the actual draft token isn’t used here to replace the placeholder? Thanks.

@njhill

@njhill njhill Dec 6, 2025

Copy link
Copy Markdown
Member Author

Choose a reason for hiding this comment

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

@hidva this line should only apply to requests/rows in the batch which don't require the output tokens (i.e. those which don't use penalties sampling parameters). The other rows should not contain any placeholder tokens at this point.

Also async scheduling + spec decode + penalties isn't yet supported (any help with that appreciated though - see discussion in #30122).


return apply_penalties(
logits,
prompt_token_ids,
Expand Down