Skip to content

Transfer V1 completion logprobs without copying #1804

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
xeophon merged 1 commit into from
Jun 21, 2026
Merged

Transfer V1 completion logprobs without copying #1804

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: 2 additions & 1 deletion verifiers/v1/graph.py
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -478,7 +478,8 @@ def _commit_turn(turn: PendingTurn, response: Response) -> None:
sampled=True,
token_ids=[*gen_prompt, *comp_ids],
mask=[False] * len(gen_prompt) + [True] * len(comp_ids),
logprobs=list(tokens.completion_logprobs) if tokens else [],
# TurnTokens is discarded after commit, so transfer its logprobs without copying.
logprobs=tokens.completion_logprobs if tokens else [],

@chatgpt-codex-connector chatgpt-codex-connector Bot Jun 21, 2026

Copy link
Copy Markdown

Choose a reason for hiding this comment

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

P2 Badge Detach logprobs before committing the trace

When a caller keeps or reuses the Response/TurnTokens object after commit(), this stores the same mutable completion_logprobs list on the long-lived trace. Any later normalization, clearing, or buffer reuse by a custom train client mutates trace.nodes[-1].logprobs, and Branch.logprobs/trace dumps will then report corrupted training logprobs; the previous list(...) copy isolated committed graph state from the transient response object.

Useful? React with 👍 / 👎.

finish_reason=response.finish_reason,
usage=response.usage,
)
Expand Down
Loading