Repository navigation
fix: reuse recorded tokens by default and preserve assistant content - #977
Draft
bradhilton wants to merge 68 commits into
Draft
bradhilton wants to merge 68 commits into
bradhilton wants to merge 68 commits into
Conversation
bradhilton
marked this pull request as ready for review
September 25, 2026 20:17
This was referenced Sep 26, 2026
bradhilton
marked this pull request as draft
September 26, 2026 15:46
bradhilton
changed the title
This branch has not been deployed
This file contains hidden or bidirectional Unicode text that may be interpreted or compiled differently than what appears below. To review, open the file in an editor that reveals hidden Unicode characters.
Learn more about bidirectional Unicode characters
Sign up for free
to join this conversation on GitHub.
Already have an account?
Sign in to comment
1 participant
Add this suggestion to a batch that can be applied as a single commit.This suggestion is invalid because no changes were made to the code.Suggestions cannot be applied while the pull request is closed.Suggestions cannot be applied while viewing a subset of changes.Only one suggestion per line can be applied in a batch.Add this suggestion to a batch that can be applied as a single commit.Applying suggestions on deleted lines is not supported.You must change the existing code in this line in order to create a valid suggestion.Outdated suggestions cannot be applied.This suggestion has been applied or marked resolved.Suggestions cannot be applied from pending reviews.Suggestions cannot be applied on multi-line comments.Suggestions cannot be applied while the pull request is queued to merge.Suggestion cannot be applied right now. Please check back later.
Agent: Hayek
Use complete, proven recorded tokens and logprobs as the default training source, including length-stopped output. Preserve ownership and masks. Missing or edited records use rendering and validation; ambiguous records still refuse. The merged parser fix in #1009 preserves literal assistant content for the standard template with thinking disabled; unsupported custom finalizers and template globals remain outside that guarantee.
Validation: independent reviews and CI passed, including 3,386 unit and 1,338 Megatron tests, plus parser controls. Mutable inputs, callbacks, and error identity are covered.
Draft: performance remains unresolved. A fixture takes 1.626 seconds versus 1.663 for its predecessor and 0.886 for the older frozen implementation; repeated context work can grow superlinearly. Snapshot byte bounds also do not bound referenced graphs. No experiment adoption or live throughput claim is included.