Add final-assistant SFT mode#743
Merged
Merged
Conversation
* origin/main: feat: unify shared-prefix Megatron execution (#739) Add DeepSeek V4 Flash support for ART Megatron (#745) Add ART Megatron support for GPT OSS (#744) ci: bound GPU prewarm cleanup ci: prewarm GPU image on all GPU clusters ci: dispatch Caladan GPU image builds Fix Gemma4 MoE Triton flex stage config (#742) fix: split GPU image dependency warmup # Conflicts: # src/art/dev/model.py # src/art/serverless/backend.py
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Outdated suggestions cannot be applied.This suggestion has been applied or marked resolved.Suggestions cannot be applied from pending reviews.Suggestions cannot be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.
SFT currently trains every assistant turn in a conversation. Some datasets need earlier turns only as context and should apply loss only to the final assistant response.
This PR adds
assistant_turns="last"while keeping"all"as the default. ART separates the final turn at the rendered prompt/completion boundary instead of finding it inside a one-shot token sequence. This keeps the full conversation as context and matters because chat templates can depend on that conversation while tokenization can merge across the generation boundary. If a template does not expose a safe boundary, ART fails instead of training unintended tokens.