Fix DiffusionGemma selected-canvas SFT decoding - #777
Open
sylvesterkaczmarek wants to merge 1 commit into
Open
Conversation
Address the multi-canvas cache semantics reported in google-deepmind#773. The SFT encoder moves each cache cursor to the selected canvas start, but the decoder currently submits the full response as queries. For selected canvases after index zero this shifts and wraps K/V writes into later-canvas and prompt slots. Decode exactly canvas_size queries at the selected canvas positions, gather aligned time and self-conditioning tensors, and scatter the selected logits back only to preserve the existing full-response masked loss interface. Add regression coverage for selected query length, positions, future-canvas isolation, and preselected inputs.
sylvesterkaczmarek
marked this pull request as ready for review
August 15, 2026 18:41
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Outdated suggestions cannot be applied.This suggestion has been applied or marked resolved.Suggestions cannot be applied from pending reviews.Suggestions cannot be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.
Address the multi-canvas cache semantics reported in #773.
The SFT encoder moves each cache cursor to the selected canvas start, but the decoder currently submits the full response as queries. For selected canvases after index zero this shifts and wraps K/V writes into later-canvas and prompt slots.
Decode exactly canvas_size queries at the selected canvas positions, gather aligned time and self-conditioning tensors, and scatter the selected logits back only to preserve the existing full-response masked loss interface. Add regression coverage for selected query length, positions, future-canvas isolation, and preselected inputs.