Compute cold-start boost target after MoE suppression - #87
Open
teddytennant wants to merge 1 commit into
Open
Conversation
cold_start_target ranked the pre-suppression scores while apply_cold_start worked on the post-suppression vector, so the boost target could be the score of a candidate that had already been zeroed out of contention.
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Outdated suggestions cannot be applied.This suggestion has been applied or marked resolved.Suggestions cannot be applied from pending reviews.Suggestions cannot be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.
AuthorColdStart::applyzeroes suppressed Phoenix MoE candidates intoeffective, then takes the boost target fromcold_start_target(query, scores), the pre-suppression vector. With the shippedcold_start_slot_min/cold_start_slot_maxof 15/16 the target isranked[15]. For a holdout viewer every MoE candidate is zeroed, so that slot belongs to a post which will not be served and the cold-started post is boosted above the band.apply_cold_startalready derivespositionsandmax_cold_start_slotfromeffective, so the two halves of one decision were reading different vectors. Every existing test setsslot_maxto 1, which makes the target the global maximum and hides the difference.home-mixer has no Cargo.toml in this tree, so I transcribed
positions_among_nonzero,apply_moe_ranking_policyandcold_start_targetinto a standalone binary with the shipped defaults: 40 candidates, a quarter of them MoE, scores from 100.0 stepping down by 2.5. The target comes out 62.5 against 50.0 for the score actually at served slot 15, landing the post at served rank 11. The added test is the same arithmetic on five candidates and returns 80.0 before the change, 70.0 after.