fix(megatron): reduce reranker evaluation metrics across DP ranks - #10088
Open
Excelius-Wang wants to merge 1 commit into
Open
fix(megatron): reduce reranker evaluation metrics across DP ranks#10088Excelius-Wang wants to merge 1 commit into
Excelius-Wang wants to merge 1 commit into
Conversation
4 tasks
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Outdated suggestions cannot be applied.This suggestion has been applied or marked resolved.Suggestions cannot be applied from pending reviews.Suggestions cannot be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.
Megatron reranker validation computes MRR/NDCG from each data-parallel rank's local shard, so checkpoint selection can use a shard-dependent metric. Average the existing per-query scores across the DP group with the existing MeanMetric sum/count reduction. Weight only valid queries, including when some ranks have no valid queries.
Keep local/HF metric behavior and query-ranking calculations unchanged. Communicate two scalars per metric at evaluation completion; do not gather predictions or alter embedding/InfoNCE evaluation.
Validation: CPU unit tests and four-process Gloo tests covering unequal query counts, independent subgroups, empty/invalid shards, repeated compute/reset, DP=1 and the local/HF path. 300 comparisons match the upstream local implementation exactly, including tied logits and singleton queries. Pre-commit checks pass. Full Megatron/NCCL training was not run.