[eval] Batch rolling likelihood across documents - #96
Open
Helw150 wants to merge 1 commit into
Open
Conversation
Batch API rolling-likelihood windows across documents instead of waiting for each document before submitting the next. Paloma and UncheatableEval usually produce one window per document, so stock lm-eval serialized their endpoint calls even when num_concurrent was greater than one. Flush four windows per concurrency slot to keep the endpoint full while bounding evaluator memory. Preserve per-document sums and cache entries.
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Outdated suggestions cannot be applied.This suggestion has been applied or marked resolved.Suggestions cannot be applied from pending reviews.Suggestions cannot be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.
Batch API rolling-likelihood windows across documents before sending them to served endpoints. lm-eval v0.4.12 processes documents one at a time, so Paloma and UncheatableEval issue serial requests when most documents fit in one context window. On Nemotron 3 Nano, the old path produced 130–180 completions per minute and one request about every 0.5 seconds despite
num_concurrent=64.Buffer four windows per concurrency slot. This keeps the endpoint busy while bounding evaluator memory. Per-document log-likelihood sums and cache keys stay unchanged.