Repository navigation
Run a hosted review's batches concurrently - #159
Merged
Merged
Conversation
Batch requests carry their own prompts and are collected by index, so nothing orders them, and `max_hosted_review_batches` already prices a review's capacity at the concurrent wave count. Execution ran them one at a time outside the deterministic large-review path, which divided the generator phase by the batch count: a four-batch review gave each batch a quarter of the phase, under the review model's tail latency, and lost a batch to a timeout more often than not. Admission and execution now take the same concurrency, so the waves the schedule is checked against are the waves that run.
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
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Outdated suggestions cannot be applied.This suggestion has been applied or marked resolved.Suggestions cannot be applied from pending reviews.Suggestions cannot be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 requests carry their own prompts and are collected by index, so nothing orders them, and
max_hosted_review_batchesalready prices a review's capacity at the concurrent wave count. Execution ran them one at a time outside the deterministic large-review path, which divided the generator phase by the batch count: a four-batch review gave each batch a quarter of the phase, below the review model's p95, and lost a batch to a timeout more often than not.Admission and execution now take the same concurrency, so the waves the schedule is checked against are the waves that run, and a review that fits in one wave gets the phase rather than a fraction of it.