PRF Mitigation heuristics for HGB scalability - #34935
Draft
cakedev0 wants to merge 34 commits into
Draft
Conversation
Non regression test for the OverflowError fixed in 68e1aae: find_best_split only populated n_subsampled_features entries of the split_infos buffer, but scanned all n_allowed_features when picking the best split, reading uninitialized memory whenever max_features < 1. Co-Authored-By: Claude Sonnet 5 <[email protected]>
…positive_heuristics
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Outdated suggestions cannot be applied.This suggestion has been applied or marked resolved.Suggestions cannot be applied from pending reviews.Suggestions cannot be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.
Finally, I have something that has a uniformly positive impact on the 8 scenarios I'm testing:
(x Libgomp & libomp for each of those cases)
And that can be very/hugely beneficial in some scenarios (2-10x speed-ups):
Reference Issues/PRs
Implements item 4 partially, item 5 and item 6 of #34764 (comment)
And I'd say it also closes #14306. Though this issues does have some interesting idea that are not implemented here, especially parallelism over block of samples x features for the histogram building. Still, benchmarks show HGB does continue to scale until 16-32 threads for many medium/big datasets (as long as active wait is enabled). I personally would consider that good enough ^^
What does this implement/fix? Explain your changes.
OMP_VERBOSE=true. Claude found a way to make this fast even if it has to dosubprocess.runAI usage disclosure
Mostly not, except for the active wait detection part.
Benchmarks
WIP
https://pr-81.sklbench-dashboard-preview.pages.dev/