Open
Conversation
Collaborator
|
See if you can get the tests to run. If not, move to a branch on the repo instead of your fork. |
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Outdated suggestions cannot be applied.This suggestion has been applied or marked resolved.Suggestions cannot be applied from pending reviews.Suggestions cannot be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.
This PR adds the initial end-to-end HPC/DDP implementation for the sklearn-style interface.
Before opening the PR, I ran scaling-oriented validation on the new distributed path to confirm the training loop was stable under multi-rank execution and that performance characteristics (epoch time / throughput / step time) behaved as expected as world size increased. Those checks focused on fit and predict workloads and included basic correctness signals.
After those initial runs, I discovered an implementation mistake introduced during subsequent development on the DDP prediction/aggregation. The reason this was not obvious from the earlier scaling results is that DDP training stability and scaling can look correct even when the predict path is missing the required cross-rank aggregation details (synchronizing outputs, restoring global row ordering, and deduplicating sampler padding). I have made the necessary corrections in this branch, and am redoing the benchmark and perturbation tests with the updated package.