[mt] Split up gradient types for the GPU hist. #11798
Merged
+1,006
−304
Conversation
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
- Use reduced gradient for tree structure exploration. The PR adds a gradient container that has two different gradient types, one for tree split and the other one for leaf values. This is an optimization for vector-leaf to reduce the overhead of finding tree structure. work on build hist. notes. work on evaluation. build all nodes. inputs. getter. alloc. proto. apply. disable. Work on high-level tests. lint. cleanup. test policy. lazy gen. Fix scan. Cleanup. Cleanup single split. Revert. Remove sync. check. cleanup. notes. error message. Cleanup grp type, notes. notes. leaf sum. leaf weight. cleanup. work on the path. update. update. sth to run. parameters. sort. probing. change n targets. Expand tree. q set leaves. print. fix. nidx. fix. copy root sum. copy root sum. Fix sum. work on a simple python test. note. sort leaves. cleanup. test. cleanup. comment. cleanup. check. note. try to find an interface. notes. remove. check. clenaup. Add debugging utilities. set root. allow smaller weight. Cleanup. Cleanup. cleanup. Doc, cleanup. Remove allocations. move container. unify the do boost. Remove the update method.
Contributor
There was a problem hiding this comment.
Pull Request Overview
This PR introduces support for reduced gradients in multi-target tree construction, enabling different gradients for split evaluation versus leaf value calculation. This is part of an experimental feature for improving multi-target learning.
Key changes:
- Introduces a new
GradientContainerstruct to hold both split gradients and optional value gradients - Adds experimental
TreeObjectivePython class withsplit_gradmethod for custom gradient reduction - Refactors tree updaters and booster code to use
GradientContainerinstead of raw gradient matrices - Renames
SetLeaftoSetRootfor multi-target trees to clarify semantics
Reviewed Changes
Copilot reviewed 54 out of 54 changed files in this pull request and generated 2 comments.
Show a summary per file
| File | Description |
|---|---|
| include/xgboost/gradient.h | New header defining GradientContainer struct for holding split and value gradients |
| include/xgboost/tree_updater.h | Updates Update signature to accept GradientContainer* instead of gradient matrix |
| include/xgboost/tree_model.h | Renames SetLeaf to SetRoot and adds SetLeaves method for batch leaf updates |
| include/xgboost/multi_target_tree_model.h | Updates multi-target tree API to use SetRoot and adds SetLeaves |
| include/xgboost/learner.h | Updates BoostOneIter signature to accept GradientContainer |
| include/xgboost/gbm.h | Updates DoBoost signature to accept GradientContainer |
| include/xgboost/objective.h | Changes parameter type from std::int32_t to bst_target_t for consistency |
| include/xgboost/linalg.h | Adds EmptyLike utility function for tensor creation |
| src/tree/updater_gpu_hist.cuh | Implements reduced gradient support with separate split/value quantizers |
| src/tree/updater_gpu_hist.cu | Updates GPU histogram builder to handle gradient container |
| src/tree/leaf_sum.cuh | New file implementing leaf gradient sum calculations for GPU |
| src/tree/leaf_sum.cu | Implementation of leaf weight calculation from value gradients |
| src/tree/multi_target_tree_model.cc | Refactors weight setting with SetRoot and implements SetLeaves |
| src/tree/tree_model.cc | Adds SetLeaves wrapper method |
| src/tree/updater_*.cc | Updates all tree updaters to accept GradientContainer |
| src/gbm/gbtree.cc | Refactors boosting logic to handle gradient containers |
| src/gbm/gblinear.cc | Updates linear booster to use gradient container |
| src/learner.cc | Updates learner to use gradient containers throughout |
| src/c_api/c_api.cc | Adds experimental XGBoosterTrainOneIterWithObj API and renames function |
| src/c_api/c_api.cu | Renames CopyGradientFromCUDAArrays to CopyGradientFromCudaArrays |
| src/common/device_helpers.cuh | Removes deprecated DebugSyncDevice function |
| src/common/device_debug.cuh | New file with debug utilities moved from device_helpers |
| src/common/algorithm.h | Removes incorrect GPU check from ArgSort |
| python-package/xgboost/objective.py | New module with experimental objective interface |
| python-package/xgboost/core.py | Implements support for tree objectives with split gradients |
| python-package/xgboost/testing/multi_target.py | Adds comprehensive tests for reduced gradient feature |
| python-package/xgboost/testing/init.py | Fixes ls_obj to handle multi-dimensional arrays |
| tests/python-gpu/test_gpu_multi_target.py | Adds test for reduced gradient on GPU |
| tests/cpp/**/test_*.cc | Updates test files to use GradientContainer |
💡 Add Copilot custom instructions for smarter, more guided reviews. Learn how to get started.
Contributor
There was a problem hiding this comment.
Pull Request Overview
Copilot reviewed 56 out of 56 changed files in this pull request and generated 9 comments.
💡 Add Copilot custom instructions for smarter, more guided reviews. Learn how to get started.
Contributor
There was a problem hiding this comment.
Pull Request Overview
Copilot reviewed 56 out of 56 changed files in this pull request and generated 5 comments.
💡 Add Copilot custom instructions for smarter, more guided reviews. Learn how to get started.
Contributor
There was a problem hiding this comment.
Pull Request Overview
Copilot reviewed 60 out of 60 changed files in this pull request and generated 4 comments.
💡 Add Copilot custom instructions for smarter, more guided reviews. Learn how to get started.
trivialfis
changed the title
Member
Author
|
cc @rongou |
rongou
approved these changes
Nov 10, 2025
Sign up for free
to join this conversation on GitHub.
Already have an account?
Sign in to comment
2 participants
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.
The PR adds a gradient container with two different gradient types: one for tree splits and the other for leaf values. This is an optimization for vector-leaf to reduce the overhead of finding the tree structure. Currently, the interface is exposed through a custom objective function by creating a specialized tree objective.
An alternative approach would be to decouple the dimension reduction function from the objective function by introducing an additional parameter, such as
grad_reducer. This is actually the first approach I tried.I don't have a strong opinion on which one we should choose. The objective approach facilitates associating the reduced gradient with the prediction and label more easily, but may be less modularized.
ref #9043