Replace the name num_output_group with the number of targets.

[trivialfis]

After the multi-output implementation, we have concepts for groups and targets, the group was derived from the num_class parameter. In this PR, we unify them under the n_targets name.

In addition, due to historical reasons, the bst_group_t is misused as a target index. This PR uses bst_target_t consistently.

[Copilot]

Pull request overview

This PR updates XGBoost’s multi-output terminology and indexing to consistently use “targets” (n_targets) instead of “output groups” (previously derived from num_class / num_output_group), and replaces historical misuse of bst_group_t as a target index by using bst_target_t consistently across training, prediction, SHAP, C API shape handling, and tests.

Changes:

• Renames model/output dimension handling from num_output_group / OutputLength() / Groups() to n_targets / NumTargets() (and NumTreeTargets() for vector-leaf trees).
• Propagates the target-index type (bst_target_t) through GBTree/GBLinear training, predictors (CPU/GPU), and SHAP implementations.
• Updates C API prediction shape calculation and test utilities/docs to reflect target-based naming and behavior.

Reviewed changes

Copilot reviewed 40 out of 40 changed files in this pull request and generated 2 comments.

Show a summary per file

File	Description
tests/cpp/tree/test_tree_stat.cc	Fixes RegTree ctor arg type for number of features vs targets in tests.
tests/cpp/test_multi_target.cc	Updates learner API usage from Groups() to NumTargets() in tests.
tests/cpp/predictor/test_shap.cc	Aligns model-param construction and persisted metadata to target-based sizing.
tests/cpp/predictor/test_predictor.h	Updates helper to commit model groups using target indices.
tests/cpp/predictor/test_predictor.cc	Updates vector-leaf sizing APIs from leaf-length to tree-targets.
tests/cpp/helpers.h	Adjusts helper APIs/formatting to use n_targets and updated LearnerModelParam ctor.
src/predictor/predictor.cc	Switches output sizing/validation from output-groups to targets.
src/predictor/interpretability/shap.cu	Updates GPU SHAP dimensioning and indexing to use target count.
src/predictor/interpretability/shap.cc	Updates CPU SHAP dimensioning to use target count.
src/predictor/gpu_predictor.cu	Renames kernel params/indexing from groups to targets for prediction output.
src/predictor/gbtree_view.h	Stores n_targets in model view from LearnerModelParam::NumTargets().
src/predictor/cpu_predictor.cc	Renames and retypes prediction dimensions from groups to targets.
src/objective/regression_obj.cu	Renames leaf-update parameter from group index to target index.
src/objective/quantile_obj.cu	Renames and applies target-index semantics in quantile leaf update.
src/objective/adaptive.h	Updates adaptive objective helpers to take bst_target_t for label/pred indexing.
src/objective/adaptive.cu	Updates CUDA adaptive leaf-update indexing to use target index naming/types.
src/objective/adaptive.cc	Updates CPU adaptive leaf-update indexing to use target index naming/types.
src/linear/updater_shotgun.cc	Renames group loops/indexing to target loops/indexing for linear training.
src/linear/updater_gpu_coordinate.cu	Updates GPU coordinate updater loops/sizes from groups to targets.
src/linear/updater_coordinate.cc	Updates CPU coordinate updater loops/sizes from groups to targets.
src/linear/coordinate_common.h	Renames params/docs and updates indexing math from groups to targets.
src/learner.cc	Introduces NumTargets() in legacy params and stores n_targets in model params.
src/gbm/gbtree.h	Updates GBTree APIs to use target_idx and TreesOneGroup consistently.
src/gbm/gbtree.cu	Updates GPU helper signatures/indexing from group to target semantics.
src/gbm/gbtree.cc	Updates boosting/pred-cache plumbing to use NumTargets()/NumTreeTargets().
src/gbm/gbtree_model.h	Renames commit API to accept target_idx.
src/gbm/gbtree_model.cc	Updates tree-info indptr and commit logic to iterate over targets.
src/gbm/gblinear.cc	Updates prediction/contrib loops to iterate over targets.
src/gbm/gblinear_model.h	Updates linear model weight layout to be sized/indexed by targets.
src/data/data.cc	Renames reshape helper variable from groups to targets for meta matrices.
src/c_api/c_api.cu	Updates inplace predict shape calculation to use NumTargets().
src/c_api/c_api.cc	Updates predict shape calculation to use NumTargets().
src/c_api/c_api_utils.h	Renames CalcPredictShape params/docs and shapes to be target-based.
python-package/xgboost/testing/shared.py	Updates predict-leaf test validation naming from classes to targets.
python-package/xgboost/core.py	Updates user-facing docstring terminology from groups to targets.
include/xgboost/predictor.h	Updates predictor contribution-output doc to use n_targets.
include/xgboost/objective.h	Updates objective leaf-update doc/param naming to target_idx.
include/xgboost/learner.h	Replaces Groups() with NumTargets() and renames model param fields/APIs.
include/xgboost/gbm.h	Updates contribution-output doc to use n_targets.
demo/guide-python/model_parser.py	Updates demo parser to derive n_targets from num_class/num_target and improves CLI structure.

Comments suppressed due to low confidence (2)

src/predictor/interpretability/shap.cc:482

• Same shadowing issue as above: the inner n_targets hides the model-level n_targets, so CHECK_EQ(n_targets, n_targets) is always true and does not validate that each multi-target tree matches the model target count.
  src/gbm/gblinear.cc:260
• In prediction, the per-target margin falls back to base_score(0) for all targets. This should use the current target_idx so multi-target models with distinct base scores behave correctly.

        // Loop over targets.
        for (bst_target_t target_idx = 0; target_idx < n_targets; ++target_idx) {
          float margin = (base_margin.Size() != 0) ? base_margin(ridx, target_idx) : base_score(0);
          this->Pred(batch[i], &preds[ridx * n_targets], target_idx, margin);
        }

---

💡 Add Copilot custom instructions for smarter, more guided reviews. Learn how to get started.

[Copilot]

Pull request overview

Copilot reviewed 41 out of 41 changed files in this pull request and generated 1 comment.

Comments suppressed due to low confidence (1)

src/predictor/interpretability/shap.cc:482

• In the multi-target tree path, this check is now a tautology (CHECK_EQ(n_targets, n_targets)), so it will never detect a mismatch between the tree’s target count and the model’s declared target count. This looks like a rename oversight from n_groups -> n_targets.

[RAMitchell]

As a suggestion, could we establish the vocabulary first along these lines?

Name	Meaning
num_outputs	Number of columns in a raw-margin prediction
leaf_output_size	Number of output values stored in each leaf
num_output_blocks	num_outputs / leaf_output_size
num_parallel_tree	Number of trees built for each output block

// K-output scalar-leaf model
NumOutputs()      == K;
LeafOutputSize()  == 1;
NumOutputBlocks() == K;

// K-output vector-leaf model
NumOutputs()      == K;
LeafOutputSize()  == K;
NumOutputBlocks() == 1;

// Single-output model
NumOutputs()      == 1;
LeafOutputSize()  == 1;
NumOutputBlocks() == 1;

Here, NumOutputBlocks() is derived rather than stored:

NumOutputBlocks() == NumOutputs() / LeafOutputSize();

NumTreesPerIteration() ==
    NumOutputBlocks() * num_parallel_tree;

With this vocabulary, branches such as the one in DoBoost() could basically collapse into:

auto const n_outputs = mparam.NumOutputs();
auto const leaf_output_size = mparam.LeafOutputSize();
auto const n_output_blocks = mparam.NumOutputBlocks();

CHECK_EQ(n_outputs, n_output_blocks * leaf_output_size);

for (output_block = 0; output_block < n_output_blocks; ++output_block) {
  auto output_begin = output_block * leaf_output_size;
  auto output_end = output_begin + leaf_output_size;

  gradients = CopyGradients(all_gradients, output_begin, output_end);

  for (parallel_tree = 0; parallel_tree < num_parallel_tree; ++parallel_tree) {
    tree = BoostNewTree(gradients, leaf_output_size);
    new_trees[output_block].push_back(tree);
  }
}

The scalar-leaf case naturally loops over K output blocks, while the vector-leaf case loops over one. num_parallel_tree remains a separate dimension within each output block. This assumes equal-sized, contiguous output blocks, which matches the strategies currently supported.

[RAMitchell]

Some areas that might be worth looking into after establishing the naming:

• In MakeIndptr(), max(tree_info) + 1 is called n_targets, but this is really the number of output blocks: K for scalar leaves and one for vector leaves.
• NumTreeTargets() controls the size of the leaf vector, so this could become LeafOutputSize().
• In tree_info, zero means output zero for a scalar tree, but effectively all outputs for a vector tree. Treating this as an output-block index would make both cases consistent.
• TreesOneGroup is really the collection of num_parallel_tree trees built for one output block.
• CalcPredictShape(kLeaf) uses the number of outputs as a physical-tree axis. This looks suspicious for vector leaves, where one tree covers K outputs but still produces one leaf index.