Support vector leaf SHAP on GPU

[RAMitchell]

What changed

This extends the exact GPU QuadratureTreeSHAP implementation to support vector-leaf trees for both SHAP contributions and SHAP interaction contributions.

The GPU model compression now stores tree topology once and stores leaf outputs separately. For vector-leaf trees, each compressed leaf records an offset into the leaf-value array, and the kernel reads the target-specific value from that array. The per-target work is represented by lightweight compressed-tree task metadata, so split structure, categories, and branch weights are not duplicated in compressed memory.

The shared C++ SHAP test cases now always include the vector-leaf case, and the CUDA multi-target Python tests include a public API regression for multi_output_tree with pred_contribs=True and pred_interactions=True.

Why

CPU SHAP already supports vector-leaf trees. The GPU predictor still rejected vector-leaf models for exact SHAP contributions and interactions. This removes that unsupported path while preserving the existing exact QuadratureTreeSHAP traversal.

Validation

• git diff --check
• cmake --build build-cuda --target testxgboost -j35
• build-cuda/testxgboost --gtest_filter=GPUPredictor.ShapOutputCasesGPU:GPUPredictor.DartShapOutputGPU:GPUPredictor.CompareCPUShap:Predictor.ShapOutputCasesCPU
• PYTHONPATH=/home/nfs/rorym/xgboost-wt/gpu-vector-leaf-shap/python-package /home/nfs/rorym/anaconda3/bin/conda run -n xgboost python -m pytest tests/python-gpu/test_gpu_multi_target.py::test_shap_multi_output_tree

[Copilot]

Pull request overview

This PR extends the GPU exact QuadratureTreeSHAP implementation to support vector-leaf (multi-target) trees, enabling pred_contribs=True and pred_interactions=True for multi_output_tree on CUDA and aligning GPU capabilities with existing CPU SHAP support.

Changes:

• Refactors GPU SHAP model compression to store tree topology once and store leaf outputs in a separate array, with per-target task metadata.
• Enables GPU exact SHAP contributions and interaction contributions for multi-target (vector-leaf) trees, removing the previous unsupported-path checks.
• Expands C++ SHAP test coverage to always include vector-leaf cases and adds a CUDA Python regression test for multi_output_tree SHAP outputs.

Reviewed changes

Copilot reviewed 5 out of 5 changed files in this pull request and generated 4 comments.

Show a summary per file

File	Description
tests/python-gpu/test_gpu_multi_target.py	Adds a regression test validating SHAP contribs/interactions shapes and additivity for multi_output_tree on CUDA.
tests/cpp/predictor/test_shap.h	Simplifies SHAP test case builder API to always include vector-leaf cases.
tests/cpp/predictor/test_shap.cu	Runs the GPU SHAP output test suite over the unified test-case set (now including vector-leaf).
tests/cpp/predictor/test_shap.cc	Always includes a multi-target vector-leaf case in shared SHAP test cases for CPU and GPU.
src/predictor/interpretability/shap.cu	Implements vector-leaf support in GPU QuadratureTreeSHAP via new compression layout and target-aware leaf lookup in kernels.

---

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

[Copilot]

Pull request overview

Copilot reviewed 7 out of 7 changed files in this pull request and generated no new comments.