Compatibility fix for XGBoost 3.2 (#655)

* Compatibility fix for XGBoost 3.2

* Address warning from XGBoost

* Check for leaf_weights field

Author: hcho3

src/model_loader/detail/xgboost_json/delegated_handler.cc

@@ -348,6 +348,7 @@ bool RegTreeHandler::StartArray() {
   return (push_key_handler<ArrayHandler<float>>("loss_changes", loss_changes)
           || push_key_handler<ArrayHandler<float>>("sum_hessian", sum_hessian)
           || push_key_handler<ArrayHandler<float>>("base_weights", base_weights)
+          || push_key_handler<ArrayHandler<float>>("leaf_weights", leaf_weights)
           || push_key_handler<ArrayHandler<int>>("categories_segments", categories_segments)
           || push_key_handler<ArrayHandler<int>>("categories_sizes", categories_sizes)
           || push_key_handler<ArrayHandler<int>>("categories_nodes", categories_nodes)
@@ -390,10 +391,8 @@ bool RegTreeHandler::EndObject() {
   if (output.size_leaf_vector == 0) {
     output.size_leaf_vector = 1;  // In XGBoost, size_leaf_vector=0 indicates a scalar output
   }
-  if (num_nodes * output.size_leaf_vector != base_weights.size()) {
-    TREELITE_LOG(ERROR) << "Field base_weights has an incorrect dimension. Expected: "
-                        << (num_nodes * output.size_leaf_vector)
-                        << ", Actual: " << base_weights.size();
+  if (output.size_leaf_vector != 1 && leaf_weights.empty()) {
+    TREELITE_LOG(ERROR) << "Field leaf_weights must be provided for multi-target trees.";
     return false;
   }
   if (static_cast<std::size_t>(num_nodes) != left_children.size()) {
@@ -440,9 +439,10 @@ bool RegTreeHandler::EndObject() {
       if (size_leaf_vector > 1) {
         // Vector output
         std::vector<float> leafvec(size_leaf_vector);
-        std::transform(&base_weights[node_id * size_leaf_vector],
-            &base_weights[(node_id + 1) * size_leaf_vector], leafvec.begin(),
-            [](float e) { return static_cast<float>(e); });
+        auto leaf_id = right_children[node_id];
+        TREELITE_CHECK_NE(leaf_id, -1) << "Expected a leaf node at index " << node_id;
+        std::copy(&leaf_weights[leaf_id * size_leaf_vector],
+            &leaf_weights[(leaf_id + 1) * size_leaf_vector], leafvec.begin());
         model_builder.LeafVector(leafvec);
       } else {
         // Scalar leaf output
@@ -487,7 +487,7 @@ bool RegTreeHandler::is_recognized_key(std::string const& key) {
           || key == "categories" || key == "leaf_child_counts" || key == "left_children"
           || key == "right_children" || key == "parents" || key == "split_indices"
           || key == "split_type" || key == "split_conditions" || key == "default_left"
-          || key == "tree_param" || key == "id");
+          || key == "tree_param" || key == "id" || key == "leaf_weights");
 }
 
 /******************************************************************************

src/model_loader/detail/xgboost_json/delegated_handler.h

@@ -414,6 +414,7 @@ class RegTreeHandler : public OutputHandler<ParsedRegTreeParams> {
   std::vector<float> loss_changes;
   std::vector<float> sum_hessian;
   std::vector<float> base_weights;
+  std::vector<float> leaf_weights;
   std::vector<int> left_children;
   std::vector<int> right_children;
   std::vector<int> parents;

tests/python/test_xgboost_integration.py

@@ -30,6 +30,12 @@
 from .util import TemporaryDirectory, has_pandas, to_categorical
 
 
+def _get_model_filename(name, model_format):
+    if model_format == "ubjson":
+        model_format = "ubj"
+    return f"{name}.{model_format}"
+
+
 def generate_data_for_squared_log_error(n_targets: int = 1):
     """Generate data containing outliers."""
     n_rows = 4096
@@ -108,7 +114,7 @@ def test_xgb_regressor(
         num_boost_round=num_boost_round,
     )
     with TemporaryDirectory() as tmpdir:
-        model_path = pathlib.Path(tmpdir) / f"model.{model_format}"
+        model_path = pathlib.Path(tmpdir) / _get_model_filename("model", model_format)
         xgb_model.save_model(model_path)
         tl_model = treelite.frontend.load_xgboost_model(
             model_path, format_choice=model_format
@@ -179,7 +185,7 @@ def test_xgb_multiclass_classifier(
     )
 
     with TemporaryDirectory() as tmpdir:
-        model_path = pathlib.Path(tmpdir) / f"iris.{model_format}"
+        model_path = pathlib.Path(tmpdir) / _get_model_filename("iris", model_format)
         xgb_model.save_model(model_path)
         tl_model = treelite.frontend.load_xgboost_model(
             model_path, format_choice=model_format
@@ -262,10 +268,7 @@ def test_xgb_nonlinear_objective(
     )
 
     objective_tag = objective.replace(":", "_")
-    if model_format in ["json", "ubjson"]:
-        model_name = f"nonlinear_{objective_tag}.{model_format}"
-    else:
-        model_name = f"nonlinear_{objective_tag}.deprecated"
+    model_name = _get_model_filename(f"nonlinear_{objective_tag}", model_format)
     with TemporaryDirectory() as tmpdir:
         model_path = pathlib.Path(tmpdir) / model_name
         xgb_model.save_model(model_path)
@@ -458,7 +461,9 @@ def test_xgb_multi_target_binary_classifier(
         tl_model = treelite.frontend.from_xgboost(bst)
     else:
         with TemporaryDirectory() as tmpdir:
-            model_path = pathlib.Path(tmpdir) / f"multi_target.{model_format}"
+            model_path = pathlib.Path(tmpdir) / _get_model_filename(
+                "multi_target", model_format
+            )
             bst.save_model(model_path)
             tl_model = treelite.frontend.load_xgboost_model(
                 model_path, format_choice=model_format
@@ -533,7 +538,7 @@ def test_xgb_multi_target_regressor(
     )
 
     with TemporaryDirectory() as tmpdir:
-        model_path = pathlib.Path(tmpdir) / f"model.{model_format}"
+        model_path = pathlib.Path(tmpdir) / _get_model_filename("model", model_format)
         xgb_model.save_model(model_path)
         tl_model = treelite.frontend.load_xgboost_model(
             model_path, format_choice=model_format
@@ -578,7 +583,7 @@ def test_xgb_detect_format(
     expected_pred = xgb_model.predict(xgb.DMatrix(X)).reshape((X.shape[0], 1, -1))
 
     with TemporaryDirectory() as tmpdir:
-        model_path = pathlib.Path(tmpdir) / f"model.{model_format}"
+        model_path = pathlib.Path(tmpdir) / _get_model_filename("model", model_format)
         xgb_model.save_model(model_path)
         detected_format = treelite.frontend._detect_xgboost_format(model_path)
         assert detected_format == model_format