Bump version to 1.7.4. (#8805)

* [backport] Fix CPU bin compression with categorical data. (#8809)

* Fix CPU bin compression with categorical data.

* The bug causes the maximum category to be lesser than 256 or the maximum number of bins when
the input data is dense.

* Avoid test symbol.

CMakeLists.txt

@@ -1,5 +1,5 @@
 cmake_minimum_required(VERSION 3.14 FATAL_ERROR)
-project(xgboost LANGUAGES CXX C VERSION 1.7.2)
+project(xgboost LANGUAGES CXX C VERSION 1.7.4)
 include(cmake/Utils.cmake)
 list(APPEND CMAKE_MODULE_PATH "${xgboost_SOURCE_DIR}/cmake/modules")
 cmake_policy(SET CMP0022 NEW)

R-package/DESCRIPTION

@@ -1,8 +1,8 @@
 Package: xgboost
 Type: Package
 Title: Extreme Gradient Boosting
-Version: 1.7.2.1
-Date: 2022-12-08
+Version: 1.7.4.1
+Date: 2023-02-15
 Authors@R: c(
   person("Tianqi", "Chen", role = c("aut"),
          email = "tianqi.tchen@gmail.com"),
@@ -66,5 +66,5 @@ Imports:
     methods,
     data.table (>= 1.9.6),
     jsonlite (>= 1.0),
-RoxygenNote: 7.2.1
+RoxygenNote: 7.2.2
 SystemRequirements: GNU make, C++14

R-package/R/xgb.Booster.R

@@ -328,8 +328,9 @@ predict.xgb.Booster <- function(object, newdata, missing = NA, outputmargin = FA
                                 predleaf = FALSE, predcontrib = FALSE, approxcontrib = FALSE, predinteraction = FALSE,
                                 reshape = FALSE, training = FALSE, iterationrange = NULL, strict_shape = FALSE, ...) {
   object <- xgb.Booster.complete(object, saveraw = FALSE)
+
   if (!inherits(newdata, "xgb.DMatrix"))
-    newdata <- xgb.DMatrix(newdata, missing = missing)
+    newdata <- xgb.DMatrix(newdata, missing = missing, nthread = NVL(object$params[["nthread"]], -1))
   if (!is.null(object[["feature_names"]]) &&
       !is.null(colnames(newdata)) &&
       !identical(object[["feature_names"]], colnames(newdata)))

R-package/configure.ac

@@ -2,10 +2,25 @@
 
 AC_PREREQ(2.69)
 
-AC_INIT([xgboost],[1.7.2],[],[xgboost],[])
+AC_INIT([xgboost],[1.7.4],[],[xgboost],[])
 
-# Use this line to set CC variable to a C compiler
-AC_PROG_CC
+: ${R_HOME=`R RHOME`}
+if test -z "${R_HOME}"; then
+  echo "could not determine R_HOME"
+  exit 1
+fi
+
+CXX14=`"${R_HOME}/bin/R" CMD config CXX14`
+CXX14STD=`"${R_HOME}/bin/R" CMD config CXX14STD`
+CXX="${CXX14} ${CXX14STD}"
+CXXFLAGS=`"${R_HOME}/bin/R" CMD config CXXFLAGS`
+
+CC=`"${R_HOME}/bin/R" CMD config CC`
+CFLAGS=`"${R_HOME}/bin/R" CMD config CFLAGS`
+CPPFLAGS=`"${R_HOME}/bin/R" CMD config CPPFLAGS`
+
+LDFLAGS=`"${R_HOME}/bin/R" CMD config LDFLAGS`
+AC_LANG(C++)
 
 ### Check whether backtrace() is part of libc or the external lib libexecinfo
 AC_MSG_CHECKING([Backtrace lib])
@@ -40,7 +55,7 @@ then
   ac_pkg_openmp=no
   AC_MSG_CHECKING([whether OpenMP will work in a package])
   AC_LANG_CONFTEST([AC_LANG_PROGRAM([[#include <omp.h>]], [[ return (omp_get_max_threads() <= 1); ]])])
-  ${CC} -o conftest conftest.c ${CPPFLAGS} ${LDFLAGS} ${OPENMP_LIB} ${OPENMP_CXXFLAGS} 2>/dev/null && ./conftest && ac_pkg_openmp=yes
+  ${CXX} -o conftest conftest.cpp ${CPPFLAGS} ${LDFLAGS} ${OPENMP_LIB} ${OPENMP_CXXFLAGS} 2>/dev/null && ./conftest && ac_pkg_openmp=yes
   AC_MSG_RESULT([${ac_pkg_openmp}])
   if test "${ac_pkg_openmp}" = no; then
     OPENMP_CXXFLAGS=''

R-package/src/Makevars.in

@@ -23,7 +23,6 @@ PKG_LIBS = @OPENMP_CXXFLAGS@ @OPENMP_LIB@ @ENDIAN_FLAG@ @BACKTRACE_LIB@ -pthread
 OBJECTS= \
     ./xgboost_R.o \
     ./xgboost_custom.o \
-    ./xgboost_assert.o \
     ./init.o \
     $(PKGROOT)/src/metric/metric.o \
     $(PKGROOT)/src/metric/elementwise_metric.o \

R-package/src/Makevars.win

@@ -23,7 +23,6 @@ PKG_LIBS = $(SHLIB_OPENMP_CXXFLAGS)  -DDMLC_CMAKE_LITTLE_ENDIAN=1  $(SHLIB_PTHRE
 OBJECTS= \
     ./xgboost_R.o \
     ./xgboost_custom.o \
-    ./xgboost_assert.o \
     ./init.o \
     $(PKGROOT)/src/metric/metric.o \
     $(PKGROOT)/src/metric/elementwise_metric.o \

R-package/src/xgboost_assert.c

@@ -1,26 +0,0 @@
-// Copyright (c) 2014 by Contributors
-#include <stdio.h>
-#include <stdarg.h>
-#include <Rinternals.h>
-
-// implements error handling
-void XGBoostAssert_R(int exp, const char *fmt, ...) {
-  char buf[1024];
-  if (exp == 0) {
-    va_list args;
-    va_start(args, fmt);
-    vsprintf(buf, fmt, args);
-    va_end(args);
-    error("AssertError:%s\n", buf);
-  }
-}
-void XGBoostCheck_R(int exp, const char *fmt, ...) {
-  char buf[1024];
-  if (exp == 0) {
-    va_list args;
-    va_start(args, fmt);
-    vsprintf(buf, fmt, args);
-    va_end(args);
-    error("%s\n", buf);
-  }
-}

doc/tutorials/categorical.rst

@@ -138,11 +138,11 @@ Miscellaneous
 
 By default, XGBoost assumes input categories are integers starting from 0 till the number
 of categories :math:`[0, n\_categories)`. However, user might provide inputs with invalid
-values due to mistakes or missing values. It can be negative value, integer values that
-can not be accurately represented by 32-bit floating point, or values that are larger than
-actual number of unique categories.  During training this is validated but for prediction
-it's treated as the same as missing value for performance reasons.  Lastly, missing values
-are treated as the same as numerical features (using the learned split direction).
+values due to mistakes or missing values in training dataset. It can be negative value,
+integer values that can not be accurately represented by 32-bit floating point, or values
+that are larger than actual number of unique categories.  During training this is
+validated but for prediction it's treated as the same as not-chosen category for
+performance reasons.
 
 
 **********

include/xgboost/version_config.h

@@ -6,6 +6,6 @@
 
 #define XGBOOST_VER_MAJOR 1
 #define XGBOOST_VER_MINOR 7
-#define XGBOOST_VER_PATCH 2
+#define XGBOOST_VER_PATCH 4
 
 #endif  // XGBOOST_VERSION_CONFIG_H_

jvm-packages/pom.xml

@@ -6,7 +6,7 @@
 
     <groupId>ml.dmlc</groupId>
     <artifactId>xgboost-jvm_2.12</artifactId>
-    <version>1.7.2</version>
+    <version>1.7.4</version>
     <packaging>pom</packaging>
     <name>XGBoost JVM Package</name>
     <description>JVM Package for XGBoost</description>

jvm-packages/xgboost4j-example/pom.xml

@@ -6,10 +6,10 @@
     <parent>
         <groupId>ml.dmlc</groupId>
         <artifactId>xgboost-jvm_2.12</artifactId>
-        <version>1.7.2</version>
+        <version>1.7.4</version>
     </parent>
     <artifactId>xgboost4j-example_2.12</artifactId>
-    <version>1.7.2</version>
+    <version>1.7.4</version>
     <packaging>jar</packaging>
     <build>
         <plugins>
@@ -26,7 +26,7 @@
         <dependency>
             <groupId>ml.dmlc</groupId>
             <artifactId>xgboost4j-spark_${scala.binary.version}</artifactId>
-            <version>1.7.2</version>
+            <version>1.7.4</version>
         </dependency>
         <dependency>
             <groupId>org.apache.spark</groupId>
@@ -37,7 +37,7 @@
         <dependency>
             <groupId>ml.dmlc</groupId>
             <artifactId>xgboost4j-flink_${scala.binary.version}</artifactId>
-            <version>1.7.2</version>
+            <version>1.7.4</version>
         </dependency>
         <dependency>
             <groupId>org.apache.commons</groupId>

jvm-packages/xgboost4j-flink/pom.xml

@@ -6,10 +6,10 @@
     <parent>
         <groupId>ml.dmlc</groupId>
         <artifactId>xgboost-jvm_2.12</artifactId>
-        <version>1.7.2</version>
+        <version>1.7.4</version>
     </parent>
     <artifactId>xgboost4j-flink_2.12</artifactId>
-    <version>1.7.2</version>
+    <version>1.7.4</version>
     <build>
         <plugins>
             <plugin>
@@ -26,7 +26,7 @@
         <dependency>
             <groupId>ml.dmlc</groupId>
             <artifactId>xgboost4j_${scala.binary.version}</artifactId>
-            <version>1.7.2</version>
+            <version>1.7.4</version>
         </dependency>
         <dependency>
             <groupId>org.apache.commons</groupId>

jvm-packages/xgboost4j-gpu/pom.xml

@@ -6,10 +6,10 @@
     <parent>
         <groupId>ml.dmlc</groupId>
         <artifactId>xgboost-jvm_2.12</artifactId>
-        <version>1.7.2</version>
+        <version>1.7.4</version>
     </parent>
     <artifactId>xgboost4j-gpu_2.12</artifactId>
-    <version>1.7.2</version>
+    <version>1.7.4</version>
     <packaging>jar</packaging>
 
     <dependencies>

jvm-packages/xgboost4j-spark-gpu/pom.xml

@@ -6,7 +6,7 @@
     <parent>
         <groupId>ml.dmlc</groupId>
         <artifactId>xgboost-jvm_2.12</artifactId>
-        <version>1.7.2</version>
+        <version>1.7.4</version>
     </parent>
     <artifactId>xgboost4j-spark-gpu_2.12</artifactId>
     <build>
@@ -24,7 +24,7 @@
         <dependency>
             <groupId>ml.dmlc</groupId>
             <artifactId>xgboost4j-gpu_${scala.binary.version}</artifactId>
-            <version>1.7.2</version>
+            <version>1.7.4</version>
         </dependency>
         <dependency>
             <groupId>org.apache.spark</groupId>

jvm-packages/xgboost4j-spark/pom.xml

@@ -6,7 +6,7 @@
     <parent>
         <groupId>ml.dmlc</groupId>
         <artifactId>xgboost-jvm_2.12</artifactId>
-        <version>1.7.2</version>
+        <version>1.7.4</version>
     </parent>
     <artifactId>xgboost4j-spark_2.12</artifactId>
     <build>
@@ -24,7 +24,7 @@
         <dependency>
             <groupId>ml.dmlc</groupId>
             <artifactId>xgboost4j_${scala.binary.version}</artifactId>
-            <version>1.7.2</version>
+            <version>1.7.4</version>
         </dependency>
         <dependency>
             <groupId>org.apache.spark</groupId>

jvm-packages/xgboost4j-tester/get_iris.py

@@ -1,9 +1,9 @@
-from sklearn.datasets import load_iris
 import numpy as np
 import pandas
+from sklearn.datasets import load_iris
 
 X, y = load_iris(return_X_y=True)
-y = y.astype(np.int)
+y = y.astype(np.int32)
 df = pandas.DataFrame(data=X, columns=['sepal length', 'sepal width', 'petal length', 'petal width'])
 class_id_to_name = {0:'Iris-setosa', 1:'Iris-versicolor', 2:'Iris-virginica'}
 df['class'] = np.vectorize(class_id_to_name.get)(y)

jvm-packages/xgboost4j/pom.xml

@@ -6,10 +6,10 @@
     <parent>
         <groupId>ml.dmlc</groupId>
         <artifactId>xgboost-jvm_2.12</artifactId>
-        <version>1.7.2</version>
+        <version>1.7.4</version>
     </parent>
     <artifactId>xgboost4j_2.12</artifactId>
-    <version>1.7.2</version>
+    <version>1.7.4</version>
     <packaging>jar</packaging>
 
     <dependencies>

python-package/xgboost/VERSION

@@ -1 +1 @@
-1.7.2
+1.7.4

python-package/xgboost/compat.py

@@ -36,7 +36,6 @@ try:
 
     PANDAS_INSTALLED = True
 except ImportError:
-
     MultiIndex = object
     DataFrame = object
     Series = object
@@ -161,6 +160,7 @@ def concat(value: Sequence[_T]) -> _T:  # pylint: disable=too-many-return-statem
 # `importlib.utils`, except it's unclear from its document on how to use it.  This one
 # seems to be easy to understand and works out of box.
 
+
 # Copyright 2015 The TensorFlow Authors. All Rights Reserved.
 #
 # Licensed under the Apache License, Version 2.0 (the "License"); you may not use this

python-package/xgboost/core.py

@@ -2172,6 +2172,7 @@ class Booster:
         )
         return _prediction_output(shape, dims, preds, False)
 
+    # pylint: disable=too-many-statements
     def inplace_predict(
         self,
         data: DataType,
@@ -2192,10 +2193,10 @@ class Booster:
 
         .. code-block:: python
 
-            booster.set_param({'predictor': 'gpu_predictor'})
+            booster.set_param({"predictor": "gpu_predictor"})
             booster.inplace_predict(cupy_array)
 
-            booster.set_param({'predictor': 'cpu_predictor})
+            booster.set_param({"predictor": "cpu_predictor"})
             booster.inplace_predict(numpy_array)
 
         .. versionadded:: 1.1.0
@@ -2301,14 +2302,16 @@ class Booster:
             )
             return _prediction_output(shape, dims, preds, False)
         if isinstance(data, scipy.sparse.csr_matrix):
-            csr = data
+            from .data import _transform_scipy_csr
+
+            data = _transform_scipy_csr(data)
             _check_call(
                 _LIB.XGBoosterPredictFromCSR(
                     self.handle,
-                    _array_interface(csr.indptr),
-                    _array_interface(csr.indices),
-                    _array_interface(csr.data),
-                    c_bst_ulong(csr.shape[1]),
+                    _array_interface(data.indptr),
+                    _array_interface(data.indices),
+                    _array_interface(data.data),
+                    c_bst_ulong(data.shape[1]),
                     from_pystr_to_cstr(json.dumps(args)),
                     p_handle,
                     ctypes.byref(shape),

python-package/xgboost/data.py

@@ -30,6 +30,7 @@ from .core import (
     c_array,
     c_str,
     from_pystr_to_cstr,
+    make_jcargs,
 )
 
 DispatchedDataBackendReturnType = Tuple[
@@ -80,6 +81,21 @@ def _array_interface(data: np.ndarray) -> bytes:
     return interface_str
 
 
+def _transform_scipy_csr(data: DataType) -> DataType:
+    from scipy.sparse import csr_matrix
+
+    indptr, _ = _ensure_np_dtype(data.indptr, data.indptr.dtype)
+    indices, _ = _ensure_np_dtype(data.indices, data.indices.dtype)
+    values, _ = _ensure_np_dtype(data.data, data.data.dtype)
+    if (
+        indptr is not data.indptr
+        or indices is not data.indices
+        or values is not data.data
+    ):
+        data = csr_matrix((values, indices, indptr), shape=data.shape)
+    return data
+
+
 def _from_scipy_csr(
     data: DataType,
     missing: FloatCompatible,
@@ -93,18 +109,14 @@ def _from_scipy_csr(
             f"length mismatch: {len(data.indices)} vs {len(data.data)}"
         )
     handle = ctypes.c_void_p()
-    args = {
-        "missing": float(missing),
-        "nthread": int(nthread),
-    }
-    config = bytes(json.dumps(args), "utf-8")
+    data = _transform_scipy_csr(data)
     _check_call(
         _LIB.XGDMatrixCreateFromCSR(
             _array_interface(data.indptr),
             _array_interface(data.indices),
             _array_interface(data.data),
             c_bst_ulong(data.shape[1]),
-            config,
+            make_jcargs(missing=float(missing), nthread=int(nthread)),
             ctypes.byref(handle),
         )
     )
@@ -153,12 +165,13 @@ def _is_numpy_array(data: DataType) -> bool:
 
 
 def _ensure_np_dtype(
-    data: DataType,
-    dtype: Optional[NumpyDType]
+    data: DataType, dtype: Optional[NumpyDType]
 ) -> Tuple[np.ndarray, Optional[NumpyDType]]:
     if data.dtype.hasobject or data.dtype in [np.float16, np.bool_]:
-        data = data.astype(np.float32, copy=False)
         dtype = np.float32
+        data = data.astype(dtype, copy=False)
+    if not data.flags.aligned:
+        data = np.require(data, requirements="A")
     return data, dtype
 
 
@@ -1197,11 +1210,13 @@ def _proxy_transform(
         data, _ = _ensure_np_dtype(data, data.dtype)
         return data, None, feature_names, feature_types
     if _is_scipy_csr(data):
+        data = _transform_scipy_csr(data)
         return data, None, feature_names, feature_types
     if _is_pandas_df(data):
         arr, feature_names, feature_types = _transform_pandas_df(
             data, enable_categorical, feature_names, feature_types
         )
+        arr, _ = _ensure_np_dtype(arr, arr.dtype)
         return arr, None, feature_names, feature_types
     raise TypeError("Value type is not supported for data iterator:" + str(type(data)))
 

python-package/xgboost/sklearn.py

@@ -674,7 +674,7 @@ class XGBModel(XGBModelBase):
                     self.kwargs = {}
                 self.kwargs[key] = value
 
-        if hasattr(self, "_Booster"):
+        if self.__sklearn_is_fitted__():
             parameters = self.get_xgb_params()
             self.get_booster().set_param(parameters)
 
@@ -701,39 +701,12 @@ class XGBModel(XGBModelBase):
                 np.iinfo(np.int32).max
             )
 
-        def parse_parameter(value: Any) -> Optional[Union[int, float, str]]:
-            for t in (int, float, str):
-                try:
-                    ret = t(value)
-                    return ret
-                except ValueError:
-                    continue
-            return None
-
-        # Get internal parameter values
-        try:
-            config = json.loads(self.get_booster().save_config())
-            stack = [config]
-            internal = {}
-            while stack:
-                obj = stack.pop()
-                for k, v in obj.items():
-                    if k.endswith("_param"):
-                        for p_k, p_v in v.items():
-                            internal[p_k] = p_v
-                    elif isinstance(v, dict):
-                        stack.append(v)
-
-            for k, v in internal.items():
-                if k in params and params[k] is None:
-                    params[k] = parse_parameter(v)
-        except ValueError:
-            pass
         return params
 
     def get_xgb_params(self) -> Dict[str, Any]:
         """Get xgboost specific parameters."""
-        params = self.get_params()
+        params: Dict[str, Any] = self.get_params()
+
         # Parameters that should not go into native learner.
         wrapper_specific = {
             "importance_type",
@@ -750,6 +723,7 @@ class XGBModel(XGBModelBase):
         for k, v in params.items():
             if k not in wrapper_specific and not callable(v):
                 filtered[k] = v
+
         return filtered
 
     def get_num_boosting_rounds(self) -> int:
@@ -1070,7 +1044,7 @@ class XGBModel(XGBModelBase):
         # error with incompatible data type.
         # Inplace predict doesn't handle as many data types as DMatrix, but it's
         # sufficient for dask interface where input is simpiler.
-        predictor = self.get_params().get("predictor", None)
+        predictor = self.get_xgb_params().get("predictor", None)
         if predictor in ("auto", None) and self.booster != "gblinear":
             return True
         return False
@@ -1336,7 +1310,7 @@ class XGBModel(XGBModelBase):
         -------
         coef_ : array of shape ``[n_features]`` or ``[n_classes, n_features]``
         """
-        if self.get_params()["booster"] != "gblinear":
+        if self.get_xgb_params()["booster"] != "gblinear":
             raise AttributeError(
                 f"Coefficients are not defined for Booster type {self.booster}"
             )
@@ -1366,7 +1340,7 @@ class XGBModel(XGBModelBase):
         -------
         intercept_ : array of shape ``(1,)`` or ``[n_classes]``
         """
-        if self.get_params()["booster"] != "gblinear":
+        if self.get_xgb_params()["booster"] != "gblinear":
             raise AttributeError(
                 f"Intercept (bias) is not defined for Booster type {self.booster}"
             )

python-package/xgboost/spark/core.py

@@ -140,6 +140,13 @@ _unsupported_predict_params = {
 }
 
 
+# TODO: supply hint message for all other unsupported params.
+_unsupported_params_hint_message = {
+    "enable_categorical": "`xgboost.spark` estimators do not have 'enable_categorical' param, "
+    "but you can set `feature_types` param and mark categorical features with 'c' string."
+}
+
+
 class _SparkXGBParams(
     HasFeaturesCol,
     HasLabelCol,
@@ -523,7 +530,10 @@ class _SparkXGBEstimator(Estimator, _SparkXGBParams, MLReadable, MLWritable):
                     or k in _unsupported_predict_params
                     or k in _unsupported_train_params
                 ):
-                    raise ValueError(f"Unsupported param '{k}'.")
+                    err_msg = _unsupported_params_hint_message.get(
+                        k, f"Unsupported param '{k}'."
+                    )
+                    raise ValueError(err_msg)
                 _extra_params[k] = v
         _existing_extra_params = self.getOrDefault(self.arbitrary_params_dict)
         self._set(arbitrary_params_dict={**_existing_extra_params, **_extra_params})
@@ -749,6 +759,8 @@ class _SparkXGBEstimator(Estimator, _SparkXGBParams, MLReadable, MLWritable):
             "feature_weights": self.getOrDefault(self.feature_weights),
             "missing": float(self.getOrDefault(self.missing)),
         }
+        if dmatrix_kwargs["feature_types"] is not None:
+            dmatrix_kwargs["enable_categorical"] = True
         booster_params["nthread"] = cpu_per_task
         use_gpu = self.getOrDefault(self.use_gpu)
 

src/common/categorical.h

@@ -48,20 +48,21 @@ inline XGBOOST_DEVICE bool InvalidCat(float cat) {
   return cat < 0 || cat >= kMaxCat;
 }
 
-/* \brief Whether should it traverse to left branch of a tree.
+/**
+ * \brief Whether should it traverse to left branch of a tree.
  *
- *  For one hot split, go to left if it's NOT the matching category.
+ *   Go to left if it's NOT the matching category, which matches one-hot encoding.
  */
-template <bool validate = true>
-inline XGBOOST_DEVICE bool Decision(common::Span<uint32_t const> cats, float cat, bool dft_left) {
+inline XGBOOST_DEVICE bool Decision(common::Span<uint32_t const> cats, float cat) {
   KCatBitField const s_cats(cats);
-  // FIXME: Size() is not accurate since it represents the size of bit set instead of
-  // actual number of categories.
-  if (XGBOOST_EXPECT(validate && (InvalidCat(cat) || cat >= s_cats.Size()), false)) {
-    return dft_left;
+  if (XGBOOST_EXPECT(InvalidCat(cat), false)) {
+    return true;
   }
 
   auto pos = KCatBitField::ToBitPos(cat);
+  // If the input category is larger than the size of the bit field, it implies that the
+  // category is not chosen. Otherwise the bit field would have the category instead of
+  // being smaller than the category value.
   if (pos.int_pos >= cats.size()) {
     return true;
   }

src/common/column_matrix.cc

@@ -46,7 +46,7 @@ void ColumnMatrix::InitStorage(GHistIndexMatrix const& gmat, double sparse_thres
     feature_offsets_[fid] = accum_index;
   }
 
-  SetTypeSize(gmat.max_num_bins);
+  SetTypeSize(gmat.MaxNumBinPerFeat());
   auto storage_size =
       feature_offsets_.back() * static_cast<std::underlying_type_t<BinTypeSize>>(bins_type_size_);
   index_.resize(storage_size, 0);

src/common/linalg_op.h

@@ -62,7 +62,7 @@ void ElementWiseKernel(GenericParameter const* ctx, linalg::TensorView<T, D> t,
 #endif  // !defined(XGBOOST_USE_CUDA)
 
 template <typename T, std::int32_t kDim>
-auto cbegin(TensorView<T, kDim> v) {  // NOLINT
+auto cbegin(TensorView<T, kDim> const& v) {  // NOLINT
   auto it = common::MakeIndexTransformIter([&](size_t i) -> std::remove_cv_t<T> const& {
     return linalg::detail::Apply(v, linalg::UnravelIndex(i, v.Shape()));
   });
@@ -70,19 +70,19 @@ auto cbegin(TensorView<T, kDim> v) {  // NOLINT
 }
 
 template <typename T, std::int32_t kDim>
-auto cend(TensorView<T, kDim> v) {  // NOLINT
+auto cend(TensorView<T, kDim> const& v) {  // NOLINT
   return cbegin(v) + v.Size();
 }
 
 template <typename T, std::int32_t kDim>
-auto begin(TensorView<T, kDim> v) {  // NOLINT
+auto begin(TensorView<T, kDim>& v) {  // NOLINT
   auto it = common::MakeIndexTransformIter(
       [&](size_t i) -> T& { return linalg::detail::Apply(v, linalg::UnravelIndex(i, v.Shape())); });
   return it;
 }
 
 template <typename T, std::int32_t kDim>
-auto end(TensorView<T, kDim> v) {  // NOLINT
+auto end(TensorView<T, kDim>& v) {  // NOLINT
   return begin(v) + v.Size();
 }
 }  // namespace linalg

src/common/partition_builder.h

@@ -144,7 +144,7 @@ class PartitionBuilder {
         auto gidx = gidx_calc(ridx);
         bool go_left = default_left;
         if (gidx > -1) {
-          go_left = Decision(node_cats, cut_values[gidx], default_left);
+          go_left = Decision(node_cats, cut_values[gidx]);
         }
         return go_left;
       } else {
@@ -157,7 +157,7 @@ class PartitionBuilder {
       bool go_left = default_left;
       if (gidx > -1) {
         if (is_cat) {
-          go_left = Decision(node_cats, cut_values[gidx], default_left);
+          go_left = Decision(node_cats, cut_values[gidx]);
         } else {
           go_left = cut_values[gidx] <= nodes[node_in_set].split.split_value;
         }

src/data/array_interface.h

@@ -1,5 +1,5 @@
-/*!
- * Copyright 2019-2021 by Contributors
+/**
+ * Copyright 2019-2023 by XGBoost Contributors
  * \file array_interface.h
  * \brief View of __array_interface__
  */
@@ -7,9 +7,11 @@
 #define XGBOOST_DATA_ARRAY_INTERFACE_H_
 
 #include <algorithm>
-#include <cinttypes>
+#include <cstddef>  // std::size_t
+#include <cstdint>
 #include <map>
 #include <string>
+#include <type_traits>  // std::alignment_of,std::remove_pointer_t
 #include <utility>
 #include <vector>
 
@@ -394,6 +396,11 @@ class ArrayInterface {
 
     data = ArrayInterfaceHandler::ExtractData(array, n);
     static_assert(allow_mask ? D == 1 : D >= 1, "Masked ndarray is not supported.");
+
+    auto alignment = this->ElementAlignment();
+    auto ptr = reinterpret_cast<uintptr_t>(this->data);
+    CHECK_EQ(ptr % alignment, 0) << "Input pointer misalignment.";
+
     if (allow_mask) {
       common::Span<RBitField8::value_type> s_mask;
       size_t n_bits = ArrayInterfaceHandler::ExtractMask(array, &s_mask);
@@ -512,9 +519,15 @@ class ArrayInterface {
     return func(reinterpret_cast<uint64_t const *>(data));
   }
 
-  XGBOOST_DEVICE size_t ElementSize() {
-    return this->DispatchCall(
-        [](auto *p_values) { return sizeof(std::remove_pointer_t<decltype(p_values)>); });
+  XGBOOST_DEVICE std::size_t ElementSize() const {
+    return this->DispatchCall([](auto *typed_data_ptr) {
+      return sizeof(std::remove_pointer_t<decltype(typed_data_ptr)>);
+    });
+  }
+  XGBOOST_DEVICE std::size_t ElementAlignment() const {
+    return this->DispatchCall([](auto *typed_data_ptr) {
+      return std::alignment_of<std::remove_pointer_t<decltype(typed_data_ptr)>>::value;
+    });
   }
 
   template <typename T = float, typename... Index>

src/data/gradient_index.cc

@@ -20,13 +20,13 @@ GHistIndexMatrix::GHistIndexMatrix() : columns_{std::make_unique<common::ColumnM
 
 GHistIndexMatrix::GHistIndexMatrix(DMatrix *p_fmat, bst_bin_t max_bins_per_feat,
                                    double sparse_thresh, bool sorted_sketch, int32_t n_threads,
-                                   common::Span<float> hess) {
+                                   common::Span<float> hess)
+    : max_numeric_bins_per_feat{max_bins_per_feat} {
   CHECK(p_fmat->SingleColBlock());
   // We use sorted sketching for approx tree method since it's more efficient in
   // computation time (but higher memory usage).
   cut = common::SketchOnDMatrix(p_fmat, max_bins_per_feat, n_threads, sorted_sketch, hess);
 
-  max_num_bins = max_bins_per_feat;
   const uint32_t nbins = cut.Ptrs().back();
   hit_count.resize(nbins, 0);
   hit_count_tloc_.resize(n_threads * nbins, 0);
@@ -63,7 +63,7 @@ GHistIndexMatrix::GHistIndexMatrix(MetaInfo const &info, common::HistogramCuts &
     : row_ptr(info.num_row_ + 1, 0),
       hit_count(cuts.TotalBins(), 0),
       cut{std::forward<common::HistogramCuts>(cuts)},
-      max_num_bins(max_bin_per_feat),
+      max_numeric_bins_per_feat(max_bin_per_feat),
       isDense_{info.num_col_ * info.num_row_ == info.num_nonzero_} {}
 
 #if !defined(XGBOOST_USE_CUDA)
@@ -86,13 +86,13 @@ void GHistIndexMatrix::PushBatch(SparsePage const &batch, common::Span<FeatureTy
 }
 
 GHistIndexMatrix::GHistIndexMatrix(SparsePage const &batch, common::Span<FeatureType const> ft,
-                                   common::HistogramCuts const &cuts, int32_t max_bins_per_feat,
-                                   bool isDense, double sparse_thresh, int32_t n_threads) {
+                                   common::HistogramCuts cuts, int32_t max_bins_per_feat,
+                                   bool isDense, double sparse_thresh, int32_t n_threads)
+    : cut{std::move(cuts)},
+      max_numeric_bins_per_feat{max_bins_per_feat},
+      base_rowid{batch.base_rowid},
+      isDense_{isDense} {
   CHECK_GE(n_threads, 1);
-  base_rowid = batch.base_rowid;
-  isDense_ = isDense;
-  cut = cuts;
-  max_num_bins = max_bins_per_feat;
   CHECK_EQ(row_ptr.size(), 0);
   // The number of threads is pegged to the batch size. If the OMP
   // block is parallelized on anything other than the batch/block size,
@@ -127,12 +127,13 @@ INSTANTIATION_PUSH(data::SparsePageAdapterBatch)
 #undef INSTANTIATION_PUSH
 
 void GHistIndexMatrix::ResizeIndex(const size_t n_index, const bool isDense) {
-  if ((max_num_bins - 1 <= static_cast<int>(std::numeric_limits<uint8_t>::max())) && isDense) {
+  if ((MaxNumBinPerFeat() - 1 <= static_cast<int>(std::numeric_limits<uint8_t>::max())) &&
+      isDense) {
     // compress dense index to uint8
     index.SetBinTypeSize(common::kUint8BinsTypeSize);
     index.Resize((sizeof(uint8_t)) * n_index);
-  } else if ((max_num_bins - 1 > static_cast<int>(std::numeric_limits<uint8_t>::max()) &&
-              max_num_bins - 1 <= static_cast<int>(std::numeric_limits<uint16_t>::max())) &&
+  } else if ((MaxNumBinPerFeat() - 1 > static_cast<int>(std::numeric_limits<uint8_t>::max()) &&
+              MaxNumBinPerFeat() - 1 <= static_cast<int>(std::numeric_limits<uint16_t>::max())) &&
              isDense) {
     // compress dense index to uint16
     index.SetBinTypeSize(common::kUint16BinsTypeSize);

src/data/gradient_index.cu

@@ -65,7 +65,7 @@ void GetRowPtrFromEllpack(Context const* ctx, EllpackPageImpl const* page,
 
 GHistIndexMatrix::GHistIndexMatrix(Context const* ctx, MetaInfo const& info,
                                    EllpackPage const& in_page, BatchParam const& p)
-    : max_num_bins{p.max_bin} {
+    : max_numeric_bins_per_feat{p.max_bin} {
   auto page = in_page.Impl();
   isDense_ = page->is_dense;
 

src/data/gradient_index.h

@@ -133,11 +133,15 @@ class GHistIndexMatrix {
   std::vector<size_t> hit_count;
   /*! \brief The corresponding cuts */
   common::HistogramCuts cut;
-  /*! \brief max_bin for each feature. */
-  bst_bin_t max_num_bins;
+  /** \brief max_bin for each feature. */
+  bst_bin_t max_numeric_bins_per_feat;
   /*! \brief base row index for current page (used by external memory) */
   size_t base_rowid{0};
 
+  bst_bin_t MaxNumBinPerFeat() const {
+    return std::max(static_cast<bst_bin_t>(cut.MaxCategory() + 1), max_numeric_bins_per_feat);
+  }
+
   ~GHistIndexMatrix();
   /**
    * \brief Constrcutor for SimpleDMatrix.
@@ -160,7 +164,7 @@ class GHistIndexMatrix {
    * \brief Constructor for external memory.
    */
   GHistIndexMatrix(SparsePage const& page, common::Span<FeatureType const> ft,
-                   common::HistogramCuts const& cuts, int32_t max_bins_per_feat, bool is_dense,
+                   common::HistogramCuts cuts, int32_t max_bins_per_feat, bool is_dense,
                    double sparse_thresh, int32_t n_threads);
   GHistIndexMatrix();  // also for ext mem, empty ctor so that we can read the cache back.
 

src/data/gradient_index_format.cc

@@ -35,7 +35,7 @@ class GHistIndexRawFormat : public SparsePageFormat<GHistIndexMatrix> {
     if (!fi->Read(&page->hit_count)) {
       return false;
     }
-    if (!fi->Read(&page->max_num_bins)) {
+    if (!fi->Read(&page->max_numeric_bins_per_feat)) {
       return false;
     }
     if (!fi->Read(&page->base_rowid)) {
@@ -76,8 +76,8 @@ class GHistIndexRawFormat : public SparsePageFormat<GHistIndexMatrix> {
         page.hit_count.size() * sizeof(decltype(page.hit_count)::value_type) +
         sizeof(uint64_t);
     // max_bins, base row, is_dense
-    fo->Write(page.max_num_bins);
-    bytes += sizeof(page.max_num_bins);
+    fo->Write(page.max_numeric_bins_per_feat);
+    bytes += sizeof(page.max_numeric_bins_per_feat);
     fo->Write(page.base_rowid);
     bytes += sizeof(page.base_rowid);
     fo->Write(page.IsDense());

src/data/iterative_dmatrix.cc

@@ -58,6 +58,13 @@ void GetCutsFromRef(std::shared_ptr<DMatrix> ref_, bst_feature_t n_features, Bat
     }
   };
   auto ellpack = [&]() {
+    // workaround ellpack being initialized from CPU.
+    if (p.gpu_id == Context::kCpuId) {
+      p.gpu_id = ref_->Ctx()->gpu_id;
+    }
+    if (p.gpu_id == Context::kCpuId) {
+      p.gpu_id = 0;
+    }
     for (auto const& page : ref_->GetBatches<EllpackPage>(p)) {
       GetCutsFromEllpack(page, p_cuts);
       break;
@@ -172,9 +179,9 @@ void IterativeDMatrix::InitFromCPU(DataIterHandle iter_handle, float missing,
     size_t i = 0;
     while (iter.Next()) {
       if (!p_sketch) {
-        p_sketch.reset(new common::HostSketchContainer{batch_param_.max_bin,
-                                                       proxy->Info().feature_types.ConstHostSpan(),
-                                                       column_sizes, false, ctx_.Threads()});
+        p_sketch.reset(new common::HostSketchContainer{
+            batch_param_.max_bin, proxy->Info().feature_types.ConstHostSpan(), column_sizes,
+            !proxy->Info().group_ptr_.empty(), ctx_.Threads()});
       }
       HostAdapterDispatch(proxy, [&](auto const& batch) {
         proxy->Info().num_nonzero_ = batch_nnz[i];

src/data/simple_dmatrix.cc

@@ -42,6 +42,7 @@ DMatrix* SimpleDMatrix::Slice(common::Span<int32_t const> ridxs) {
     out->Info() = this->Info().Slice(ridxs);
     out->Info().num_nonzero_ = h_offset.back();
   }
+  out->ctx_ = this->ctx_;
   return out;
 }
 

src/gbm/gbtree.cc

@@ -28,6 +28,7 @@
 #include "xgboost/logging.h"
 #include "xgboost/objective.h"
 #include "xgboost/predictor.h"
+#include "xgboost/string_view.h"
 #include "xgboost/tree_updater.h"
 
 namespace xgboost {
@@ -395,23 +396,36 @@ void GBTree::LoadConfig(Json const& in) {
   tparam_.process_type = TreeProcessType::kDefault;
   int32_t const n_gpus = xgboost::common::AllVisibleGPUs();
   if (n_gpus == 0 && tparam_.predictor == PredictorType::kGPUPredictor) {
-    LOG(WARNING)
-        << "Loading from a raw memory buffer on CPU only machine.  "
-           "Changing predictor to auto.";
+    LOG(WARNING) << "Loading from a raw memory buffer on CPU only machine.  "
+                    "Changing predictor to auto.";
     tparam_.UpdateAllowUnknown(Args{{"predictor", "auto"}});
   }
+
+  auto msg = StringView{
+      R"(
+  Loading from a raw memory buffer (like pickle in Python, RDS in R) on a CPU-only
+  machine. Consider using `save_model/load_model` instead. See:
+
+    https://xgboost.readthedocs.io/en/latest/tutorials/saving_model.html
+
+  for more details about differences between saving model and serializing.)"};
+
   if (n_gpus == 0 && tparam_.tree_method == TreeMethod::kGPUHist) {
     tparam_.UpdateAllowUnknown(Args{{"tree_method", "hist"}});
-    LOG(WARNING)
-        << "Loading from a raw memory buffer on CPU only machine.  "
-           "Changing tree_method to hist.";
+    LOG(WARNING) << msg << "  Changing `tree_method` to `hist`.";
   }
 
   auto const& j_updaters = get<Object const>(in["updater"]);
   updaters_.clear();
+
   for (auto const& kv : j_updaters) {
-    std::unique_ptr<TreeUpdater> up(
-        TreeUpdater::Create(kv.first, ctx_, model_.learner_model_param->task));
+    auto name = kv.first;
+    if (n_gpus == 0 && name == "grow_gpu_hist") {
+      name = "grow_quantile_histmaker";
+      LOG(WARNING) << "Changing updater from `grow_gpu_hist` to `grow_quantile_histmaker`.";
+    }
+    std::unique_ptr<TreeUpdater> up{
+        TreeUpdater::Create(name, ctx_, model_.learner_model_param->task)};
     up->LoadConfig(kv.second);
     updaters_.push_back(std::move(up));
   }

src/predictor/predict_fn.h

@@ -18,9 +18,7 @@ inline XGBOOST_DEVICE bst_node_t GetNextNode(const RegTree::Node &node, const bs
     if (has_categorical && common::IsCat(cats.split_type, nid)) {
       auto node_categories =
           cats.categories.subspan(cats.node_ptr[nid].beg, cats.node_ptr[nid].size);
-      return common::Decision<true>(node_categories, fvalue, node.DefaultLeft())
-                 ? node.LeftChild()
-                 : node.RightChild();
+      return common::Decision(node_categories, fvalue) ? node.LeftChild() : node.RightChild();
     } else {
       return node.LeftChild() + !(fvalue < node.SplitCond());
     }

src/tree/gpu_hist/evaluate_splits.cu

@@ -248,8 +248,10 @@ class EvaluateSplitAgent {
 
 template <int kBlockSize>
 __global__ __launch_bounds__(kBlockSize) void EvaluateSplitsKernel(
-    bst_feature_t number_active_features, common::Span<const EvaluateSplitInputs> d_inputs,
-    const EvaluateSplitSharedInputs shared_inputs, common::Span<bst_feature_t> sorted_idx,
+    bst_feature_t max_active_features,
+    common::Span<const EvaluateSplitInputs> d_inputs,
+    const EvaluateSplitSharedInputs shared_inputs,
+    common::Span<bst_feature_t> sorted_idx,
     const TreeEvaluator::SplitEvaluator<GPUTrainingParam> evaluator,
     common::Span<DeviceSplitCandidate> out_candidates) {
   // Aligned && shared storage for best_split
@@ -263,11 +265,15 @@ __global__ __launch_bounds__(kBlockSize) void EvaluateSplitsKernel(
   __syncthreads();
 
   // Allocate blocks to one feature of one node
-  const auto input_idx = blockIdx.x / number_active_features;
+  const auto input_idx = blockIdx.x / max_active_features;
   const EvaluateSplitInputs &inputs = d_inputs[input_idx];
   // One block for each feature. Features are sampled, so fidx != blockIdx.x
-
-  int fidx = inputs.feature_set[blockIdx.x % number_active_features];
+  // Some blocks may not have any feature to work on, simply return
+  int feature_offset = blockIdx.x % max_active_features;
+  if (feature_offset >= inputs.feature_set.size()) {
+    return;
+  }
+  int fidx = inputs.feature_set[feature_offset];
 
   using AgentT = EvaluateSplitAgent<kBlockSize>;
   __shared__ typename AgentT::TempStorage temp_storage;
@@ -338,7 +344,8 @@ __device__ void SetCategoricalSplit(const EvaluateSplitSharedInputs &shared_inpu
 }
 
 void GPUHistEvaluator::LaunchEvaluateSplits(
-    bst_feature_t number_active_features, common::Span<const EvaluateSplitInputs> d_inputs,
+    bst_feature_t max_active_features,
+    common::Span<const EvaluateSplitInputs> d_inputs,
     EvaluateSplitSharedInputs shared_inputs,
     TreeEvaluator::SplitEvaluator<GPUTrainingParam> evaluator,
     common::Span<DeviceSplitCandidate> out_splits) {
@@ -346,20 +353,25 @@ void GPUHistEvaluator::LaunchEvaluateSplits(
     this->SortHistogram(d_inputs, shared_inputs, evaluator);
   }
 
-  size_t combined_num_features = number_active_features * d_inputs.size();
-  dh::TemporaryArray<DeviceSplitCandidate> feature_best_splits(combined_num_features);
+  size_t combined_num_features = max_active_features * d_inputs.size();
+  dh::TemporaryArray<DeviceSplitCandidate> feature_best_splits(
+      combined_num_features, DeviceSplitCandidate());
 
   // One block for each feature
   uint32_t constexpr kBlockThreads = 32;
-  dh::LaunchKernel {static_cast<uint32_t>(combined_num_features), kBlockThreads, 0}(
-      EvaluateSplitsKernel<kBlockThreads>, number_active_features, d_inputs,
-      shared_inputs, this->SortedIdx(d_inputs.size(), shared_inputs.feature_values.size()),
+  dh::LaunchKernel{static_cast<uint32_t>(combined_num_features), kBlockThreads,
+                   0}(
+      EvaluateSplitsKernel<kBlockThreads>, max_active_features, d_inputs,
+      shared_inputs,
+      this->SortedIdx(d_inputs.size(), shared_inputs.feature_values.size()),
       evaluator, dh::ToSpan(feature_best_splits));
 
   // Reduce to get best candidate for left and right child over all features
-  auto reduce_offset = dh::MakeTransformIterator<size_t>(
-      thrust::make_counting_iterator(0llu),
-      [=] __device__(size_t idx) -> size_t { return idx * number_active_features; });
+  auto reduce_offset =
+      dh::MakeTransformIterator<size_t>(thrust::make_counting_iterator(0llu),
+                                        [=] __device__(size_t idx) -> size_t {
+                                          return idx * max_active_features;
+                                        });
   size_t temp_storage_bytes = 0;
   auto num_segments = out_splits.size();
   cub::DeviceSegmentedReduce::Sum(nullptr, temp_storage_bytes, feature_best_splits.data(),
@@ -386,15 +398,16 @@ void GPUHistEvaluator::CopyToHost(const std::vector<bst_node_t> &nidx) {
 }
 
 void GPUHistEvaluator::EvaluateSplits(
-    const std::vector<bst_node_t> &nidx, bst_feature_t number_active_features,
-    common::Span<const EvaluateSplitInputs> d_inputs, EvaluateSplitSharedInputs shared_inputs,
+    const std::vector<bst_node_t> &nidx, bst_feature_t max_active_features,
+    common::Span<const EvaluateSplitInputs> d_inputs,
+    EvaluateSplitSharedInputs shared_inputs,
     common::Span<GPUExpandEntry> out_entries) {
   auto evaluator = this->tree_evaluator_.template GetEvaluator<GPUTrainingParam>();
 
   dh::TemporaryArray<DeviceSplitCandidate> splits_out_storage(d_inputs.size());
   auto out_splits = dh::ToSpan(splits_out_storage);
-  this->LaunchEvaluateSplits(number_active_features, d_inputs, shared_inputs, evaluator,
-                             out_splits);
+  this->LaunchEvaluateSplits(max_active_features, d_inputs, shared_inputs,
+                             evaluator, out_splits);
 
   auto d_sorted_idx = this->SortedIdx(d_inputs.size(), shared_inputs.feature_values.size());
   auto d_entries = out_entries;

src/tree/gpu_hist/evaluate_splits.cuh

@@ -170,13 +170,18 @@ class GPUHistEvaluator {
       TreeEvaluator::SplitEvaluator<GPUTrainingParam> evaluator);
 
   // impl of evaluate splits, contains CUDA kernels so it's public
-  void LaunchEvaluateSplits(bst_feature_t number_active_features,common::Span<const EvaluateSplitInputs> d_inputs,EvaluateSplitSharedInputs shared_inputs,
-                      TreeEvaluator::SplitEvaluator<GPUTrainingParam> evaluator,
-                      common::Span<DeviceSplitCandidate> out_splits);
+  void LaunchEvaluateSplits(
+      bst_feature_t max_active_features,
+      common::Span<const EvaluateSplitInputs> d_inputs,
+      EvaluateSplitSharedInputs shared_inputs,
+      TreeEvaluator::SplitEvaluator<GPUTrainingParam> evaluator,
+      common::Span<DeviceSplitCandidate> out_splits);
   /**
    * \brief Evaluate splits for left and right nodes.
    */
-  void EvaluateSplits(const std::vector<bst_node_t> &nidx,bst_feature_t number_active_features,common::Span<const EvaluateSplitInputs> d_inputs,
+  void EvaluateSplits(const std::vector<bst_node_t> &nidx,
+                      bst_feature_t max_active_features,
+                      common::Span<const EvaluateSplitInputs> d_inputs,
                       EvaluateSplitSharedInputs shared_inputs,
                       common::Span<GPUExpandEntry> out_splits);
   /**

src/tree/updater_gpu_hist.cu

@@ -188,7 +188,8 @@ struct GPUHistMakerDevice {
   common::Span<GradientPair> gpair;
 
   dh::device_vector<int> monotone_constraints;
-  dh::device_vector<float> update_predictions;
+  // node idx for each sample
+  dh::device_vector<bst_node_t> positions;
 
   TrainParam param;
 
@@ -318,24 +319,27 @@ struct GPUHistMakerDevice {
       auto right_sampled_features = column_sampler.GetFeatureSet(tree.GetDepth(right_nidx));
       right_sampled_features->SetDevice(ctx_->gpu_id);
       common::Span<bst_feature_t> right_feature_set =
-          interaction_constraints.Query(right_sampled_features->DeviceSpan(), left_nidx);
-      h_node_inputs[i * 2] = {left_nidx, candidate.depth + 1, candidate.split.left_sum,
-                              left_feature_set, hist.GetNodeHistogram(left_nidx)};
-      h_node_inputs[i * 2 + 1] = {right_nidx, candidate.depth + 1, candidate.split.right_sum,
-                                  right_feature_set, hist.GetNodeHistogram(right_nidx)};
+          interaction_constraints.Query(right_sampled_features->DeviceSpan(),
+                                        right_nidx);
+      h_node_inputs[i * 2] = {left_nidx, candidate.depth + 1,
+                              candidate.split.left_sum, left_feature_set,
+                              hist.GetNodeHistogram(left_nidx)};
+      h_node_inputs[i * 2 + 1] = {right_nidx, candidate.depth + 1,
+                                  candidate.split.right_sum, right_feature_set,
+                                  hist.GetNodeHistogram(right_nidx)};
     }
-    bst_feature_t number_active_features = h_node_inputs[0].feature_set.size();
+    bst_feature_t max_active_features = 0;
     for (auto input : h_node_inputs) {
-      CHECK_EQ(input.feature_set.size(), number_active_features)
-          << "Current implementation assumes that the number of active features "
-             "(after sampling) in any node is the same";
+      max_active_features = std::max(max_active_features,
+                                     bst_feature_t(input.feature_set.size()));
     }
-    dh::safe_cuda(cudaMemcpyAsync(d_node_inputs.data().get(), h_node_inputs.data(),
-                                  h_node_inputs.size() * sizeof(EvaluateSplitInputs),
-                                  cudaMemcpyDefault));
+    dh::safe_cuda(cudaMemcpyAsync(
+        d_node_inputs.data().get(), h_node_inputs.data(),
+        h_node_inputs.size() * sizeof(EvaluateSplitInputs), cudaMemcpyDefault));
 
-    this->evaluator_.EvaluateSplits(nidx, number_active_features, dh::ToSpan(d_node_inputs),
-                                    shared_inputs, dh::ToSpan(entries));
+    this->evaluator_.EvaluateSplits(nidx, max_active_features,
+                                    dh::ToSpan(d_node_inputs), shared_inputs,
+                                    dh::ToSpan(entries));
     dh::safe_cuda(cudaMemcpyAsync(pinned_candidates_out.data(),
                                   entries.data().get(), sizeof(GPUExpandEntry) * entries.size(),
                                   cudaMemcpyDeviceToHost));
@@ -403,8 +407,7 @@ struct GPUHistMakerDevice {
             go_left = data.split_node.DefaultLeft();
           } else {
             if (data.split_type == FeatureType::kCategorical) {
-              go_left = common::Decision<false>(data.node_cats.Bits(), cut_value,
-                                                data.split_node.DefaultLeft());
+              go_left = common::Decision(data.node_cats.Bits(), cut_value);
             } else {
               go_left = cut_value <= data.split_node.SplitCond();
             }
@@ -424,7 +427,7 @@ struct GPUHistMakerDevice {
         LOG(FATAL) << "Current objective function can not be used with external memory.";
       }
       p_out_position->Resize(0);
-      update_predictions.clear();
+      positions.clear();
       return;
     }
 
@@ -459,8 +462,6 @@ struct GPUHistMakerDevice {
                               HostDeviceVector<bst_node_t>* p_out_position) {
     auto d_matrix = page->GetDeviceAccessor(ctx_->gpu_id);
     auto d_gpair = this->gpair;
-    update_predictions.resize(row_partitioner->GetRows().size());
-    auto d_update_predictions = dh::ToSpan(update_predictions);
     p_out_position->SetDevice(ctx_->gpu_id);
     p_out_position->Resize(row_partitioner->GetRows().size());
 
@@ -481,7 +482,7 @@ struct GPUHistMakerDevice {
           if (common::IsCat(d_feature_types, position)) {
             auto node_cats = categories.subspan(categories_segments[position].beg,
                                                 categories_segments[position].size);
-            go_left = common::Decision<false>(node_cats, element, node.DefaultLeft());
+            go_left = common::Decision(node_cats, element);
           } else {
             go_left = element <= node.SplitCond();
           }
@@ -495,32 +496,45 @@ struct GPUHistMakerDevice {
         node = d_nodes[position];
       }
 
-      d_update_predictions[row_id] = node.LeafValue();
       return position;
     };  // NOLINT
 
     auto d_out_position = p_out_position->DeviceSpan();
     row_partitioner->FinalisePosition(d_out_position, new_position_op);
 
+    auto s_position = p_out_position->ConstDeviceSpan();
+    positions.resize(s_position.size());
+    dh::safe_cuda(cudaMemcpyAsync(positions.data().get(), s_position.data(),
+                                  s_position.size_bytes(), cudaMemcpyDeviceToDevice));
+
     dh::LaunchN(row_partitioner->GetRows().size(), [=] __device__(size_t idx) {
       bst_node_t position = d_out_position[idx];
-      d_update_predictions[idx] = d_nodes[position].LeafValue();
       bool is_row_sampled = d_gpair[idx].GetHess() - .0f == 0.f;
       d_out_position[idx] = is_row_sampled ? ~position : position;
     });
   }
 
   bool UpdatePredictionCache(linalg::VectorView<float> out_preds_d, RegTree const* p_tree) {
-    if (update_predictions.empty()) {
+    if (positions.empty()) {
       return false;
     }
+
     CHECK(p_tree);
     dh::safe_cuda(cudaSetDevice(ctx_->gpu_id));
     CHECK_EQ(out_preds_d.DeviceIdx(), ctx_->gpu_id);
-    auto d_update_predictions = dh::ToSpan(update_predictions);
-    CHECK_EQ(out_preds_d.Size(), d_update_predictions.size());
-    dh::LaunchN(out_preds_d.Size(), [=] XGBOOST_DEVICE(size_t idx) mutable {
-      out_preds_d(idx) += d_update_predictions[idx];
+
+    auto d_position = dh::ToSpan(positions);
+    CHECK_EQ(out_preds_d.Size(), d_position.size());
+
+    auto const& h_nodes = p_tree->GetNodes();
+    dh::caching_device_vector<RegTree::Node> nodes(h_nodes.size());
+    dh::safe_cuda(cudaMemcpyAsync(nodes.data().get(), h_nodes.data(),
+                                  h_nodes.size() * sizeof(RegTree::Node), cudaMemcpyHostToDevice));
+    auto d_nodes = dh::ToSpan(nodes);
+    dh::LaunchN(d_position.size(), [=] XGBOOST_DEVICE(std::size_t idx) mutable {
+      bst_node_t nidx = d_position[idx];
+      auto weight = d_nodes[nidx].LeafValue();
+      out_preds_d(idx) += weight;
     });
     return true;
   }
@@ -863,6 +877,7 @@ class GPUHistMaker : public TreeUpdater {
   std::unique_ptr<GPUHistMakerDevice<GradientSumT>> maker;  // NOLINT
 
   char const* Name() const override { return "grow_gpu_hist"; }
+  bool HasNodePosition() const override { return true; }
 
  private:
   bool initialised_{false};

tests/buildkite/test-cpp-gpu.sh

@@ -4,7 +4,7 @@ set -euo pipefail
 
 source tests/buildkite/conftest.sh
 
-echo "--- Run Google Tests with CUDA, using 4 GPUs"
+echo "--- Run Google Tests with CUDA, using a GPU"
 buildkite-agent artifact download "build/testxgboost" . --step build-cuda
 chmod +x build/testxgboost
 tests/ci_build/ci_build.sh gpu nvidia-docker \
@@ -12,11 +12,12 @@ tests/ci_build/ci_build.sh gpu nvidia-docker \
   --build-arg RAPIDS_VERSION_ARG=$RAPIDS_VERSION \
   build/testxgboost
 
-echo "--- Run Google Tests with CUDA, using 4 GPUs, RMM enabled"
-rm -rfv build/
-buildkite-agent artifact download "build/testxgboost" . --step build-cuda-with-rmm
-chmod +x build/testxgboost
-tests/ci_build/ci_build.sh rmm nvidia-docker \
-  --build-arg CUDA_VERSION_ARG=$CUDA_VERSION \
-  --build-arg RAPIDS_VERSION_ARG=$RAPIDS_VERSION bash -c \
-  "source activate gpu_test && build/testxgboost --use-rmm-pool"
+# Disabled until https://github.com/dmlc/xgboost/issues/8619 is resolved
+# echo "--- Run Google Tests with CUDA, using a GPU, RMM enabled"
+# rm -rfv build/
+# buildkite-agent artifact download "build/testxgboost" . --step build-cuda-with-rmm
+# chmod +x build/testxgboost
+# tests/ci_build/ci_build.sh rmm nvidia-docker \
+#   --build-arg CUDA_VERSION_ARG=$CUDA_VERSION \
+#   --build-arg RAPIDS_VERSION_ARG=$RAPIDS_VERSION bash -c \
+#   "source activate gpu_test && build/testxgboost --use-rmm-pool"

tests/ci_build/conda_env/cpu_test.yml

@@ -36,7 +36,8 @@ dependencies:
 - cloudpickle
 - shap
 - modin
+# TODO: Replace it with pyspark>=3.4 once 3.4 released.
+# - https://ml-team-public-read.s3.us-west-2.amazonaws.com/pyspark-3.4.0.dev0.tar.gz
+- pyspark>=3.3.1
 - pip:
   - datatable
-  # TODO: Replace it with pyspark>=3.4 once 3.4 released.
-  - https://ml-team-public-read.s3.us-west-2.amazonaws.com/pyspark-3.4.0.dev0.tar.gz

tests/cpp/common/test_categorical.cc

@@ -1,11 +1,14 @@
 /*!
- * Copyright 2021 by XGBoost Contributors
+ * Copyright 2021-2022 by XGBoost Contributors
  */
 #include <gtest/gtest.h>
+#include <xgboost/json.h>
+#include <xgboost/learner.h>
 
 #include <limits>
 
 #include "../../../src/common/categorical.h"
+#include "../helpers.h"
 
 namespace xgboost {
 namespace common {
@@ -15,29 +18,76 @@ TEST(Categorical, Decision) {
 
   ASSERT_TRUE(common::InvalidCat(a));
   std::vector<uint32_t> cats(256, 0);
-  ASSERT_TRUE(Decision(cats, a, true));
+  ASSERT_TRUE(Decision(cats, a));
 
   // larger than size
   a = 256;
-  ASSERT_TRUE(Decision(cats, a, true));
+  ASSERT_TRUE(Decision(cats, a));
 
   // negative
   a = -1;
-  ASSERT_TRUE(Decision(cats, a, true));
+  ASSERT_TRUE(Decision(cats, a));
 
   CatBitField bits{cats};
   bits.Set(0);
   a = -0.5;
-  ASSERT_TRUE(Decision(cats, a, true));
+  ASSERT_TRUE(Decision(cats, a));
 
   // round toward 0
   a = 0.5;
-  ASSERT_FALSE(Decision(cats, a, true));
+  ASSERT_FALSE(Decision(cats, a));
 
   // valid
   a = 13;
   bits.Set(a);
-  ASSERT_FALSE(Decision(bits.Bits(), a, true));
+  ASSERT_FALSE(Decision(bits.Bits(), a));
+}
+
+/**
+ * Test for running inference with input category greater than the one stored in tree.
+ */
+TEST(Categorical, MinimalSet) {
+  std::size_t constexpr kRows = 256, kCols = 1, kCat = 3;
+  std::vector<FeatureType> types{FeatureType::kCategorical};
+  auto Xy =
+      RandomDataGenerator{kRows, kCols, 0.0}.Type(types).MaxCategory(kCat).GenerateDMatrix(true);
+
+  std::unique_ptr<Learner> learner{Learner::Create({Xy})};
+  learner->SetParam("max_depth", "1");
+  learner->SetParam("tree_method", "hist");
+  learner->Configure();
+  learner->UpdateOneIter(0, Xy);
+
+  Json model{Object{}};
+  learner->SaveModel(&model);
+  auto tree = model["learner"]["gradient_booster"]["model"]["trees"][0];
+  ASSERT_GE(get<I32Array const>(tree["categories"]).size(), 1);
+  auto v = get<I32Array const>(tree["categories"])[0];
+
+  HostDeviceVector<float> predt;
+  {
+    std::vector<float> data{static_cast<float>(kCat),
+                            static_cast<float>(kCat + 1), 32.0f, 33.0f, 34.0f};
+    auto test = GetDMatrixFromData(data, data.size(), kCols);
+    learner->Predict(test, false, &predt, 0, 0, false, /*pred_leaf=*/true);
+    ASSERT_EQ(predt.Size(), data.size());
+    auto const& h_predt = predt.ConstHostSpan();
+    for (auto v : h_predt) {
+      ASSERT_EQ(v, 1);  // left child of root node
+    }
+  }
+
+  {
+    std::unique_ptr<Learner> learner{Learner::Create({Xy})};
+    learner->LoadModel(model);
+    std::vector<float> data = {static_cast<float>(v)};
+    auto test = GetDMatrixFromData(data, data.size(), kCols);
+    learner->Predict(test, false, &predt, 0, 0, false, /*pred_leaf=*/true);
+    auto const& h_predt = predt.ConstHostSpan();
+    for (auto v : h_predt) {
+      ASSERT_EQ(v, 2);  // right child of root node
+    }
+  }
 }
 }  // namespace common
 }  // namespace xgboost

tests/cpp/data/test_array_interface.cc

@@ -1,10 +1,12 @@
-/*!
- * Copyright 2020-2021 by XGBoost Contributors
+/**
+ * Copyright 2020-2023 by XGBoost Contributors
  */
 #include <gtest/gtest.h>
 #include <xgboost/host_device_vector.h>
 #include "../helpers.h"
 #include "../../../src/data/array_interface.h"
+#include "dmlc/logging.h"
+#include "xgboost/json.h"
 
 namespace xgboost {
 TEST(ArrayInterface, Initialize) {
@@ -71,6 +73,14 @@ TEST(ArrayInterface, Error) {
   column["mask"]["data"] = Null{};
   common::Span<RBitField8::value_type> s_mask;
   EXPECT_THROW(ArrayInterfaceHandler::ExtractMask(column_obj, &s_mask), dmlc::Error);
+
+  get<Object>(column).erase("mask");
+  // misaligned.
+  j_data = {Json(Integer(reinterpret_cast<Integer::Int>(
+                reinterpret_cast<char const*>(storage.ConstHostPointer()) + 1))),
+            Json(Boolean(false))};
+  column["data"] = j_data;
+  EXPECT_THROW({ ArrayInterface<1> arr{column}; }, dmlc::Error);
 }
 
 TEST(ArrayInterface, GetElement) {

tests/cpp/data/test_gradient_index.cc

@@ -68,6 +68,30 @@ TEST(GradientIndex, FromCategoricalBasic) {
   }
 }
 
+TEST(GradientIndex, FromCategoricalLarge) {
+  size_t constexpr kRows = 1000, kCats = 512, kCols = 1;
+  bst_bin_t max_bins = 8;
+  auto x = GenerateRandomCategoricalSingleColumn(kRows, kCats);
+  auto m = GetDMatrixFromData(x, kRows, 1);
+  Context ctx;
+
+  auto &h_ft = m->Info().feature_types.HostVector();
+  h_ft.resize(kCols, FeatureType::kCategorical);
+
+  BatchParam p{max_bins, 0.8};
+  {
+    GHistIndexMatrix gidx(m.get(), max_bins, p.sparse_thresh, false, Context{}.Threads(), {});
+    ASSERT_TRUE(gidx.index.GetBinTypeSize() == common::kUint16BinsTypeSize);
+  }
+  {
+    for (auto const &page : m->GetBatches<GHistIndexMatrix>(p)) {
+      common::HistogramCuts cut = page.cut;
+      GHistIndexMatrix gidx{m->Info(), std::move(cut), max_bins};
+      ASSERT_EQ(gidx.MaxNumBinPerFeat(), kCats);
+    }
+  }
+}
+
 TEST(GradientIndex, PushBatch) {
   size_t constexpr kRows = 64, kCols = 4;
   bst_bin_t max_bins = 64;

tests/cpp/data/test_simple_dmatrix.cc

@@ -1,13 +1,19 @@
-// Copyright by Contributors
+/**
+ * Copyright 2016-2023 by XGBoost Contributors
+ */
 #include <xgboost/data.h>
 
-#include <array>
+#include <array>   // std::array
+#include <limits>  // std::numeric_limits
+#include <memory>  // std::unique_ptr
 
-#include "../../../src/data/adapter.h"
-#include "../../../src/data/simple_dmatrix.h"
-#include "../filesystem.h"  // dmlc::TemporaryDirectory
-#include "../helpers.h"
+#include "../../../src/data/adapter.h"         // ArrayAdapter
+#include "../../../src/data/simple_dmatrix.h"  // SimpleDMatrix
+#include "../filesystem.h"                     // dmlc::TemporaryDirectory
+#include "../helpers.h"                        // RandomDataGenerator,CreateSimpleTestData
 #include "xgboost/base.h"
+#include "xgboost/host_device_vector.h"  // HostDeviceVector
+#include "xgboost/string_view.h"         // StringView
 
 using namespace xgboost;  // NOLINT
 
@@ -298,6 +304,17 @@ TEST(SimpleDMatrix, Slice) {
   ASSERT_EQ(out->Info().num_col_, out->Info().num_col_);
   ASSERT_EQ(out->Info().num_row_, ridxs.size());
   ASSERT_EQ(out->Info().num_nonzero_, ridxs.size() * kCols);  // dense
+
+  {
+    HostDeviceVector<float> data;
+    auto arr_str = RandomDataGenerator{kRows, kCols, 0.0}.GenerateArrayInterface(&data);
+    auto adapter = data::ArrayAdapter{StringView{arr_str}};
+    auto n_threads = 2;
+    std::unique_ptr<DMatrix> p_fmat{
+        DMatrix::Create(&adapter, std::numeric_limits<float>::quiet_NaN(), n_threads, "")};
+    std::unique_ptr<DMatrix> slice{p_fmat->Slice(ridxs)};
+    ASSERT_LE(slice->Ctx()->Threads(), n_threads);
+  }
 }
 
 TEST(SimpleDMatrix, SaveLoadBinary) {

tests/cpp/tree/test_node_partition.cc

@@ -0,0 +1,24 @@
+/**
+ * Copyright 2023 by XGBoost contributors
+ */
+#include <gtest/gtest.h>
+#include <xgboost/task.h>
+#include <xgboost/tree_updater.h>
+
+namespace xgboost {
+TEST(Updater, HasNodePosition) {
+  Context ctx;
+  ObjInfo task{ObjInfo::kRegression, true, true};
+  std::unique_ptr<TreeUpdater> up{TreeUpdater::Create("grow_histmaker", &ctx, task)};
+  ASSERT_TRUE(up->HasNodePosition());
+
+  up.reset(TreeUpdater::Create("grow_quantile_histmaker", &ctx, task));
+  ASSERT_TRUE(up->HasNodePosition());
+
+#if defined(XGBOOST_USE_CUDA)
+  ctx.gpu_id = 0;
+  up.reset(TreeUpdater::Create("grow_gpu_hist", &ctx, task));
+  ASSERT_TRUE(up->HasNodePosition());
+#endif  // defined(XGBOOST_USE_CUDA)
+}
+}  // namespace xgboost

tests/python-gpu/test_device_quantile_dmatrix.py

@@ -139,3 +139,17 @@ class TestDeviceQuantileDMatrix:
             booster.predict(xgb.DMatrix(d_m.get_data())),
             atol=1e-6,
         )
+
+    def test_ltr(self) -> None:
+        import cupy as cp
+        X, y, qid, w = tm.make_ltr(100, 3, 3, 5)
+        # make sure GPU is used to run sketching.
+        cpX = cp.array(X)
+        Xy_qdm = xgb.QuantileDMatrix(cpX, y, qid=qid, weight=w)
+        Xy = xgb.DMatrix(X, y, qid=qid, weight=w)
+        xgb.train({"tree_method": "gpu_hist", "objective": "rank:ndcg"}, Xy)
+
+        from_dm = xgb.QuantileDMatrix(X, weight=w, ref=Xy)
+        from_qdm = xgb.QuantileDMatrix(X, weight=w, ref=Xy_qdm)
+
+        assert tm.predictor_equal(from_qdm, from_dm)

tests/python-gpu/test_gpu_interaction_constraints.py

@@ -1,8 +1,14 @@
-import numpy as np
 import sys
+
+import numpy as np
+import pandas as pd
+
+import xgboost as xgb
+
 sys.path.append("tests/python")
 # Don't import the test class, otherwise they will run twice.
 import test_interaction_constraints as test_ic  # noqa
+
 rng = np.random.RandomState(1994)
 
 
@@ -10,7 +16,34 @@ class TestGPUInteractionConstraints:
     cputest = test_ic.TestInteractionConstraints()
 
     def test_interaction_constraints(self):
-        self.cputest.run_interaction_constraints(tree_method='gpu_hist')
+        self.cputest.run_interaction_constraints(tree_method="gpu_hist")
 
     def test_training_accuracy(self):
-        self.cputest.training_accuracy(tree_method='gpu_hist')
+        self.cputest.training_accuracy(tree_method="gpu_hist")
+
+    # case where different number of features can occur in the evaluator
+    def test_issue_8730(self):
+        X = pd.DataFrame(
+            zip(range(0, 100), range(200, 300), range(300, 400), range(400, 500)),
+            columns=["A", "B", "C", "D"],
+        )
+        y = np.array([*([0] * 50), *([1] * 50)])
+        dm = xgb.DMatrix(X, label=y)
+
+        params = {
+            "eta": 0.16095019509249486,
+            "min_child_weight": 1,
+            "subsample": 0.688567929338029,
+            "colsample_bynode": 0.7,
+            "gamma": 5.666579817418348e-06,
+            "lambda": 0.14943712232059794,
+            "grow_policy": "depthwise",
+            "max_depth": 3,
+            "tree_method": "gpu_hist",
+            "interaction_constraints": [["A", "B"], ["B", "D", "C"], ["C", "D"]],
+            "objective": "count:poisson",
+            "eval_metric": "poisson-nloglik",
+            "verbosity": 0,
+        }
+
+        xgb.train(params, dm, num_boost_round=100)

tests/python-gpu/test_gpu_prediction.py

@@ -338,13 +338,21 @@ class TestGPUPredict:
     @given(predict_parameter_strategy, tm.dataset_strategy)
     @settings(deadline=None, max_examples=20, print_blob=True)
     def test_predict_leaf_gbtree(self, param, dataset):
+        # Unsupported for random forest
+        if param.get("num_parallel_tree", 1) > 1 and dataset.name.endswith("-l1"):
+            return
+
         param['booster'] = 'gbtree'
         param['tree_method'] = 'gpu_hist'
         self.run_predict_leaf_booster(param, 10, dataset)
 
     @given(predict_parameter_strategy, tm.dataset_strategy)
     @settings(deadline=None, max_examples=20, print_blob=True)
-    def test_predict_leaf_dart(self, param, dataset):
+    def test_predict_leaf_dart(self, param: dict, dataset: tm.TestDataset) -> None:
+        # Unsupported for random forest
+        if param.get("num_parallel_tree", 1) > 1 and dataset.name.endswith("-l1"):
+            return
+
         param['booster'] = 'dart'
         param['tree_method'] = 'gpu_hist'
         self.run_predict_leaf_booster(param, 10, dataset)

tests/python/test_dmatrix.py

@@ -326,7 +326,7 @@ class TestDMatrix:
         nrow = 100
         ncol = 1000
         x = rand(nrow, ncol, density=0.0005, format='csr', random_state=rng)
-        assert x.indices.max() < ncol - 1
+        assert x.indices.max() < ncol
         x.data[:] = 1
         dtrain = xgb.DMatrix(x, label=rng.binomial(1, 0.3, nrow))
         assert (dtrain.num_row(), dtrain.num_col()) == (nrow, ncol)

tests/python/test_quantile_dmatrix.py

@@ -9,7 +9,9 @@ from testing import (
     make_batches,
     make_batches_sparse,
     make_categorical,
+    make_ltr,
     make_sparse_regression,
+    predictor_equal,
 )
 
 import xgboost as xgb
@@ -218,6 +220,16 @@ class TestQuantileDMatrix:
         b = booster.predict(qXy)
         np.testing.assert_allclose(a, b)
 
+    def test_ltr(self) -> None:
+        X, y, qid, w = make_ltr(100, 3, 3, 5)
+        Xy_qdm = xgb.QuantileDMatrix(X, y, qid=qid, weight=w)
+        Xy = xgb.DMatrix(X, y, qid=qid, weight=w)
+        xgb.train({"tree_method": "hist", "objective": "rank:ndcg"}, Xy)
+
+        from_qdm = xgb.QuantileDMatrix(X, weight=w, ref=Xy_qdm)
+        from_dm = xgb.QuantileDMatrix(X, weight=w, ref=Xy)
+        assert predictor_equal(from_qdm, from_dm)
+
     # we don't test empty Quantile DMatrix in single node construction.
     @given(
         strategies.integers(1, 1000),

tests/python/test_spark/test_spark_local.py

@@ -41,6 +41,16 @@ logging.getLogger("py4j").setLevel(logging.INFO)
 pytestmark = testing.timeout(60)
 
 
+def no_sparse_unwrap():
+    try:
+        from pyspark.sql.functions import unwrap_udt
+
+    except ImportError:
+        return {"reason": "PySpark<3.4", "condition": True}
+
+    return {"reason": "PySpark<3.4", "condition": False}
+
+
 class XgboostLocalTest(SparkTestCase):
     def setUp(self):
         logging.getLogger().setLevel("INFO")
@@ -985,6 +995,7 @@ class XgboostLocalTest(SparkTestCase):
         model = classifier.fit(self.cls_df_train)
         model.transform(self.cls_df_test).collect()
 
+    @pytest.mark.skipif(**no_sparse_unwrap())
     def test_regressor_with_sparse_optim(self):
         regressor = SparkXGBRegressor(missing=0.0)
         model = regressor.fit(self.reg_df_sparse_train)
@@ -1001,6 +1012,7 @@ class XgboostLocalTest(SparkTestCase):
         for row1, row2 in zip(pred_result, pred_result2):
             self.assertTrue(np.isclose(row1.prediction, row2.prediction, atol=1e-3))
 
+    @pytest.mark.skipif(**no_sparse_unwrap())
     def test_classifier_with_sparse_optim(self):
         cls = SparkXGBClassifier(missing=0.0)
         model = cls.fit(self.cls_df_sparse_train)

tests/python/test_updaters.py

@@ -458,6 +458,22 @@ class TestTreeMethod:
         config_0 = json.loads(booster_0.save_config())
         np.testing.assert_allclose(get_score(config_0), get_score(config_1) + 1)
 
+        evals_result: Dict[str, Dict[str, list]] = {}
+        xgb.train(
+            {
+                "tree_method": tree_method,
+                "objective": "reg:absoluteerror",
+                "subsample": 0.8
+            },
+            Xy,
+            num_boost_round=10,
+            evals=[(Xy, "Train")],
+            evals_result=evals_result,
+        )
+        mae = evals_result["Train"]["mae"]
+        assert mae[-1] < 20.0
+        assert tm.non_increasing(mae)
+
     @pytest.mark.skipif(**tm.no_sklearn())
     @pytest.mark.parametrize(
         "tree_method,weighted", [

tests/python/test_with_pandas.py

@@ -112,7 +112,6 @@ class TestPandas:
 
         # test Index as columns
         df = pd.DataFrame([[1, 1.1], [2, 2.2]], columns=pd.Index([1, 2]))
-        print(df.columns, isinstance(df.columns, pd.Index))
         Xy = xgb.DMatrix(df)
         np.testing.assert_equal(np.array(Xy.feature_names), np.array(["1", "2"]))
 

tests/python/test_with_shap.py

@@ -4,7 +4,7 @@ import pytest
 
 try:
     import shap
-except ImportError:
+except Exception:
     shap = None
     pass
 

tests/python/test_with_sklearn.py

@@ -2,6 +2,7 @@ import collections
 import importlib.util
 import json
 import os
+import pickle
 import random
 import tempfile
 from typing import Callable, Optional
@@ -636,26 +637,74 @@ def test_sklearn_n_jobs():
 
 def test_parameters_access():
     from sklearn import datasets
-    params = {'updater': 'grow_gpu_hist', 'subsample': .5, 'n_jobs': -1}
+
+    params = {"updater": "grow_gpu_hist", "subsample": 0.5, "n_jobs": -1}
     clf = xgb.XGBClassifier(n_estimators=1000, **params)
-    assert clf.get_params()['updater'] == 'grow_gpu_hist'
-    assert clf.get_params()['subsample'] == .5
-    assert clf.get_params()['n_estimators'] == 1000
+    assert clf.get_params()["updater"] == "grow_gpu_hist"
+    assert clf.get_params()["subsample"] == 0.5
+    assert clf.get_params()["n_estimators"] == 1000
 
     clf = xgb.XGBClassifier(n_estimators=1, nthread=4)
     X, y = datasets.load_iris(return_X_y=True)
     clf.fit(X, y)
 
     config = json.loads(clf.get_booster().save_config())
-    assert int(config['learner']['generic_param']['nthread']) == 4
+    assert int(config["learner"]["generic_param"]["nthread"]) == 4
 
     clf.set_params(nthread=16)
     config = json.loads(clf.get_booster().save_config())
-    assert int(config['learner']['generic_param']['nthread']) == 16
+    assert int(config["learner"]["generic_param"]["nthread"]) == 16
 
     clf.predict(X)
     config = json.loads(clf.get_booster().save_config())
-    assert int(config['learner']['generic_param']['nthread']) == 16
+    assert int(config["learner"]["generic_param"]["nthread"]) == 16
+
+    clf = xgb.XGBClassifier(n_estimators=2)
+    assert clf.tree_method is None
+    assert clf.get_params()["tree_method"] is None
+    clf.fit(X, y)
+    assert clf.get_params()["tree_method"] is None
+
+    def save_load(clf: xgb.XGBClassifier) -> xgb.XGBClassifier:
+        with tempfile.TemporaryDirectory() as tmpdir:
+            path = os.path.join(tmpdir, "model.json")
+            clf.save_model(path)
+            clf = xgb.XGBClassifier()
+            clf.load_model(path)
+        return clf
+
+    def get_tm(clf: xgb.XGBClassifier) -> str:
+        tm = json.loads(clf.get_booster().save_config())["learner"]["gradient_booster"][
+            "gbtree_train_param"
+        ]["tree_method"]
+        return tm
+
+    assert get_tm(clf) == "exact"
+
+    clf = pickle.loads(pickle.dumps(clf))
+
+    assert clf.tree_method is None
+    assert clf.n_estimators == 2
+    assert clf.get_params()["tree_method"] is None
+    assert clf.get_params()["n_estimators"] == 2
+    assert get_tm(clf) == "exact"  # preserved for pickle
+
+    clf = save_load(clf)
+
+    assert clf.tree_method is None
+    assert clf.n_estimators == 2
+    assert clf.get_params()["tree_method"] is None
+    assert clf.get_params()["n_estimators"] == 2
+    assert get_tm(clf) == "auto"  # discarded for save/load_model
+
+    clf.set_params(tree_method="hist")
+    assert clf.get_params()["tree_method"] == "hist"
+    clf = pickle.loads(pickle.dumps(clf))
+    assert clf.get_params()["tree_method"] == "hist"
+    clf = save_load(clf)
+    # FIXME(jiamingy): We should remove this behavior once we remove parameters
+    # serialization for skl save/load_model.
+    assert clf.get_params()["tree_method"] == "hist"
 
 
 def test_kwargs_error():
@@ -695,13 +744,19 @@ def test_sklearn_clone():
 
 def test_sklearn_get_default_params():
     from sklearn.datasets import load_digits
+
     digits_2class = load_digits(n_class=2)
-    X = digits_2class['data']
-    y = digits_2class['target']
+    X = digits_2class["data"]
+    y = digits_2class["target"]
     cls = xgb.XGBClassifier()
-    assert cls.get_params()['base_score'] is None
+    assert cls.get_params()["base_score"] is None
     cls.fit(X[:4, ...], y[:4, ...])
-    assert cls.get_params()['base_score'] is not None
+    base_score = float(
+        json.loads(cls.get_booster().save_config())["learner"]["learner_model_param"][
+            "base_score"
+        ]
+    )
+    np.testing.assert_equal(base_score, 0.5)
 
 
 def run_validation_weights(model):
@@ -1029,9 +1084,9 @@ def test_pandas_input():
 
     clf_isotonic = CalibratedClassifierCV(model, cv="prefit", method="isotonic")
     clf_isotonic.fit(train, target)
-    assert isinstance(
-        clf_isotonic.calibrated_classifiers_[0].base_estimator, xgb.XGBClassifier
-    )
+    clf = clf_isotonic.calibrated_classifiers_[0]
+    est = clf.estimator if hasattr(clf, "estimator") else clf.base_estimator
+    assert isinstance(est, xgb.XGBClassifier)
     np.testing.assert_allclose(np.array(clf_isotonic.classes_), np.array([0, 1]))
 
 

tests/python/testing.py

@@ -466,7 +466,22 @@ def make_categorical(
     return df, label
 
 
-def _cat_sampled_from():
+def make_ltr(
+    n_samples: int, n_features: int, n_query_groups: int, max_rel: int
+) -> Tuple[np.ndarray, np.ndarray, np.ndarray, np.ndarray]:
+    """Make a dataset for testing LTR."""
+    rng = np.random.default_rng(1994)
+    X = rng.normal(0, 1.0, size=n_samples * n_features).reshape(n_samples, n_features)
+    y = rng.integers(0, max_rel, size=n_samples)
+    qid = rng.integers(0, n_query_groups, size=n_samples)
+    w = rng.normal(0, 1.0, size=n_query_groups)
+    w -= np.min(w)
+    w /= np.max(w)
+    qid = np.sort(qid)
+    return X, y, qid, w
+
+
+def _cat_sampled_from() -> strategies.SearchStrategy:
     @strategies.composite
     def _make_cat(draw):
         n_samples = draw(strategies.integers(2, 512))
@@ -775,6 +790,19 @@ class DirectoryExcursion:
                 os.remove(f)
 
 
+def predictor_equal(lhs: xgb.DMatrix, rhs: xgb.DMatrix) -> bool:
+    """Assert whether two DMatrices contain the same predictors."""
+    lcsr = lhs.get_data()
+    rcsr = rhs.get_data()
+    return all(
+        (
+            np.array_equal(lcsr.data, rcsr.data),
+            np.array_equal(lcsr.indices, rcsr.indices),
+            np.array_equal(lcsr.indptr, rcsr.indptr),
+        )
+    )
+
+
 @contextmanager
 def captured_output():
     """Reassign stdout temporarily in order to test printed statements