Bump version to 1.7.6 (#9305)

* [backport] Fix monotone constraints on CPU. (#9122)

.github/workflows/main.yml

@@ -148,66 +148,13 @@ jobs:
       run: |
         LINT_LANG=cpp make lint
 
-  doxygen:
-    runs-on: ubuntu-latest
-    name: Generate C/C++ API doc using Doxygen
-    steps:
-    - uses: actions/checkout@v2
-      with:
-        submodules: 'true'
-    - uses: actions/setup-python@v2
-      with:
-        python-version: "3.8"
-        architecture: 'x64'
-    - name: Install system packages
-      run: |
-        sudo apt-get install -y --no-install-recommends doxygen graphviz ninja-build
-        python -m pip install wheel setuptools
-        python -m pip install awscli
-    - name: Run Doxygen
-      run: |
-        mkdir build
-        cd build
-        cmake .. -DBUILD_C_DOC=ON -GNinja
-        ninja -v doc_doxygen
-    - name: Extract branch name
-      shell: bash
-      run: echo "##[set-output name=branch;]$(echo ${GITHUB_REF#refs/heads/})"
-      id: extract_branch
-      if: github.ref == 'refs/heads/master' || contains(github.ref, 'refs/heads/release_')
-    - name: Publish
-      run: |
-        cd build/
-        tar cvjf ${{ steps.extract_branch.outputs.branch }}.tar.bz2 doc_doxygen/
-        python -m awscli s3 cp ./${{ steps.extract_branch.outputs.branch }}.tar.bz2 s3://xgboost-docs/doxygen/ --acl public-read
-      if: github.ref == 'refs/heads/master' || contains(github.ref, 'refs/heads/release_')
-      env:
-        AWS_ACCESS_KEY_ID: ${{ secrets.AWS_ACCESS_KEY_ID_IAM_S3_UPLOADER }}
-        AWS_SECRET_ACCESS_KEY: ${{ secrets.AWS_SECRET_ACCESS_KEY_IAM_S3_UPLOADER }}
-
-  sphinx:
-    runs-on: ubuntu-latest
-    name: Build docs using Sphinx
-    steps:
-    - uses: actions/checkout@v2
-      with:
-        submodules: 'true'
-    - uses: actions/setup-python@v2
-      with:
-        python-version: "3.8"
-        architecture: 'x64'
-    - name: Install system packages
-      run: |
-        sudo apt-get install -y --no-install-recommends graphviz
-        python -m pip install wheel setuptools
-        python -m pip install -r doc/requirements.txt
-    - name: Extract branch name
-      shell: bash
-      run: echo "##[set-output name=branch;]$(echo ${GITHUB_REF#refs/heads/})"
-      id: extract_branch
-      if: github.ref == 'refs/heads/master' || contains(github.ref, 'refs/heads/release_')
-    - name: Run Sphinx
-      run: |
-        make -C doc html
-      env:
-        SPHINX_GIT_BRANCH: ${{ steps.extract_branch.outputs.branch }}
+        python3 dmlc-core/scripts/lint.py --exclude_path \
+            python-package/xgboost/dmlc-core \
+            python-package/xgboost/include \
+            python-package/xgboost/lib \
+            python-package/xgboost/rabit \
+            python-package/xgboost/src \
+            --pylint-rc python-package/.pylintrc \
+            xgboost \
+            cpp \
+            include src python-package

CMakeLists.txt

@@ -1,5 +1,5 @@
-cmake_minimum_required(VERSION 3.14 FATAL_ERROR)
-project(xgboost LANGUAGES CXX C VERSION 1.7.3)
+cmake_minimum_required(VERSION 3.18 FATAL_ERROR)
+project(xgboost LANGUAGES CXX C VERSION 1.7.6)
 include(cmake/Utils.cmake)
 list(APPEND CMAKE_MODULE_PATH "${xgboost_SOURCE_DIR}/cmake/modules")
 cmake_policy(SET CMP0022 NEW)
@@ -168,9 +168,6 @@ find_package(Threads REQUIRED)
 
 if (USE_OPENMP)
   if (APPLE)
-    # Require CMake 3.16+ on Mac OSX, as previous versions of CMake had trouble locating
-    # OpenMP on Mac. See https://github.com/dmlc/xgboost/pull/5146#issuecomment-568312706
-    cmake_minimum_required(VERSION 3.16)
     find_package(OpenMP)
     if (NOT OpenMP_FOUND)
       # Try again with extra path info; required for libomp 15+ from Homebrew

R-package/CMakeLists.txt

@@ -31,7 +31,7 @@ if (USE_OPENMP)
 endif (USE_OPENMP)
 set_target_properties(
   xgboost-r PROPERTIES
-  CXX_STANDARD 14
+  CXX_STANDARD 17
   CXX_STANDARD_REQUIRED ON
   POSITION_INDEPENDENT_CODE ON)
 

R-package/DESCRIPTION

@@ -1,8 +1,8 @@
 Package: xgboost
 Type: Package
 Title: Extreme Gradient Boosting
-Version: 1.7.3.1
-Date: 2023-01-06
+Version: 1.7.6.1
+Date: 2023-06-16
 Authors@R: c(
   person("Tianqi", "Chen", role = c("aut"),
          email = "tianqi.tchen@gmail.com"),
@@ -66,5 +66,6 @@ Imports:
     methods,
     data.table (>= 1.9.6),
     jsonlite (>= 1.0),
-RoxygenNote: 7.2.2
-SystemRequirements: GNU make, C++14
+RoxygenNote: 7.2.3
+Encoding: UTF-8
+SystemRequirements: GNU make, C++17

R-package/LICENSE

@@ -1,9 +1,9 @@
-Copyright (c) 2014 by Tianqi Chen and Contributors 
+Copyright (c) 2014-2023, Tianqi Chen and XBGoost Contributors
 
 Licensed under the Apache License, Version 2.0 (the "License");
 you may not use this file except in compliance with the License.
 You may obtain a copy of the License at
-    
+
    http://www.apache.org/licenses/LICENSE-2.0
 
 Unless required by applicable law or agreed to in writing, software

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/R/xgb.plot.tree.R

@@ -34,7 +34,7 @@
 #' The branches that also used for missing values are marked as bold
 #' (as in "carrying extra capacity").
 #'
-#' This function uses \href{http://www.graphviz.org/}{GraphViz} as a backend of DiagrammeR.
+#' This function uses \href{https://www.graphviz.org/}{GraphViz} as a backend of DiagrammeR.
 #'
 #' @return
 #'

R-package/configure.ac

@@ -2,10 +2,25 @@
 
 AC_PREREQ(2.69)
 
-AC_INIT([xgboost],[1.7.3],[],[xgboost],[])
+AC_INIT([xgboost],[1.7.6],[],[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
+
+CXX17=`"${R_HOME}/bin/R" CMD config CXX17`
+CXX17STD=`"${R_HOME}/bin/R" CMD config CXX17STD`
+CXX="${CXX17} ${CXX17STD}"
+CXXFLAGS=`"${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/man/xgb.plot.tree.Rd

@@ -67,7 +67,7 @@ The "Yes" branches are marked by the "< split_value" label.
 The branches that also used for missing values are marked as bold
 (as in "carrying extra capacity").
 
-This function uses \href{http://www.graphviz.org/}{GraphViz} as a backend of DiagrammeR.
+This function uses \href{https://www.graphviz.org/}{GraphViz} as a backend of DiagrammeR.
 }
 \examples{
 data(agaricus.train, package='xgboost')

R-package/src/Makevars.in

@@ -3,7 +3,7 @@ PKGROOT=../../
 ENABLE_STD_THREAD=1
 # _*_ mode: Makefile; _*_
 
-CXX_STD = CXX14
+CXX_STD = CXX17
 
 XGB_RFLAGS = -DXGBOOST_STRICT_R_MODE=1 -DDMLC_LOG_BEFORE_THROW=0\
            -DDMLC_ENABLE_STD_THREAD=$(ENABLE_STD_THREAD) -DDMLC_DISABLE_STDIN=1\
@@ -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

@@ -3,7 +3,7 @@ PKGROOT=../../
 ENABLE_STD_THREAD=0
 # _*_ mode: Makefile; _*_
 
-CXX_STD = CXX14
+CXX_STD = CXX17
 
 XGB_RFLAGS = -DXGBOOST_STRICT_R_MODE=1 -DDMLC_LOG_BEFORE_THROW=0\
            -DDMLC_ENABLE_STD_THREAD=$(ENABLE_STD_THREAD) -DDMLC_DISABLE_STDIN=1\
@@ -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);
-  }
-}

cmake/Utils.cmake

@@ -178,17 +178,10 @@ function(xgboost_set_cuda_flags target)
       $<$<COMPILE_LANGUAGE:CUDA>:-Xcompiler=/utf-8>)
   endif (MSVC)
 
-  if (PLUGIN_RMM)
-    set_target_properties(${target} PROPERTIES
-      CUDA_STANDARD 17
-      CUDA_STANDARD_REQUIRED ON
-      CUDA_SEPARABLE_COMPILATION OFF)
-  else ()
-    set_target_properties(${target} PROPERTIES
-      CUDA_STANDARD 14
-      CUDA_STANDARD_REQUIRED ON
-      CUDA_SEPARABLE_COMPILATION OFF)
-  endif (PLUGIN_RMM)
+  set_target_properties(${target} PROPERTIES
+    CUDA_STANDARD 17
+    CUDA_STANDARD_REQUIRED ON
+    CUDA_SEPARABLE_COMPILATION OFF)
 endfunction(xgboost_set_cuda_flags)
 
 macro(xgboost_link_nccl target)
@@ -205,17 +198,10 @@ endmacro(xgboost_link_nccl)
 
 # compile options
 macro(xgboost_target_properties target)
-  if (PLUGIN_RMM)
-    set_target_properties(${target} PROPERTIES
-      CXX_STANDARD 17
-      CXX_STANDARD_REQUIRED ON
-      POSITION_INDEPENDENT_CODE ON)
-  else ()
-    set_target_properties(${target} PROPERTIES
-      CXX_STANDARD 14
-      CXX_STANDARD_REQUIRED ON
-      POSITION_INDEPENDENT_CODE ON)
-  endif (PLUGIN_RMM)
+  set_target_properties(${target} PROPERTIES
+    CXX_STANDARD 17
+    CXX_STANDARD_REQUIRED ON
+    POSITION_INDEPENDENT_CODE ON)
 
   if (HIDE_CXX_SYMBOLS)
     #-- Hide all C++ symbols

demo/c-api/CMakeLists.txt

@@ -1,4 +1,4 @@
-cmake_minimum_required(VERSION 3.13)
+cmake_minimum_required(VERSION 3.18)
 project(xgboost-c-examples)
 
 add_subdirectory(basic)

demo/c-api/external-memory/CMakeLists.txt

@@ -1,4 +1,4 @@
-cmake_minimum_required(VERSION 3.13)
+cmake_minimum_required(VERSION 3.18)
 project(external-memory-demo LANGUAGES C VERSION 0.0.1)
 
 find_package(xgboost REQUIRED)

demo/c-api/inference/CMakeLists.txt

@@ -1,4 +1,4 @@
-cmake_minimum_required(VERSION 3.13)
+cmake_minimum_required(VERSION 3.18)
 project(inference-demo LANGUAGES C VERSION 0.0.1)
 find_package(xgboost REQUIRED)
 

doc/c++.rst

@@ -8,5 +8,5 @@ As a result it's changing quite often and we don't maintain its stability.  Alon
 plugin system (see ``plugin/example`` in XGBoost's source tree), users can utilize some
 existing c++ headers for gaining more access to the internal of XGBoost.
 
-* `C++ interface documentation (latest master branch) <https://xgboost.readthedocs.io/en/latest/dev/files.html>`_
+* `C++ interface documentation (latest master branch) <./dev/files.html>`_
 * `C++ interface documentation (last stable release) <https://xgboost.readthedocs.io/en/stable/dev/files.html>`_

doc/c.rst

@@ -10,7 +10,7 @@ simply look at function comments in ``include/xgboost/c_api.h``. The reference i
 to sphinx with the help of breathe, which doesn't contain links to examples but might be
 easier to read. For the original doxygen pages please visit:
 
-* `C API documentation (latest master branch) <https://xgboost.readthedocs.io/en/latest/dev/c__api_8h.html>`_
+* `C API documentation (latest master branch) <./dev/c__api_8h.html>`_
 * `C API documentation (last stable release) <https://xgboost.readthedocs.io/en/stable/dev/c__api_8h.html>`_
 
 ***************

doc/conf.py

@@ -11,54 +11,107 @@
 #
 # All configuration values have a default; values that are commented out
 # serve to show the default.
-from subprocess import call
-from sh.contrib import git
-import urllib.request
-from urllib.error import HTTPError
-import sys
-import re
 import os
+import re
+import shutil
 import subprocess
+import sys
+import tarfile
+import urllib.request
+import warnings
+from urllib.error import HTTPError
 
-git_branch = os.getenv('SPHINX_GIT_BRANCH', default=None)
+from sh.contrib import git
+
+CURR_PATH = os.path.dirname(os.path.abspath(os.path.expanduser(__file__)))
+PROJECT_ROOT = os.path.normpath(os.path.join(CURR_PATH, os.path.pardir))
+TMP_DIR = os.path.join(CURR_PATH, "tmp")
+DOX_DIR = "doxygen"
+
+
+def run_doxygen():
+    """Run the doxygen make command in the designated folder."""
+    curdir = os.path.normpath(os.path.abspath(os.path.curdir))
+    if os.path.exists(TMP_DIR):
+        print(f"Delete directory {TMP_DIR}")
+        shutil.rmtree(TMP_DIR)
+    else:
+        print(f"Create directory {TMP_DIR}")
+        os.mkdir(TMP_DIR)
+    try:
+        os.chdir(PROJECT_ROOT)
+        if not os.path.exists(DOX_DIR):
+            os.mkdir(DOX_DIR)
+        os.chdir(os.path.join(PROJECT_ROOT, DOX_DIR))
+        print(
+            "Build doxygen at {}".format(
+                os.path.join(PROJECT_ROOT, DOX_DIR, "doc_doxygen")
+            )
+        )
+        subprocess.check_call(["cmake", "..", "-DBUILD_C_DOC=ON", "-GNinja"])
+        subprocess.check_call(["ninja", "doc_doxygen"])
+
+        src = os.path.join(PROJECT_ROOT, DOX_DIR, "doc_doxygen", "html")
+        dest = os.path.join(TMP_DIR, "dev")
+        print(f"Copy directory {src} -> {dest}")
+        shutil.copytree(src, dest)
+    except OSError as e:
+        sys.stderr.write("doxygen execution failed: %s" % e)
+    finally:
+        os.chdir(curdir)
+
+
+def is_readthedocs_build():
+    if os.environ.get("READTHEDOCS", None) == "True":
+        return True
+    warnings.warn(
+        "Skipping Doxygen build... You won't have documentation for C/C++ functions. "
+        "Set environment variable READTHEDOCS=True if you want to build Doxygen. "
+        "(If you do opt in, make sure to install Doxygen, Graphviz, CMake, and C++ compiler "
+        "on your system.)"
+    )
+    return False
+
+
+if is_readthedocs_build():
+    run_doxygen()
+
+
+git_branch = os.getenv("SPHINX_GIT_BRANCH", default=None)
 if not git_branch:
     # If SPHINX_GIT_BRANCH environment variable is not given, run git
     # to determine branch name
     git_branch = [
-        re.sub(r'origin/', '', x.lstrip(' ')) for x in str(
-            git.branch('-r', '--contains', 'HEAD')).rstrip('\n').split('\n')
+        re.sub(r"origin/", "", x.lstrip(" "))
+        for x in str(git.branch("-r", "--contains", "HEAD")).rstrip("\n").split("\n")
     ]
-    git_branch = [x for x in git_branch if 'HEAD' not in x]
+    git_branch = [x for x in git_branch if "HEAD" not in x]
 else:
     git_branch = [git_branch]
-print('git_branch = {}'.format(git_branch[0]))
+print("git_branch = {}".format(git_branch[0]))
 
 try:
     filename, _ = urllib.request.urlretrieve(
-        'https://s3-us-west-2.amazonaws.com/xgboost-docs/{}.tar.bz2'.format(
-            git_branch[0]))
-    call(
-        'if [ -d tmp ]; then rm -rf tmp; fi; mkdir -p tmp/jvm; cd tmp/jvm; tar xvf {}'
-        .format(filename),
-        shell=True)
+        f"https://s3-us-west-2.amazonaws.com/xgboost-docs/{git_branch[0]}.tar.bz2"
+    )
+    if not os.path.exists(TMP_DIR):
+        print(f"Create directory {TMP_DIR}")
+        os.mkdir(TMP_DIR)
+    jvm_doc_dir = os.path.join(TMP_DIR, "jvm")
+    if os.path.exists(jvm_doc_dir):
+        print(f"Delete directory {jvm_doc_dir}")
+        shutil.rmtree(jvm_doc_dir)
+    print(f"Create directory {jvm_doc_dir}")
+    os.mkdir(jvm_doc_dir)
+
+    with tarfile.open(filename, "r:bz2") as t:
+        t.extractall(jvm_doc_dir)
 except HTTPError:
-    print('JVM doc not found. Skipping...')
-try:
-    filename, _ = urllib.request.urlretrieve(
-        'https://s3-us-west-2.amazonaws.com/xgboost-docs/doxygen/{}.tar.bz2'.
-        format(git_branch[0]))
-    call(
-        'mkdir -p tmp/dev; cd tmp/dev; tar xvf {}; mv doc_doxygen/html/* .; rm -rf doc_doxygen'
-        .format(filename),
-        shell=True)
-except HTTPError:
-    print('C API doc not found. Skipping...')
+    print("JVM doc not found. Skipping...")
 
 # If extensions (or modules to document with autodoc) are in another directory,
 # add these directories to sys.path here. If the directory is relative to the
 # documentation root, use os.path.abspath to make it absolute, like shown here.
-CURR_PATH = os.path.dirname(os.path.abspath(os.path.expanduser(__file__)))
-PROJECT_ROOT = os.path.normpath(os.path.join(CURR_PATH, os.path.pardir))
 libpath = os.path.join(PROJECT_ROOT, "python-package/")
 sys.path.insert(0, libpath)
 sys.path.insert(0, CURR_PATH)
@@ -81,50 +134,56 @@ release = xgboost.__version__
 # Add any Sphinx extension module names here, as strings. They can be
 # extensions coming with Sphinx (named 'sphinx.ext.*') or your custom ones
 extensions = [
-    'matplotlib.sphinxext.plot_directive',
-    'sphinx.ext.autodoc',
-    'sphinx.ext.napoleon',
-    'sphinx.ext.mathjax',
-    'sphinx.ext.intersphinx',
+    "matplotlib.sphinxext.plot_directive",
+    "sphinxcontrib.jquery",
+    "sphinx.ext.autodoc",
+    "sphinx.ext.napoleon",
+    "sphinx.ext.mathjax",
+    "sphinx.ext.intersphinx",
     "sphinx_gallery.gen_gallery",
-    'breathe',
-    'recommonmark'
+    "breathe",
+    "recommonmark",
 ]
 
 sphinx_gallery_conf = {
     # path to your example scripts
     "examples_dirs": ["../demo/guide-python", "../demo/dask", "../demo/aft_survival"],
     # path to where to save gallery generated output
-    "gallery_dirs": ["python/examples", "python/dask-examples", "python/survival-examples"],
+    "gallery_dirs": [
+        "python/examples",
+        "python/dask-examples",
+        "python/survival-examples",
+    ],
     "matplotlib_animations": True,
 }
 
 autodoc_typehints = "description"
 
-graphviz_output_format = 'png'
-plot_formats = [('svg', 300), ('png', 100), ('hires.png', 300)]
+graphviz_output_format = "png"
+plot_formats = [("svg", 300), ("png", 100), ("hires.png", 300)]
 plot_html_show_source_link = False
 plot_html_show_formats = False
 
 # Breathe extension variables
-DOX_DIR = "doxygen"
-breathe_projects = {
-    "xgboost": os.path.join(PROJECT_ROOT, DOX_DIR, "doc_doxygen/xml")
-}
+breathe_projects = {}
+if is_readthedocs_build():
+    breathe_projects = {
+        "xgboost": os.path.join(PROJECT_ROOT, DOX_DIR, "doc_doxygen/xml")
+    }
 breathe_default_project = "xgboost"
 
 # Add any paths that contain templates here, relative to this directory.
-templates_path = ['_templates']
+templates_path = ["_templates"]
 
 # The suffix(es) of source filenames.
 # You can specify multiple suffix as a list of string:
-source_suffix = ['.rst', '.md']
+source_suffix = [".rst", ".md"]
 
 # The encoding of source files.
 # source_encoding = 'utf-8-sig'
 
 # The master toctree document.
-master_doc = 'index'
+master_doc = "index"
 
 # The language for content autogenerated by Sphinx. Refer to documentation
 # for a list of supported languages.
@@ -133,7 +192,7 @@ master_doc = 'index'
 # Usually you set "language" from the command line for these cases.
 language = "en"
 
-autoclass_content = 'both'
+autoclass_content = "both"
 
 # There are two options for replacing |today|: either, you set today to some
 # non-false value, then it is used:
@@ -143,8 +202,10 @@ autoclass_content = 'both'
 
 # List of patterns, relative to source directory, that match files and
 # directories to ignore when looking for source files.
-exclude_patterns = ['_build']
-html_extra_path = ['./tmp']
+exclude_patterns = ["_build"]
+html_extra_path = []
+if is_readthedocs_build():
+    html_extra_path = [TMP_DIR]
 
 # The reST default role (used for this markup: `text`) to use for all
 # documents.
@@ -162,7 +223,7 @@ html_extra_path = ['./tmp']
 # show_authors = False
 
 # The name of the Pygments (syntax highlighting) style to use.
-pygments_style = 'sphinx'
+pygments_style = "sphinx"
 
 # A list of ignored prefixes for module index sorting.
 # modindex_common_prefix = []
@@ -185,27 +246,24 @@ html_logo = "https://raw.githubusercontent.com/dmlc/dmlc.github.io/master/img/lo
 
 html_css_files = ["css/custom.css"]
 
-html_sidebars = {
-  '**': ['logo-text.html', 'globaltoc.html', 'searchbox.html']
-}
+html_sidebars = {"**": ["logo-text.html", "globaltoc.html", "searchbox.html"]}
 
 # Add any paths that contain custom static files (such as style sheets) here,
 # relative to this directory. They are copied after the builtin static files,
 # so a file named "default.css" will overwrite the builtin "default.css".
-html_static_path = ['_static']
+html_static_path = ["_static"]
 
 # Output file base name for HTML help builder.
-htmlhelp_basename = project + 'doc'
+htmlhelp_basename = project + "doc"
 
 # -- Options for LaTeX output ---------------------------------------------
-latex_elements = {
-}
+latex_elements = {}
 
 # Grouping the document tree into LaTeX files. List of tuples
 # (source start file, target name, title,
 #  author, documentclass [howto, manual, or own class]).
 latex_documents = [
-  (master_doc, '%s.tex' % project, project, author, 'manual'),
+    (master_doc, "%s.tex" % project, project, author, "manual"),
 ]
 
 intersphinx_mapping = {
@@ -220,30 +278,5 @@ intersphinx_mapping = {
 }
 
 
-# hook for doxygen
-def run_doxygen():
-    """Run the doxygen make command in the designated folder."""
-    curdir = os.path.normpath(os.path.abspath(os.path.curdir))
-    try:
-        os.chdir(PROJECT_ROOT)
-        if not os.path.exists(DOX_DIR):
-            os.mkdir(DOX_DIR)
-        os.chdir(os.path.join(PROJECT_ROOT, DOX_DIR))
-        subprocess.check_call(["cmake", "..", "-DBUILD_C_DOC=ON", "-GNinja"])
-        subprocess.check_call(["ninja", "doc_doxygen"])
-    except OSError as e:
-        sys.stderr.write("doxygen execution failed: %s" % e)
-    finally:
-        os.chdir(curdir)
-
-
-def generate_doxygen_xml(app):
-    """Run the doxygen make commands if we're on the ReadTheDocs server"""
-    read_the_docs_build = os.environ.get('READTHEDOCS', None) == 'True'
-    if read_the_docs_build:
-        run_doxygen()
-
-
 def setup(app):
-    app.add_css_file('custom.css')
-    app.connect("builder-inited", generate_doxygen_xml)
+    app.add_css_file("custom.css")

doc/python/python_api.rst

@@ -173,3 +173,13 @@ PySpark API
     :members:
     :inherited-members:
     :show-inheritance:
+
+.. autoclass:: xgboost.spark.SparkXGBRanker
+    :members:
+    :inherited-members:
+    :show-inheritance:
+
+.. autoclass:: xgboost.spark.SparkXGBRankerModel
+    :members:
+    :inherited-members:
+    :show-inheritance:

doc/tutorials/c_api_tutorial.rst

@@ -45,7 +45,7 @@ Use ``find_package()`` and ``target_link_libraries()`` in your application's CMa
 
 .. code-block:: cmake
 
-    cmake_minimum_required(VERSION 3.13)
+    cmake_minimum_required(VERSION 3.18)
     project(your_project_name LANGUAGES C CXX VERSION your_project_version)
     find_package(xgboost REQUIRED)
     add_executable(your_project_name /path/to/project_file.c)

doc/tutorials/spark_estimator.rst

@@ -43,10 +43,10 @@ in spark estimator, and some parameters are replaced with pyspark specific param
 such as `weight_col`, `validation_indicator_col`, `use_gpu`, for details please see
 `SparkXGBRegressor` doc.
 
-The following code snippet shows how to train a spark xgboost regressor model,
-first we need to prepare a training dataset as a spark dataframe contains
-"label" column and "features" column(s), the "features" column(s) must be `pyspark.ml.linalg.Vector`
-type or spark array type or a list of feature column names.
+The following code snippet shows how to train a spark xgboost regressor model, first we
+need to prepare a training dataset as a spark dataframe contains "label" column and
+"features" column(s), the "features" column(s) must be ``pyspark.ml.linalg.Vector`` type
+or spark array type or a list of feature column names.
 
 
 .. code-block:: python
@@ -54,10 +54,10 @@ type or spark array type or a list of feature column names.
   xgb_regressor_model = xgb_regressor.fit(train_spark_dataframe)
 
 
-The following code snippet shows how to predict test data using a spark xgboost regressor model,
-first we need to prepare a test dataset as a spark dataframe contains
-"features" and "label" column, the "features" column must be `pyspark.ml.linalg.Vector`
-type or spark array type.
+The following code snippet shows how to predict test data using a spark xgboost regressor
+model, first we need to prepare a test dataset as a spark dataframe contains "features"
+and "label" column, the "features" column must be ``pyspark.ml.linalg.Vector`` type or
+spark array type.
 
 .. code-block:: python
 
@@ -107,8 +107,8 @@ virtualenv and pip:
   python -m venv xgboost_env
   source xgboost_env/bin/activate
   pip install pyarrow pandas venv-pack xgboost
-  # https://rapids.ai/pip.html#install
-  pip install cudf-cu11 --extra-index-url=https://pypi.ngc.nvidia.com
+  # https://docs.rapids.ai/install#pip-install
+  pip install cudf-cu11 --extra-index-url=https://pypi.nvidia.com
   venv-pack -o xgboost_env.tar.gz
 
 With Conda:
@@ -240,7 +240,7 @@ additional spark configurations and dependencies:
     --master spark://<master-ip>:7077 \
     --conf spark.executor.resource.gpu.amount=1 \
     --conf spark.task.resource.gpu.amount=1 \
-    --packages com.nvidia:rapids-4-spark_2.12:22.08.0 \
+    --packages com.nvidia:rapids-4-spark_2.12:23.04.0 \
     --conf spark.plugins=com.nvidia.spark.SQLPlugin \
     --conf spark.sql.execution.arrow.maxRecordsPerBatch=1000000 \
     --archives xgboost_env.tar.gz#environment \

include/xgboost/tree_model.h

@@ -508,7 +508,7 @@ class RegTree : public Model {
      * \brief drop the trace after fill, must be called after fill.
      * \param inst The sparse instance to drop.
      */
-    void Drop(const SparsePage::Inst& inst);
+    void Drop();
     /*!
      * \brief returns the size of the feature vector
      * \return the size of the feature vector
@@ -709,13 +709,10 @@ inline void RegTree::FVec::Fill(const SparsePage::Inst& inst) {
   has_missing_ = data_.size() != feature_count;
 }
 
-inline void RegTree::FVec::Drop(const SparsePage::Inst& inst) {
-  for (auto const& entry : inst) {
-    if (entry.index >= data_.size()) {
-      continue;
-    }
-    data_[entry.index].flag = -1;
-  }
+inline void RegTree::FVec::Drop() {
+  Entry e{};
+  e.flag = -1;
+  std::fill_n(data_.data(), data_.size(), e);
   has_missing_ = true;
 }
 

include/xgboost/version_config.h

@@ -6,6 +6,6 @@
 
 #define XGBOOST_VER_MAJOR 1
 #define XGBOOST_VER_MINOR 7
-#define XGBOOST_VER_PATCH 3
+#define XGBOOST_VER_PATCH 6
 
 #endif  // XGBOOST_VERSION_CONFIG_H_

jvm-packages/pom.xml

@@ -6,7 +6,7 @@
 
     <groupId>ml.dmlc</groupId>
     <artifactId>xgboost-jvm_2.12</artifactId>
-    <version>1.7.3</version>
+    <version>1.7.6</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>1.7.3</version>
+        <version>1.7.6</version>
     </parent>
     <artifactId>xgboost4j-example_2.12</artifactId>
-    <version>1.7.3</version>
+    <version>1.7.6</version>
     <packaging>jar</packaging>
     <build>
         <plugins>
@@ -26,7 +26,7 @@
         <dependency>
             <groupId>ml.dmlc</groupId>
             <artifactId>xgboost4j-spark_${scala.binary.version}</artifactId>
-            <version>1.7.3</version>
+            <version>1.7.6</version>
         </dependency>
         <dependency>
             <groupId>org.apache.spark</groupId>
@@ -37,7 +37,7 @@
         <dependency>
             <groupId>ml.dmlc</groupId>
             <artifactId>xgboost4j-flink_${scala.binary.version}</artifactId>
-            <version>1.7.3</version>
+            <version>1.7.6</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>1.7.3</version>
+        <version>1.7.6</version>
     </parent>
     <artifactId>xgboost4j-flink_2.12</artifactId>
-    <version>1.7.3</version>
+    <version>1.7.6</version>
     <build>
         <plugins>
             <plugin>
@@ -26,7 +26,7 @@
         <dependency>
             <groupId>ml.dmlc</groupId>
             <artifactId>xgboost4j_${scala.binary.version}</artifactId>
-            <version>1.7.3</version>
+            <version>1.7.6</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>1.7.3</version>
+        <version>1.7.6</version>
     </parent>
     <artifactId>xgboost4j-gpu_2.12</artifactId>
-    <version>1.7.3</version>
+    <version>1.7.6</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>1.7.3</version>
+        <version>1.7.6</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>1.7.3</version>
+            <version>1.7.6</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>1.7.3</version>
+        <version>1.7.6</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>1.7.3</version>
+            <version>1.7.6</version>
         </dependency>
         <dependency>
             <groupId>org.apache.spark</groupId>

jvm-packages/xgboost4j/pom.xml

@@ -6,10 +6,10 @@
     <parent>
         <groupId>ml.dmlc</groupId>
         <artifactId>xgboost-jvm_2.12</artifactId>
-        <version>1.7.3</version>
+        <version>1.7.6</version>
     </parent>
     <artifactId>xgboost4j_2.12</artifactId>
-    <version>1.7.3</version>
+    <version>1.7.6</version>
     <packaging>jar</packaging>
 
     <dependencies>

plugin/CMakeLists.txt

@@ -15,7 +15,7 @@ if (PLUGIN_UPDATER_ONEAPI)
   target_link_libraries(oneapi_plugin PUBLIC -fsycl)
   set_target_properties(oneapi_plugin PROPERTIES
     COMPILE_FLAGS -fsycl
-    CXX_STANDARD 14
+    CXX_STANDARD 17
     CXX_STANDARD_REQUIRED ON
     POSITION_INDEPENDENT_CODE ON)
   if (USE_OPENMP)

python-package/xgboost/VERSION

@@ -1 +1 @@
-1.7.3
+1.7.6

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/spark/__init__.py

@@ -1,6 +1,5 @@
 # type: ignore
-"""PySpark XGBoost integration interface
-"""
+"""PySpark XGBoost integration interface"""
 
 try:
     import pyspark
@@ -11,6 +10,7 @@ from .estimator import (
     SparkXGBClassifier,
     SparkXGBClassifierModel,
     SparkXGBRanker,
+    SparkXGBRankerModel,
     SparkXGBRegressor,
     SparkXGBRegressorModel,
 )
@@ -21,4 +21,5 @@ __all__ = [
     "SparkXGBRegressor",
     "SparkXGBRegressorModel",
     "SparkXGBRanker",
+    "SparkXGBRankerModel",
 ]

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)
 
@@ -854,7 +866,11 @@ class _SparkXGBEstimator(Estimator, _SparkXGBParams, MLReadable, MLWritable):
         result_xgb_model = self._convert_to_sklearn_model(
             bytearray(booster, "utf-8"), config
         )
-        return self._copyValues(self._create_pyspark_model(result_xgb_model))
+        spark_model = self._create_pyspark_model(result_xgb_model)
+        # According to pyspark ML convention, the model uid should be the same
+        # with estimator uid.
+        spark_model._resetUid(self.uid)
+        return self._copyValues(spark_model)
 
     def write(self):
         """

rabit/CMakeLists.txt

@@ -1,4 +1,4 @@
-cmake_minimum_required(VERSION 3.3)
+cmake_minimum_required(VERSION 3.18)
 
 find_package(Threads REQUIRED)
 

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/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);
@@ -148,10 +149,28 @@ common::ColumnMatrix const &GHistIndexMatrix::Transpose() const {
   return *columns_;
 }
 
+bst_bin_t GHistIndexMatrix::GetGindex(size_t ridx, size_t fidx) const {
+  auto begin = RowIdx(ridx);
+  if (IsDense()) {
+    return static_cast<bst_bin_t>(index[begin + fidx]);
+  }
+  auto end = RowIdx(ridx + 1);
+  auto const& cut_ptrs = cut.Ptrs();
+  auto f_begin = cut_ptrs[fidx];
+  auto f_end = cut_ptrs[fidx + 1];
+  return BinarySearchBin(begin, end, index, f_begin, f_end);
+}
+
 float GHistIndexMatrix::GetFvalue(size_t ridx, size_t fidx, bool is_cat) const {
   auto const &values = cut.Values();
   auto const &mins = cut.MinValues();
   auto const &ptrs = cut.Ptrs();
+  return this->GetFvalue(ptrs, values, mins, ridx, fidx, is_cat);
+}
+
+float GHistIndexMatrix::GetFvalue(std::vector<std::uint32_t> const &ptrs,
+                                  std::vector<float> const &values, std::vector<float> const &mins,
+                                  bst_row_t ridx, bst_feature_t fidx, bool is_cat) const {
   if (is_cat) {
     auto f_begin = ptrs[fidx];
     auto f_end = ptrs[fidx + 1];
@@ -171,24 +190,27 @@ float GHistIndexMatrix::GetFvalue(size_t ridx, size_t fidx, bool is_cat) const {
     }
     return common::HistogramCuts::NumericBinValue(ptrs, values, mins, fidx, bin_idx);
   };
-
-  if (columns_->GetColumnType(fidx) == common::kDenseColumn) {
-    if (columns_->AnyMissing()) {
+  switch (columns_->GetColumnType(fidx)) {
+    case common::kDenseColumn: {
+      if (columns_->AnyMissing()) {
+        return common::DispatchBinType(columns_->GetTypeSize(), [&](auto dtype) {
+          auto column = columns_->DenseColumn<decltype(dtype), true>(fidx);
+          return get_bin_val(column);
+        });
+      } else {
+        return common::DispatchBinType(columns_->GetTypeSize(), [&](auto dtype) {
+          auto column = columns_->DenseColumn<decltype(dtype), false>(fidx);
+          auto bin_idx = column[ridx];
+          return common::HistogramCuts::NumericBinValue(ptrs, values, mins, fidx, bin_idx);
+        });
+      }
+    }
+    case common::kSparseColumn: {
       return common::DispatchBinType(columns_->GetTypeSize(), [&](auto dtype) {
-        auto column = columns_->DenseColumn<decltype(dtype), true>(fidx);
-        return get_bin_val(column);
-      });
-    } else {
-      return common::DispatchBinType(columns_->GetTypeSize(), [&](auto dtype) {
-        auto column = columns_->DenseColumn<decltype(dtype), false>(fidx);
+        auto column = columns_->SparseColumn<decltype(dtype)>(fidx, 0);
         return get_bin_val(column);
       });
     }
-  } else {
-    return common::DispatchBinType(columns_->GetTypeSize(), [&](auto dtype) {
-      auto column = columns_->SparseColumn<decltype(dtype)>(fidx, 0);
-      return get_bin_val(column);
-    });
   }
 
   SPAN_CHECK(false);

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.
 
@@ -223,7 +227,12 @@ class GHistIndexMatrix {
 
   common::ColumnMatrix const& Transpose() const;
 
+  bst_bin_t GetGindex(size_t ridx, size_t fidx) const;
+
   float GetFvalue(size_t ridx, size_t fidx, bool is_cat) const;
+  float GetFvalue(std::vector<std::uint32_t> const& ptrs, std::vector<float> const& values,
+                  std::vector<float> const& mins, bst_row_t ridx, bst_feature_t fidx,
+                  bool is_cat) const;
 
  private:
   std::unique_ptr<common::ColumnMatrix> columns_;

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;
@@ -103,6 +110,7 @@ void IterativeDMatrix::InitFromCPU(DataIterHandle iter_handle, float missing,
       size_t n_threads = ctx_.Threads();
       size_t n_features = column_sizes.size();
       linalg::Tensor<size_t, 2> column_sizes_tloc({n_threads, n_features}, Context::kCpuId);
+      column_sizes_tloc.Data()->Fill(0);
       auto view = column_sizes_tloc.HostView();
       common::ParallelFor(value.Size(), n_threads, common::Sched::Static(256), [&](auto i) {
         auto const& line = value.GetLine(i);
@@ -172,9 +180,9 @@ 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/predictor/cpu_predictor.cc

@@ -63,7 +63,7 @@ bst_float PredValue(const SparsePage::Inst &inst,
       psum += (*trees[i])[nidx].LeafValue();
     }
   }
-  p_feats->Drop(inst);
+  p_feats->Drop();
   return psum;
 }
 
@@ -116,13 +116,11 @@ void FVecFill(const size_t block_size, const size_t batch_offset, const int num_
   }
 }
 
-template <typename DataView>
-void FVecDrop(const size_t block_size, const size_t batch_offset, DataView* batch,
-              const size_t fvec_offset, std::vector<RegTree::FVec>* p_feats) {
+void FVecDrop(std::size_t const block_size, std::size_t const fvec_offset,
+              std::vector<RegTree::FVec> *p_feats) {
   for (size_t i = 0; i < block_size; ++i) {
     RegTree::FVec &feats = (*p_feats)[fvec_offset + i];
-    const SparsePage::Inst inst = (*batch)[batch_offset + i];
-    feats.Drop(inst);
+    feats.Drop();
   }
 }
 
@@ -142,11 +140,15 @@ struct SparsePageView {
 struct GHistIndexMatrixView {
  private:
   GHistIndexMatrix const &page_;
-  uint64_t n_features_;
+  std::uint64_t const n_features_;
   common::Span<FeatureType const> ft_;
   common::Span<Entry> workspace_;
   std::vector<size_t> current_unroll_;
 
+  std::vector<std::uint32_t> const& ptrs_;
+  std::vector<float> const& mins_;
+  std::vector<float> const& values_;
+
  public:
   size_t base_rowid;
 
@@ -159,6 +161,9 @@ struct GHistIndexMatrixView {
         ft_{ft},
         workspace_{workplace},
         current_unroll_(n_threads > 0 ? n_threads : 1, 0),
+        ptrs_{_page.cut.Ptrs()},
+        mins_{_page.cut.MinValues()},
+        values_{_page.cut.Values()},
         base_rowid{_page.base_rowid} {}
 
   SparsePage::Inst operator[](size_t r) {
@@ -167,7 +172,7 @@ struct GHistIndexMatrixView {
     size_t non_missing{beg};
 
     for (bst_feature_t c = 0; c < n_features_; ++c) {
-      float f = page_.GetFvalue(r, c, common::IsCat(ft_, c));
+      float f = page_.GetFvalue(ptrs_, values_, mins_, r, c, common::IsCat(ft_, c));
       if (!common::CheckNAN(f)) {
         workspace_[non_missing] = Entry{c, f};
         ++non_missing;
@@ -250,10 +255,9 @@ void PredictBatchByBlockOfRowsKernel(
     FVecFill(block_size, batch_offset, num_feature, &batch, fvec_offset,
              p_thread_temp);
     // process block of rows through all trees to keep cache locality
-    PredictByAllTrees(model, tree_begin, tree_end, out_preds,
-                      batch_offset + batch.base_rowid, num_group, thread_temp,
-                      fvec_offset, block_size);
-    FVecDrop(block_size, batch_offset, &batch, fvec_offset, p_thread_temp);
+    PredictByAllTrees(model, tree_begin, tree_end, out_preds, batch_offset + batch.base_rowid,
+                      num_group, thread_temp, fvec_offset, block_size);
+    FVecDrop(block_size, fvec_offset, p_thread_temp);
   });
 }
 
@@ -470,7 +474,7 @@ class CPUPredictor : public Predictor {
           bst_node_t tid = GetLeafIndex<true, true>(tree, feats, cats);
           preds[ridx * ntree_limit + j] = static_cast<bst_float>(tid);
         }
-        feats.Drop(page[i]);
+        feats.Drop();
       });
     }
   }
@@ -544,7 +548,7 @@ class CPUPredictor : public Predictor {
                   (tree_weights == nullptr ? 1 : (*tree_weights)[j]);
             }
           }
-          feats.Drop(page[i]);
+          feats.Drop();
           // add base margin to BIAS
           if (base_margin.Size() != 0) {
             CHECK_EQ(base_margin.Shape(1), ngroup);

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/hist/evaluate_splits.h

@@ -389,6 +389,7 @@ class HistEvaluator {
     tree_evaluator_.AddSplit(candidate.nid, left_child, right_child,
                              tree[candidate.nid].SplitIndex(), left_weight,
                              right_weight);
+    evaluator = tree_evaluator_.GetEvaluator();
 
     auto max_node = std::max(left_child, tree[candidate.nid].RightChild());
     max_node = std::max(candidate.nid, max_node);

src/tree/split_evaluator.h

@@ -48,6 +48,8 @@ class TreeEvaluator {
       monotone_.HostVector().resize(n_features, 0);
       has_constraint_ = false;
     } else {
+      CHECK_LE(p.monotone_constraints.size(), n_features)
+          << "The size of monotone constraint should be less or equal to the number of features.";
       monotone_.HostVector() = p.monotone_constraints;
       monotone_.HostVector().resize(n_features, 0);
       // Initialised to some small size, can grow if needed

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;
 
@@ -285,7 +286,7 @@ struct GPUHistMakerDevice {
         matrix.feature_segments,
         matrix.gidx_fvalue_map,
         matrix.min_fvalue,
-        matrix.is_dense
+        matrix.is_dense && !collective::IsDistributed()
     };
     auto split = this->evaluator_.EvaluateSingleSplit(inputs, shared_inputs);
     return split;
@@ -299,12 +300,14 @@ struct GPUHistMakerDevice {
     std::vector<bst_node_t> nidx(2 * candidates.size());
     auto h_node_inputs = pinned2.GetSpan<EvaluateSplitInputs>(2 * candidates.size());
     auto matrix = page->GetDeviceAccessor(ctx_->gpu_id);
-    EvaluateSplitSharedInputs shared_inputs{
-        GPUTrainingParam{param}, *quantiser, feature_types,     matrix.feature_segments,
-        matrix.gidx_fvalue_map,  matrix.min_fvalue,
-        matrix.is_dense
-    };
+    EvaluateSplitSharedInputs shared_inputs{GPUTrainingParam{param}, *quantiser, feature_types,
+                                            matrix.feature_segments, matrix.gidx_fvalue_map,
+                                            matrix.min_fvalue,
+                                            // is_dense represents the local data
+                                            matrix.is_dense && !collective::IsDistributed()};
     dh::TemporaryArray<GPUExpandEntry> entries(2 * candidates.size());
+    // Store the feature set ptrs so they dont go out of scope before the kernel is called
+    std::vector<std::shared_ptr<HostDeviceVector<bst_feature_t>>> feature_sets;
     for (size_t i = 0; i < candidates.size(); i++) {
       auto candidate = candidates.at(i);
       int left_nidx = tree[candidate.nid].LeftChild();
@@ -313,29 +316,34 @@ struct GPUHistMakerDevice {
       nidx[i * 2 + 1] = right_nidx;
       auto left_sampled_features = column_sampler.GetFeatureSet(tree.GetDepth(left_nidx));
       left_sampled_features->SetDevice(ctx_->gpu_id);
+      feature_sets.emplace_back(left_sampled_features);
       common::Span<bst_feature_t> left_feature_set =
           interaction_constraints.Query(left_sampled_features->DeviceSpan(), left_nidx);
       auto right_sampled_features = column_sampler.GetFeatureSet(tree.GetDepth(right_nidx));
       right_sampled_features->SetDevice(ctx_->gpu_id);
+      feature_sets.emplace_back(right_sampled_features);
       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));
@@ -423,7 +431,7 @@ 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;
     }
 
@@ -458,8 +466,6 @@ 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());
 
@@ -494,32 +500,45 @@ 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;
   }
@@ -862,6 +881,7 @@ 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};

src/tree/updater_quantile_hist.cc

@@ -78,7 +78,7 @@ CPUExpandEntry QuantileHistMaker::Builder::InitRoot(
 
   {
     GradientPairPrecise grad_stat;
-    if (p_fmat->IsDense()) {
+    if (p_fmat->IsDense() && !collective::IsDistributed()) {
       /**
        * Specialized code for dense data: For dense data (with no missing value), the sum
        * of gradient histogram is equal to snode[nid]

src/tree/updater_refresh.cc

@@ -89,7 +89,7 @@ class TreeRefresher : public TreeUpdater {
                      dmlc::BeginPtr(stemp[tid]) + offset);
             offset += tree->param.num_nodes;
           }
-          feats.Drop(inst);
+          feats.Drop();
         });
       }
       // aggregate the statistics

tests/buildkite/build-containers.sh

@@ -23,10 +23,15 @@ case "${container}" in
   gpu|rmm)
     BUILD_ARGS="$BUILD_ARGS --build-arg CUDA_VERSION_ARG=$CUDA_VERSION"
     BUILD_ARGS="$BUILD_ARGS --build-arg RAPIDS_VERSION_ARG=$RAPIDS_VERSION"
+    if [[ $container == "rmm" ]]
+    then
+      BUILD_ARGS="$BUILD_ARGS --build-arg NCCL_VERSION_ARG=$NCCL_VERSION"
+    fi
     ;;
 
   gpu_build_centos7|jvm_gpu_build)
     BUILD_ARGS="$BUILD_ARGS --build-arg CUDA_VERSION_ARG=$CUDA_VERSION"
+    BUILD_ARGS="$BUILD_ARGS --build-arg NCCL_VERSION_ARG=$NCCL_VERSION"
     ;;
 
   *)

tests/buildkite/build-cuda-with-rmm.sh

@@ -15,7 +15,8 @@ fi
 
 command_wrapper="tests/ci_build/ci_build.sh rmm docker --build-arg "`
                 `"CUDA_VERSION_ARG=$CUDA_VERSION --build-arg "`
-                `"RAPIDS_VERSION_ARG=$RAPIDS_VERSION"
+                `"RAPIDS_VERSION_ARG=$RAPIDS_VERSION --build-arg "`
+                `"NCCL_VERSION_ARG=$NCCL_VERSION"
 
 echo "--- Build libxgboost from the source"
 $command_wrapper tests/ci_build/build_via_cmake.sh --conda-env=gpu_test -DUSE_CUDA=ON \

tests/buildkite/build-cuda.sh

@@ -16,7 +16,8 @@ else
 fi
 
 command_wrapper="tests/ci_build/ci_build.sh gpu_build_centos7 docker --build-arg "`
-                `"CUDA_VERSION_ARG=$CUDA_VERSION"
+                `"CUDA_VERSION_ARG=$CUDA_VERSION --build-arg "`
+                `"NCCL_VERSION_ARG=$NCCL_VERSION"
 
 echo "--- Build libxgboost from the source"
 $command_wrapper tests/ci_build/prune_libnccl.sh

tests/buildkite/build-jvm-packages-gpu.sh

@@ -14,5 +14,7 @@ else
 fi
 
 tests/ci_build/ci_build.sh jvm_gpu_build nvidia-docker \
-  --build-arg CUDA_VERSION_ARG=${CUDA_VERSION} tests/ci_build/build_jvm_packages.sh \
+  --build-arg CUDA_VERSION_ARG=${CUDA_VERSION} \
+  --build-arg NCCL_VERSION_ARG=${NCCL_VERSION} \
+  tests/ci_build/build_jvm_packages.sh \
   ${SPARK_VERSION} -Duse.cuda=ON ${arch_flag}

tests/buildkite/build-win64-gpu.ps1

@@ -12,10 +12,10 @@ if ( $is_release_branch -eq 0 ) {
 }
 mkdir build
 cd build
-cmake .. -G"Visual Studio 15 2017 Win64" -DUSE_CUDA=ON -DCMAKE_VERBOSE_MAKEFILE=ON `
-  -DGOOGLE_TEST=ON -DUSE_DMLC_GTEST=ON -DCMAKE_UNITY_BUILD=ON ${arch_flag}
+cmake .. -G"Visual Studio 17 2022" -A x64 -DUSE_CUDA=ON -DCMAKE_VERBOSE_MAKEFILE=ON `
+  -DGOOGLE_TEST=ON -DUSE_DMLC_GTEST=ON ${arch_flag}
 $msbuild = -join @(
-  "C:\\Program Files (x86)\\Microsoft Visual Studio\\2017\\Community\\MSBuild\\15.0"
+  "C:\\Program Files\\Microsoft Visual Studio\\2022\\Community\\MSBuild\\Current"
   "\\Bin\\MSBuild.exe"
 )
 & $msbuild xgboost.sln /m /p:Configuration=Release /nodeReuse:false

tests/buildkite/conftest.sh

@@ -22,9 +22,10 @@ function set_buildkite_env_vars_in_container {
 
 set -x
 
-CUDA_VERSION=11.0.3
-RAPIDS_VERSION=22.10
-SPARK_VERSION=3.0.1
+CUDA_VERSION=11.8.0
+NCCL_VERSION=2.16.5-1
+RAPIDS_VERSION=23.02
+SPARK_VERSION=3.1.1
 JDK_VERSION=8
 
 if [[ -z ${BUILDKITE:-} ]]

tests/buildkite/deploy-jvm-packages.sh

@@ -9,5 +9,6 @@ then
   echo "--- Deploy JVM packages to xgboost-maven-repo S3 repo"
   tests/ci_build/ci_build.sh jvm_gpu_build docker \
     --build-arg CUDA_VERSION_ARG=${CUDA_VERSION} \
+    --build-arg NCCL_VERSION_ARG=${NCCL_VERSION} \
     tests/ci_build/deploy_jvm_packages.sh ${SPARK_VERSION}
 fi

tests/buildkite/infrastructure/aws-stack-creator/create_stack.py

@@ -2,12 +2,16 @@ import argparse
 import copy
 import os
 import re
+import sys
 
 import boto3
 import botocore
 from metadata import AMI_ID, COMMON_STACK_PARAMS, STACK_PARAMS
 
 current_dir = os.path.dirname(__file__)
+sys.path.append(os.path.join(current_dir, ".."))
+
+from common_blocks.utils import create_or_update_stack, wait
 
 TEMPLATE_URL = "https://s3.amazonaws.com/buildkite-aws-stack/latest/aws-stack.yml"
 
@@ -68,72 +72,7 @@ def get_full_stack_id(stack_id):
     return f"buildkite-{stack_id}-autoscaling-group"
 
 
-def stack_exists(args, *, stack_name):
-    client = boto3.client("cloudformation", region_name=args.aws_region)
-    waiter = client.get_waiter("stack_exists")
-    try:
-        waiter.wait(StackName=stack_name, WaiterConfig={"MaxAttempts": 1})
-        return True
-    except botocore.exceptions.WaiterError as e:
-        return False
-
-
-def create_or_update_stack(
-    args, *, stack_name, template_url=None, template_body=None, params=None
-):
-    kwargs = {
-        "StackName": stack_name,
-        "Capabilities": [
-            "CAPABILITY_IAM",
-            "CAPABILITY_NAMED_IAM",
-            "CAPABILITY_AUTO_EXPAND",
-        ],
-    }
-    if template_url:
-        kwargs["TemplateURL"] = template_url
-    if template_body:
-        kwargs["TemplateBody"] = template_body
-    if params:
-        kwargs["Parameters"] = params
-
-    client = boto3.client("cloudformation", region_name=args.aws_region)
-
-    if stack_exists(args, stack_name=stack_name):
-        print(f"Stack {stack_name} already exists. Updating...")
-        try:
-            response = client.update_stack(**kwargs)
-            return {"StackName": stack_name, "Action": "update"}
-        except botocore.exceptions.ClientError as e:
-            if e.response["Error"]["Code"] == "ValidationError" and re.search(
-                "No updates are to be performed", e.response["Error"]["Message"]
-            ):
-                print(f"No update was made to {stack_name}")
-                return {"StackName": stack_name, "Action": "noop"}
-            else:
-                raise e
-    else:
-        kwargs.update({"OnFailure": "ROLLBACK", "EnableTerminationProtection": False})
-        response = client.create_stack(**kwargs)
-        return {"StackName": stack_name, "Action": "create"}
-
-
-def wait(promise):
-    client = boto3.client("cloudformation", region_name=args.aws_region)
-    stack_name = promise["StackName"]
-    print(f"Waiting for {stack_name}...")
-    if promise["Action"] == "create":
-        waiter = client.get_waiter("stack_create_complete")
-        waiter.wait(StackName=stack_name)
-        print(f"Finished creating stack {stack_name}")
-    elif promise["Action"] == "update":
-        waiter = client.get_waiter("stack_update_complete")
-        waiter.wait(StackName=stack_name)
-        print(f"Finished updating stack {stack_name}")
-    elif promise["Action"] != "noop":
-        raise ValueError(f"Invalid promise {promise}")
-
-
-def create_agent_iam_policy(args):
+def create_agent_iam_policy(args, *, client):
     policy_stack_name = "buildkite-agent-iam-policy"
     print(f"Creating stack {policy_stack_name} for agent IAM policy...")
     with open(
@@ -142,9 +81,9 @@ def create_agent_iam_policy(args):
     ) as f:
         policy_template = f.read()
     promise = create_or_update_stack(
-        args, stack_name=policy_stack_name, template_body=policy_template
+        args, client=client, stack_name=policy_stack_name, template_body=policy_template
     )
-    wait(promise)
+    wait(promise, client=client)
 
     cf = boto3.resource("cloudformation", region_name=args.aws_region)
     policy = cf.StackResource(policy_stack_name, "BuildkiteAgentManagedPolicy")
@@ -152,10 +91,10 @@ def create_agent_iam_policy(args):
 
 
 def main(args):
-    agent_iam_policy = create_agent_iam_policy(args)
-
     client = boto3.client("cloudformation", region_name=args.aws_region)
 
+    agent_iam_policy = create_agent_iam_policy(args, client=client)
+
     promises = []
 
     for stack_id in AMI_ID:
@@ -167,13 +106,17 @@ def main(args):
         )
 
         promise = create_or_update_stack(
-            args, stack_name=stack_id_full, template_url=TEMPLATE_URL, params=params
+            args,
+            client=client,
+            stack_name=stack_id_full,
+            template_url=TEMPLATE_URL,
+            params=params,
         )
         promises.append(promise)
         print(f"CI stack {stack_id_full} is in progress in the background")
 
     for promise in promises:
-        wait(promise)
+        wait(promise, client=client)
 
 
 if __name__ == "__main__":

tests/buildkite/infrastructure/aws-stack-creator/metadata.py

@@ -1,27 +1,27 @@
 AMI_ID = {
     # Managed by XGBoost team
     "linux-amd64-gpu": {
-        "us-west-2": "ami-00ed92bd37f77bc33",
+        "us-west-2": "ami-094271bed4788ddb5",
     },
     "linux-amd64-mgpu": {
-        "us-west-2": "ami-00ed92bd37f77bc33",
+        "us-west-2": "ami-094271bed4788ddb5",
     },
     "windows-gpu": {
-        "us-west-2": "ami-0a1a2ea551a07ad5f",
+        "us-west-2": "ami-0839681594a1d7627",
     },
     "windows-cpu": {
-        "us-west-2": "ami-0a1a2ea551a07ad5f",
+        "us-west-2": "ami-0839681594a1d7627",
     },
     # Managed by BuildKite
     # from https://s3.amazonaws.com/buildkite-aws-stack/latest/aws-stack.yml
     "linux-amd64-cpu": {
-        "us-west-2": "ami-075d4c25d5f0c17c1",
+        "us-west-2": "ami-00f2127550cf03658",
     },
     "pipeline-loader": {
-        "us-west-2": "ami-075d4c25d5f0c17c1",
+        "us-west-2": "ami-00f2127550cf03658",
     },
     "linux-arm64-cpu": {
-        "us-west-2": "ami-0952c6fb6db9a9891",
+        "us-west-2": "ami-0c5789068f4a2d1b5",
     },
 }
 

tests/buildkite/infrastructure/common_blocks/utils.py

@@ -0,0 +1,97 @@
+import re
+
+import boto3
+import botocore
+
+
+def stack_exists(args, *, stack_name):
+    client = boto3.client("cloudformation", region_name=args.aws_region)
+    waiter = client.get_waiter("stack_exists")
+    try:
+        waiter.wait(StackName=stack_name, WaiterConfig={"MaxAttempts": 1})
+        return True
+    except botocore.exceptions.WaiterError as e:
+        return False
+
+
+def create_or_update_stack(
+    args, *, client, stack_name, template_url=None, template_body=None, params=None
+):
+    kwargs = {
+        "StackName": stack_name,
+        "Capabilities": [
+            "CAPABILITY_IAM",
+            "CAPABILITY_NAMED_IAM",
+            "CAPABILITY_AUTO_EXPAND",
+        ],
+    }
+    if template_url:
+        kwargs["TemplateURL"] = template_url
+    if template_body:
+        kwargs["TemplateBody"] = template_body
+    if params:
+        kwargs["Parameters"] = params
+
+    if stack_exists(args, stack_name=stack_name):
+        print(f"Stack {stack_name} already exists. Updating...")
+        try:
+            response = client.update_stack(**kwargs)
+            return {"StackName": stack_name, "Action": "update"}
+        except botocore.exceptions.ClientError as e:
+            if e.response["Error"]["Code"] == "ValidationError" and re.search(
+                "No updates are to be performed", e.response["Error"]["Message"]
+            ):
+                print(f"No update was made to {stack_name}")
+                return {"StackName": stack_name, "Action": "noop"}
+            else:
+                raise e
+    else:
+        kwargs.update({"OnFailure": "ROLLBACK", "EnableTerminationProtection": False})
+        response = client.create_stack(**kwargs)
+        return {"StackName": stack_name, "Action": "create"}
+
+
+def replace_stack(
+    args, *, client, stack_name, template_url=None, template_body=None, params=None
+):
+    """Delete an existing stack and create a new stack with identical name"""
+
+    if not stack_exists(args, stack_name=stack_name):
+        raise ValueError(f"Stack {stack_name} does not exist")
+    r = client.delete_stack(StackName=stack_name)
+    delete_waiter = client.get_waiter("stack_delete_complete")
+    delete_waiter.wait(StackName=stack_name)
+
+    kwargs = {
+        "StackName": stack_name,
+        "Capabilities": [
+            "CAPABILITY_IAM",
+            "CAPABILITY_NAMED_IAM",
+            "CAPABILITY_AUTO_EXPAND",
+        ],
+        "OnFailure": "ROLLBACK",
+        "EnableTerminationProtection": False,
+    }
+    if template_url:
+        kwargs["TemplateURL"] = template_url
+    if template_body:
+        kwargs["TemplateBody"] = template_body
+    if params:
+        kwargs["Parameters"] = params
+    response = client.create_stack(**kwargs)
+    return {"StackName": stack_name, "Action": "create"}
+
+
+def wait(promise, *, client):
+    stack_name = promise["StackName"]
+    print(f"Waiting for {stack_name}...")
+    if promise["Action"] == "create":
+        waiter = client.get_waiter("stack_create_complete")
+        waiter.wait(StackName=stack_name)
+        print(f"Finished creating stack {stack_name}")
+    elif promise["Action"] == "update":
+        waiter = client.get_waiter("stack_update_complete")
+        waiter.wait(StackName=stack_name)
+        print(f"Finished updating stack {stack_name}")
+    elif promise["Action"] != "noop":
+        raise ValueError(f"Invalid promise {promise}")

tests/buildkite/infrastructure/worker-image-pipeline/create_worker_image_pipelines.py

@@ -2,6 +2,7 @@ import argparse
 import copy
 import json
 import os
+import sys
 from urllib.request import urlopen
 
 import boto3
@@ -9,6 +10,9 @@ import cfn_flip
 from metadata import IMAGE_PARAMS
 
 current_dir = os.path.dirname(__file__)
+sys.path.append(os.path.join(current_dir, ".."))
+
+from common_blocks.utils import replace_stack, wait
 
 BUILDKITE_CF_TEMPLATE_URL = (
     "https://s3.amazonaws.com/buildkite-aws-stack/latest/aws-stack.yml"
@@ -47,6 +51,9 @@ def main(args):
 
     ami_mapping = get_ami_mapping()
 
+    client = boto3.client("cloudformation", region_name=args.aws_region)
+    promises = []
+
     for stack_id in IMAGE_PARAMS:
         stack_id_full = get_full_stack_id(stack_id)
         print(f"Creating EC2 image builder stack {stack_id_full}...")
@@ -55,28 +62,20 @@ def main(args):
             stack_id=stack_id, aws_region=args.aws_region, ami_mapping=ami_mapping
         )
 
-        client = boto3.client("cloudformation", region_name=args.aws_region)
-        response = client.create_stack(
-            StackName=stack_id_full,
-            TemplateBody=ec2_image_pipeline_template,
-            Capabilities=[
-                "CAPABILITY_IAM",
-                "CAPABILITY_NAMED_IAM",
-                "CAPABILITY_AUTO_EXPAND",
-            ],
-            OnFailure="ROLLBACK",
-            EnableTerminationProtection=False,
-            Parameters=params,
+        promise = replace_stack(
+            args,
+            client=client,
+            stack_name=stack_id_full,
+            template_body=ec2_image_pipeline_template,
+            params=params,
         )
+        promises.append(promise)
         print(
             f"EC2 image builder stack {stack_id_full} is in progress in the background"
         )
 
-    for stack_id in IMAGE_PARAMS:
-        stack_id_full = get_full_stack_id(stack_id)
-        waiter = client.get_waiter("stack_create_complete")
-        waiter.wait(StackName=stack_id_full)
-        print(f"EC2 image builder stack {stack_id_full} is now finished.")
+    for promise in promises:
+        wait(promise, client=client)
 
 
 if __name__ == "__main__":

tests/buildkite/infrastructure/worker-image-pipeline/ec2-image-builder-pipeline-template.yml

@@ -58,7 +58,7 @@ Resources:
   BootstrapComponent:
     Type: AWS::ImageBuilder::Component
     Properties:
-      Name: !Sub "${AWS::StackName}-bootstrap-component"
+      Name: !Join ["-", [!Ref AWS::StackName, "bootstrap-component", !Select [2, !Split ['/', !Ref AWS::StackId]]]]
       Platform: !Ref InstanceOperatingSystem
       Version: "1.0.0"
       Description: Execute a bootstrap script.
@@ -67,7 +67,7 @@ Resources:
   Recipe:
     Type: AWS::ImageBuilder::ImageRecipe
     Properties:
-      Name: !Sub "${AWS::StackName}-image"
+      Name: !Join ["-", [!Ref AWS::StackName, "image", !Select [2, !Split ['/', !Ref AWS::StackId]]]]
       Components:
         - ComponentArn: !Ref BootstrapComponent
       ParentImage: !Ref BaseImageId
@@ -83,7 +83,7 @@ Resources:
   Infrastructure:
     Type: AWS::ImageBuilder::InfrastructureConfiguration
     Properties:
-      Name: !Sub "${AWS::StackName}-image-pipeline-infrastructure"
+      Name: !Join ["-", [!Ref AWS::StackName, "image-pipeline-infrastructure", !Select [2, !Split ['/', !Ref AWS::StackId]]]]
       InstanceProfileName: !Ref InstanceProfile
       InstanceTypes:
         - !Ref InstanceType
@@ -93,7 +93,7 @@ Resources:
   Distribution:
     Type: AWS::ImageBuilder::DistributionConfiguration
     Properties:
-      Name: !Sub "${AWS::StackName}-image-pipeline-distribution-config"
+      Name: !Join ["-", [!Ref AWS::StackName, "image-pipeline-distribution-config", !Select [2, !Split ['/', !Ref AWS::StackId]]]]
       Distributions:
         - Region: !Ref AWS::Region
           AmiDistributionConfiguration: {}
@@ -102,7 +102,7 @@ Resources:
   Pipeline:
     Type: AWS::ImageBuilder::ImagePipeline
     Properties:
-      Name: !Sub "${AWS::StackName}-image-pipeline"
+      Name: !Join ["-", [!Ref AWS::StackName, "image-pipeline", !Select [2, !Split ['/', !Ref AWS::StackId]]]]
       DistributionConfigurationArn: !Ref Distribution
       ImageRecipeArn: !Ref Recipe
       InfrastructureConfigurationArn: !Ref Infrastructure

tests/buildkite/infrastructure/worker-image-pipeline/metadata.py

@@ -13,6 +13,6 @@ IMAGE_PARAMS = {
         "BootstrapScript": "windows-gpu-bootstrap.yml",
         "InstanceType": "g4dn.2xlarge",
         "InstanceOperatingSystem": "Windows",
-        "VolumeSize": "80",  # in GiBs
+        "VolumeSize": "120",  # in GiBs
     },
 }

tests/buildkite/infrastructure/worker-image-pipeline/windows-gpu-bootstrap.yml

@@ -15,9 +15,9 @@ phases:
               choco --version
               choco feature enable -n=allowGlobalConfirmation
 
-              # CMake 3.18
-              Write-Host '>>> Installing CMake 3.18...'
-              choco install cmake --version 3.18.0 --installargs "ADD_CMAKE_TO_PATH=System"
+              # CMake 3.25
+              Write-Host '>>> Installing CMake 3.25...'
+              choco install cmake --version 3.25.2 --installargs "ADD_CMAKE_TO_PATH=System"
               if ($LASTEXITCODE -ne 0) { throw "Last command failed" }
 
               # Notepad++
@@ -45,18 +45,18 @@ phases:
               choco install graphviz
               if ($LASTEXITCODE -ne 0) { throw "Last command failed" }
 
-              # Install Visual Studio Community 2017 (15.9)
-              Write-Host '>>> Installing Visual Studio 2017 Community (15.9)...'
-              choco install visualstudio2017community --version 15.9.23.0 `
+              # Install Visual Studio 2022 Community
+              Write-Host '>>> Installing Visual Studio 2022 Community...'
+              choco install visualstudio2022community `
                   --params "--wait --passive --norestart"
               if ($LASTEXITCODE -ne 0) { throw "Last command failed" }
-              choco install visualstudio2017-workload-nativedesktop --params `
+              choco install visualstudio2022-workload-nativedesktop --params `
                   "--wait --passive --norestart --includeOptional"
               if ($LASTEXITCODE -ne 0) { throw "Last command failed" }
 
-              # Install CUDA 11.0
-              Write-Host '>>> Installing CUDA 11.0...'
-              choco install cuda --version 11.0.3
+              # Install CUDA 11.8
+              Write-Host '>>> Installing CUDA 11.8...'
+              choco install cuda --version=11.8.0.52206
               if ($LASTEXITCODE -ne 0) { throw "Last command failed" }
 
               # Install Python packages

tests/buildkite/pipeline.yml

@@ -22,11 +22,11 @@ steps:
       queue: linux-amd64-cpu
   - wait
   #### -------- BUILD --------
-  - label: ":console: Run clang-tidy"
-    command: "tests/buildkite/run-clang-tidy.sh"
-    key: run-clang-tidy
-    agents:
-      queue: linux-amd64-cpu
+  # - label: ":console: Run clang-tidy"
+  #   command: "tests/buildkite/run-clang-tidy.sh"
+  #   key: run-clang-tidy
+  #  agents:
+  #    queue: linux-amd64-cpu
   - wait
   - label: ":console: Build CPU"
     command: "tests/buildkite/build-cpu.sh"

tests/buildkite/test-cpp-gpu.sh

@@ -20,4 +20,5 @@ tests/ci_build/ci_build.sh gpu nvidia-docker \
 # 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 \
+#   --build-arg NCCL_VERSION_ARG=$NCCL_VERSION bash -c \
 #   "source activate gpu_test && build/testxgboost --use-rmm-pool"

tests/ci_build/Dockerfile.clang_tidy

@@ -15,8 +15,8 @@ RUN \
     add-apt-repository -u 'deb http://apt.llvm.org/bionic/ llvm-toolchain-bionic-11 main' && \
     apt-get update && \
     apt-get install -y llvm-11 clang-tidy-11 clang-11 && \
-    wget -nv -nc https://cmake.org/files/v3.14/cmake-3.14.0-Linux-x86_64.sh --no-check-certificate && \
-    bash cmake-3.14.0-Linux-x86_64.sh --skip-license --prefix=/usr
+    wget -nv -nc https://cmake.org/files/v3.18/cmake-3.18.0-Linux-x86_64.sh --no-check-certificate && \
+    bash cmake-3.18.0-Linux-x86_64.sh --skip-license --prefix=/usr
 
 # Set default clang-tidy version
 RUN \

tests/ci_build/Dockerfile.cpu

@@ -12,8 +12,8 @@ RUN \
     apt-get update && \
     apt-get install -y tar unzip wget git build-essential doxygen graphviz llvm libasan2 libidn11 ninja-build gcc-8 g++-8 openjdk-8-jdk-headless && \
     # CMake
-    wget -nv -nc https://cmake.org/files/v3.14/cmake-3.14.0-Linux-x86_64.sh --no-check-certificate && \
-    bash cmake-3.14.0-Linux-x86_64.sh --skip-license --prefix=/usr && \
+    wget -nv -nc https://cmake.org/files/v3.18/cmake-3.18.0-Linux-x86_64.sh --no-check-certificate && \
+    bash cmake-3.18.0-Linux-x86_64.sh --skip-license --prefix=/usr && \
     # Python
     wget -nv https://github.com/conda-forge/miniforge/releases/latest/download/Mambaforge-Linux-x86_64.sh && \
     bash Mambaforge-Linux-x86_64.sh -b -p /opt/python

tests/ci_build/Dockerfile.gpu

@@ -22,10 +22,10 @@ ENV PATH=/opt/python/bin:$PATH
 RUN \
     conda install -c conda-forge mamba && \
     mamba create -n gpu_test -c rapidsai-nightly -c rapidsai -c nvidia -c conda-forge -c defaults \
-        python=3.9 cudf=$RAPIDS_VERSION_ARG* rmm=$RAPIDS_VERSION_ARG* cudatoolkit=$CUDA_VERSION_ARG \
+        python=3.10 cudf=$RAPIDS_VERSION_ARG* rmm=$RAPIDS_VERSION_ARG* cudatoolkit=$CUDA_VERSION_ARG \
         dask dask-cuda=$RAPIDS_VERSION_ARG* dask-cudf=$RAPIDS_VERSION_ARG* cupy \
         numpy pytest pytest-timeout scipy scikit-learn pandas matplotlib wheel python-kubernetes urllib3 graphviz hypothesis \
-        pyspark cloudpickle cuda-python=11.7.0 && \
+        pyspark cloudpickle cuda-python && \
     mamba clean --all && \
     conda run --no-capture-output -n gpu_test pip install buildkite-test-collector
 

tests/ci_build/Dockerfile.gpu_build_centos7

@@ -1,6 +1,7 @@
 ARG CUDA_VERSION_ARG
 FROM nvidia/cuda:$CUDA_VERSION_ARG-devel-centos7
 ARG CUDA_VERSION_ARG
+ARG NCCL_VERSION_ARG
 
 # Install all basic requirements
 RUN \
@@ -21,7 +22,7 @@ RUN \
 # NCCL2 (License: https://docs.nvidia.com/deeplearning/sdk/nccl-sla/index.html)
 RUN \
     export CUDA_SHORT=`echo $CUDA_VERSION_ARG | grep -o -E '[0-9]+\.[0-9]'` && \
-    export NCCL_VERSION=2.13.4-1 && \
+    export NCCL_VERSION=$NCCL_VERSION_ARG && \
     wget -nv -nc https://developer.download.nvidia.com/compute/machine-learning/repos/rhel7/x86_64/nvidia-machine-learning-repo-rhel7-1.0.0-1.x86_64.rpm && \
     rpm -i nvidia-machine-learning-repo-rhel7-1.0.0-1.x86_64.rpm && \
     yum -y update && \

tests/ci_build/Dockerfile.gpu_build_r_centos7

@@ -36,8 +36,8 @@ RUN \
     bash Miniconda3.sh -b -p /opt/python && \
     /opt/python/bin/python -m pip install auditwheel awscli && \
     # CMake
-    wget -nv -nc https://cmake.org/files/v3.14/cmake-3.14.0-Linux-x86_64.sh --no-check-certificate && \
-    bash cmake-3.14.0-Linux-x86_64.sh --skip-license --prefix=/usr
+    wget -nv -nc https://cmake.org/files/v3.18/cmake-3.18.0-Linux-x86_64.sh --no-check-certificate && \
+    bash cmake-3.18.0-Linux-x86_64.sh --skip-license --prefix=/usr
 
 ENV GOSU_VERSION 1.10
 

tests/ci_build/Dockerfile.jvm

@@ -12,8 +12,8 @@ RUN \
     wget -nv -nc -O Miniconda3.sh https://repo.anaconda.com/miniconda/Miniconda3-latest-Linux-x86_64.sh && \
     bash Miniconda3.sh -b -p /opt/python && \
     # CMake
-    wget -nv -nc https://cmake.org/files/v3.14/cmake-3.14.0-Linux-x86_64.sh --no-check-certificate && \
-    bash cmake-3.14.0-Linux-x86_64.sh --skip-license --prefix=/usr && \
+    wget -nv -nc https://cmake.org/files/v3.18/cmake-3.18.0-Linux-x86_64.sh --no-check-certificate && \
+    bash cmake-3.18.0-Linux-x86_64.sh --skip-license --prefix=/usr && \
     # Maven
     wget -nv -nc https://archive.apache.org/dist/maven/maven-3/3.6.1/binaries/apache-maven-3.6.1-bin.tar.gz && \
     tar xvf apache-maven-3.6.1-bin.tar.gz -C /opt && \

tests/ci_build/Dockerfile.jvm_gpu_build

@@ -1,6 +1,7 @@
 ARG CUDA_VERSION_ARG
 FROM nvidia/cuda:$CUDA_VERSION_ARG-devel-centos7
 ARG CUDA_VERSION_ARG
+ARG NCCL_VERSION_ARG
 
 # Install all basic requirements
 RUN \
@@ -14,8 +15,8 @@ RUN \
     wget -nv -nc -O Miniconda3.sh https://repo.anaconda.com/miniconda/Miniconda3-latest-Linux-x86_64.sh && \
     bash Miniconda3.sh -b -p /opt/python && \
     # CMake
-    wget -nv -nc https://cmake.org/files/v3.14/cmake-3.14.0-Linux-x86_64.sh --no-check-certificate && \
-    bash cmake-3.14.0-Linux-x86_64.sh --skip-license --prefix=/usr && \
+    wget -nv -nc https://cmake.org/files/v3.18/cmake-3.18.0-Linux-x86_64.sh --no-check-certificate && \
+    bash cmake-3.18.0-Linux-x86_64.sh --skip-license --prefix=/usr && \
     # Maven
     wget -nv -nc https://archive.apache.org/dist/maven/maven-3/3.6.1/binaries/apache-maven-3.6.1-bin.tar.gz && \
     tar xvf apache-maven-3.6.1-bin.tar.gz -C /opt && \
@@ -24,7 +25,7 @@ RUN \
 # NCCL2 (License: https://docs.nvidia.com/deeplearning/sdk/nccl-sla/index.html)
 RUN \
     export CUDA_SHORT=`echo $CUDA_VERSION_ARG | grep -o -E '[0-9]+\.[0-9]'` && \
-    export NCCL_VERSION=2.13.4-1 && \
+    export NCCL_VERSION=$NCCL_VERSION_ARG && \
     yum-config-manager --add-repo http://developer.download.nvidia.com/compute/cuda/repos/rhel7/x86_64/cuda-rhel7.repo && \
     yum -y update && \
     yum install -y libnccl-${NCCL_VERSION}+cuda${CUDA_SHORT} libnccl-devel-${NCCL_VERSION}+cuda${CUDA_SHORT} libnccl-static-${NCCL_VERSION}+cuda${CUDA_SHORT}

tests/ci_build/Dockerfile.rmm

@@ -1,7 +1,8 @@
 ARG CUDA_VERSION_ARG
-FROM nvidia/cuda:$CUDA_VERSION_ARG-devel-ubuntu18.04
+FROM nvidia/cuda:$CUDA_VERSION_ARG-devel-ubuntu20.04
 ARG CUDA_VERSION_ARG
 ARG RAPIDS_VERSION_ARG
+ARG NCCL_VERSION_ARG
 
 # Environment
 ENV DEBIAN_FRONTEND noninteractive
@@ -19,7 +20,7 @@ RUN \
 # NCCL2 (License: https://docs.nvidia.com/deeplearning/sdk/nccl-sla/index.html)
 RUN \
     export CUDA_SHORT=`echo $CUDA_VERSION_ARG | grep -o -E '[0-9]+\.[0-9]'` && \
-    export NCCL_VERSION=2.13.4-1 && \
+    export NCCL_VERSION=$NCCL_VERSION_ARG && \
     apt-get update && \
     apt-get install -y --allow-downgrades --allow-change-held-packages libnccl2=${NCCL_VERSION}+cuda${CUDA_SHORT} libnccl-dev=${NCCL_VERSION}+cuda${CUDA_SHORT}
 
@@ -29,7 +30,7 @@ ENV PATH=/opt/python/bin:$PATH
 RUN \
     conda install -c conda-forge mamba && \
     mamba create -n gpu_test -c rapidsai-nightly -c rapidsai -c nvidia -c conda-forge -c defaults \
-        python=3.9 rmm=$RAPIDS_VERSION_ARG* cudatoolkit=$CUDA_VERSION_ARG cmake && \
+        python=3.10 rmm=$RAPIDS_VERSION_ARG* cudatoolkit=$CUDA_VERSION_ARG cmake && \
     mamba clean --all
 
 ENV GOSU_VERSION 1.10

tests/ci_build/build_r_pkg_with_cuda_win64.sh

@@ -15,7 +15,7 @@ mv xgboost/ xgboost_rpack/
 
 mkdir build
 cd build
-cmake .. -G"Visual Studio 15 2017 Win64" -DUSE_CUDA=ON -DR_LIB=ON -DLIBR_HOME="c:\\Program Files\\R\\R-3.6.3"
+cmake .. -G"Visual Studio 17 2022" -A x64 -DUSE_CUDA=ON -DR_LIB=ON -DLIBR_HOME="c:\\Program Files\\R\\R-3.6.3"
 cmake --build . --config Release --parallel
 cd ..
 

tests/ci_build/conda_env/aarch64_test.yml

@@ -31,6 +31,5 @@ dependencies:
 - pyspark
 - cloudpickle
 - pip:
-  - shap
   - awscli
   - auditwheel

tests/ci_build/conda_env/cpu_test.yml

@@ -34,9 +34,9 @@ dependencies:
 - pyarrow
 - protobuf
 - 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/ci_build/insert_vcomp140.py

@@ -1,19 +1,102 @@
-import sys
-import re
-import zipfile
+import argparse
+import base64
 import glob
+import hashlib
+import os
+import pathlib
+import re
+import shutil
+import tempfile
 
-if len(sys.argv) != 2:
-    print('Usage: {} [wheel]'.format(sys.argv[0]))
-    sys.exit(1)
+VCOMP140_PATH = "C:\\Windows\\System32\\vcomp140.dll"
 
-vcomp140_path = 'C:\\Windows\\System32\\vcomp140.dll'
 
-for wheel_path in sorted(glob.glob(sys.argv[1])):
-    m = re.search(r'xgboost-(.*)-py3', wheel_path)
-    assert m, f'wheel_path = {wheel_path}'
-    version = m.group(1)
+def get_sha256sum(path):
+    return (
+        base64.urlsafe_b64encode(hashlib.sha256(open(path, "rb").read()).digest())
+        .decode("latin1")
+        .rstrip("=")
+    )
 
-    print(f"Inserting vcomp140.dll into {wheel_path}...")
-    with zipfile.ZipFile(wheel_path, 'a') as f:
-        f.write(vcomp140_path, 'xgboost-{}.data/data/xgboost/vcomp140.dll'.format(version))
+
+def update_record(*, wheel_content_dir, xgboost_version):
+    vcomp140_size = os.path.getsize(VCOMP140_PATH)
+    vcomp140_hash = get_sha256sum(VCOMP140_PATH)
+
+    record_path = wheel_content_dir / pathlib.Path(
+        f"xgboost-{xgboost_version}.dist-info/RECORD"
+    )
+    with open(record_path, "r") as f:
+        record_content = f.read()
+    record_content += f"xgboost-{xgboost_version}.data/data/xgboost/vcomp140.dll,"
+    record_content += f"sha256={vcomp140_hash},{vcomp140_size}\n"
+    with open(record_path, "w") as f:
+        f.write(record_content)
+
+
+def main(args):
+    candidates = list(sorted(glob.glob(args.wheel_path)))
+    for wheel_path in candidates:
+        print(f"Processing wheel {wheel_path}")
+        m = re.search(r"xgboost-(.*)\+.*-py3", wheel_path)
+        if not m:
+            raise ValueError(f"Wheel {wheel_path} has unexpected name")
+        version = m.group(1)
+        print(f"  Detected version for {wheel_path}: {version}")
+        print(f"  Inserting vcomp140.dll into {wheel_path}...")
+        with tempfile.TemporaryDirectory() as tempdir:
+            wheel_content_dir = pathlib.Path(tempdir) / "wheel_content"
+            print(f"    Extract {wheel_path} into {wheel_content_dir}")
+            shutil.unpack_archive(
+                wheel_path, extract_dir=wheel_content_dir, format="zip"
+            )
+            data_dir = wheel_content_dir / pathlib.Path(
+                f"xgboost-{version}.data/data/xgboost"
+            )
+            data_dir.mkdir(parents=True, exist_ok=True)
+
+            print(f"    Copy {VCOMP140_PATH} -> {data_dir}")
+            shutil.copy(VCOMP140_PATH, data_dir)
+
+            print(f"    Update RECORD")
+            update_record(wheel_content_dir=wheel_content_dir, xgboost_version=version)
+
+            print(f"    Content of {wheel_content_dir}:")
+            for e in sorted(wheel_content_dir.rglob("*")):
+                if e.is_file():
+                    r = e.relative_to(wheel_content_dir)
+                    print(f"      {r}")
+
+            print(f"    Create new wheel...")
+            new_wheel_tmp_path = pathlib.Path(tempdir) / "new_wheel"
+            shutil.make_archive(
+                str(new_wheel_tmp_path.resolve()),
+                format="zip",
+                root_dir=wheel_content_dir,
+            )
+            new_wheel_tmp_path = new_wheel_tmp_path.resolve().with_suffix(".zip")
+            new_wheel_tmp_path = new_wheel_tmp_path.rename(
+                new_wheel_tmp_path.with_suffix(".whl")
+            )
+            print(f"    Created new wheel {new_wheel_tmp_path}")
+
+            # Rename the old wheel with suffix .bak
+            # The new wheel takes the name of the old wheel
+            wheel_path_obj = pathlib.Path(wheel_path).resolve()
+            backup_path = wheel_path_obj.with_suffix(".whl.bak")
+            print(f"    Rename {wheel_path_obj} -> {backup_path}")
+            wheel_path_obj.replace(backup_path)
+            print(f"    Rename {new_wheel_tmp_path} -> {wheel_path_obj}")
+            new_wheel_tmp_path.replace(wheel_path_obj)
+
+            shutil.rmtree(wheel_content_dir)
+
+
+if __name__ == "__main__":
+    parser = argparse.ArgumentParser()
+    parser.add_argument(
+        "wheel_path", type=str, help="Path to wheel (wildcard permitted)"
+    )
+    args = parser.parse_args()
+
+    main(args)

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_constraints.cc

@@ -6,6 +6,9 @@
 #include <string>
 
 #include "../../../src/tree/constraints.h"
+#include "../../../src/tree/hist/evaluate_splits.h"
+#include "../../../src/tree/hist/expand_entry.h"
+#include "../helpers.h"
 
 namespace xgboost {
 namespace tree {
@@ -56,5 +59,38 @@ TEST(CPUFeatureInteractionConstraint, Basic) {
   ASSERT_FALSE(constraints.Query(1, 5));
 }
 
+TEST(CPUMonoConstraint, Basic) {
+  std::size_t kRows{64}, kCols{16};
+  Context ctx;
+
+  TrainParam param;
+  std::vector<std::int32_t> mono(kCols, 1);
+  I32Array arr;
+  for (std::size_t i = 0; i < kCols; ++i) {
+    arr.GetArray().push_back(mono[i]);
+  }
+  Json jarr{std::move(arr)};
+  std::string str_mono;
+  Json::Dump(jarr, &str_mono);
+  str_mono.front() = '(';
+  str_mono.back() = ')';
+
+  param.UpdateAllowUnknown(Args{{"monotone_constraints", str_mono}});
+
+  auto Xy = RandomDataGenerator{kRows, kCols, 0.0}.GenerateDMatrix(true);
+  auto sampler = std::make_shared<common::ColumnSampler>();
+
+  HistEvaluator<CPUExpandEntry> evalutor{param, Xy->Info(), ctx.Threads(), sampler};
+  evalutor.InitRoot(GradStats{2.0, 2.0});
+
+  SplitEntry split;
+  split.Update(1.0f, 0, 3.0, false, false, GradStats{1.0, 1.0}, GradStats{1.0, 1.0});
+  CPUExpandEntry entry{0, 0, split};
+  RegTree tree;
+  tree.param.UpdateAllowUnknown(Args{{"num_feature", std::to_string(kCols)}});
+  evalutor.ApplyTreeSplit(entry, &tree);
+
+  ASSERT_TRUE(evalutor.Evaluator().has_constraint);
+}
 }  // namespace tree
 }  // namespace xgboost

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

@@ -216,6 +216,7 @@ class TestGPUPredict:
     def test_inplace_predict_cupy(self):
         self.run_inplace_predict_cupy(0)
 
+    @pytest.mark.xfail
     @pytest.mark.skipif(**tm.no_cupy())
     @pytest.mark.mgpu
     def test_inplace_predict_cupy_specified_device(self):
@@ -338,13 +339,21 @@ 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-gpu/test_gpu_updaters.py

@@ -215,3 +215,34 @@ class TestGPUUpdaters:
     @pytest.mark.parametrize("weighted", [True, False])
     def test_adaptive(self, weighted) -> None:
         self.cputest.run_adaptive("gpu_hist", weighted)
+
+    @pytest.mark.skipif(**tm.no_pandas())
+    def test_issue8824(self):
+        # column sampling by node crashes because shared pointers go out of scope
+        import pandas as pd
+
+        data = pd.DataFrame(np.random.rand(1024, 8))
+        data.columns = "x" + data.columns.astype(str)
+        features = data.columns
+        data["y"] = data.sum(axis=1) < 4
+        dtrain = xgb.DMatrix(data[features], label=data["y"])
+        model = xgb.train(
+            dtrain=dtrain,
+            params={
+                "max_depth": 5,
+                "learning_rate": 0.05,
+                "objective": "binary:logistic",
+                "tree_method": "gpu_hist",
+                "colsample_bytree": 0.5,
+                "colsample_bylevel": 0.5,
+                "colsample_bynode": 0.5,  # Causes issues
+                "reg_alpha": 0.05,
+                "reg_lambda": 0.005,
+                "seed": 66,
+                "subsample": 0.5,
+                "gamma": 0.2,
+                "predictor": "auto",
+                "eval_metric": "auc",
+            },
+            num_boost_round=150,
+        )

tests/python/test_basic_models.py

@@ -578,7 +578,7 @@ class TestModels:
         y = rng.randn(rows)
         feature_names = ["test_feature_" + str(i) for i in range(cols)]
         X_pd = pd.DataFrame(X, columns=feature_names)
-        X_pd.iloc[:, 3] = X_pd.iloc[:, 3].astype(np.int32)
+        X_pd[f"test_feature_{3}"] = X_pd.iloc[:, 3].astype(np.int32)
 
         Xy = xgb.DMatrix(X_pd, y)
         assert Xy.feature_types[3] == "int"

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

@@ -75,7 +75,10 @@ class TestPandas:
         np.testing.assert_array_equal(result, exp)
         dm = xgb.DMatrix(dummies)
         assert dm.feature_names == ['B', 'A_X', 'A_Y', 'A_Z']
-        assert dm.feature_types == ['int', 'int', 'int', 'int']
+        if int(pd.__version__[0]) >= 2:
+            assert dm.feature_types == ['int', 'i', 'i', 'i']
+        else:
+            assert dm.feature_types == ['int', 'int', 'int', 'int']
         assert dm.num_row() == 3
         assert dm.num_col() == 4