Merge remote-tracking branch 'origin/master' into skbuild

Author: hcho3

doc/conf.py

@@ -378,6 +378,9 @@ def is_readthedocs_build():
 
 # -- Options for HTML output ----------------------------------------------
 
+# See RTD document for the effect of using canonical URL on SEO.
+html_baseurl: str = os.environ.get("READTHEDOCS_CANONICAL_URL", "")
+
 # The theme to use for HTML and HTML Help pages.  See the documentation for
 # a list of builtin themes.
 html_theme = "sphinx_rtd_theme"

doc/parameter.rst

@@ -535,7 +535,10 @@ Parameter for using Expectile Loss (``reg:expectileerror``)
 
     .. versionadded:: 3.3.0
 
-    .. note:: Multiple alphas must be sorted in ascending order. Unlike the quantile objective, expectile does not suffer from curve crossing.
+    .. note::
+
+        Multiple alphas must be sorted in ascending order. Unlike the quantile objective, expectile does not suffer from curve crossing.
+        When predicting with ``output_margin=True`` and multiple alphas, the first margin corresponds to the smallest alpha; subsequent margins are reparameterized gaps between consecutive expectile predictions, use normal prediction to obtain the actual expectile values.
 
 Parameter for using AFT Survival Loss (``survival:aft``) and Negative Log Likelihood of AFT metric (``aft-nloglik``)
 ====================================================================================================================

jvm-packages/create_jni.py

@@ -1,20 +1,20 @@
 #!/usr/bin/env python
+"""Build the native XGBoost4J JNI library."""
+
 import argparse
-import errno
-import glob
 import os
 import platform
+import shlex
 import shutil
 import subprocess
 import sys
-from contextlib import contextmanager
-
-# Monkey-patch the API inconsistency between Python2.X and 3.X.
-if sys.platform.startswith("linux"):
-    sys.platform = "linux"
+from pathlib import Path
+from typing import Sequence
 
+ROOT = Path(__file__).resolve().parents[1]
+JVM_PACKAGES = Path(__file__).resolve().parent
 
-CONFIG = {
+DEFAULT_CONFIG = {
     "USE_OPENMP": "ON",
     "USE_CUDA": "OFF",
     "USE_NCCL": "OFF",
@@ -24,121 +24,121 @@
 }
 
 
-@contextmanager
-def cd(path):
-    path = normpath(path)
-    cwd = os.getcwd()
-    os.chdir(path)
-    print("cd " + path, flush=True)
-    try:
-        yield path
-    finally:
-        os.chdir(cwd)
+def run(command: Sequence[str], *, cwd: Path | None = None) -> None:
+    """Run a shell command."""
+    print(shlex.join(command), flush=True)
+    subprocess.run(command, cwd=cwd, check=True, env=os.environ)
 
 
-def maybe_makedirs(path):
-    path = normpath(path)
-    print("mkdir -p " + path, flush=True)
-    try:
-        os.makedirs(path)
-    except OSError as e:
-        if e.errno != errno.EEXIST:
-            raise
+def mkdir(path: Path) -> None:
+    """Create a directory if it does not already exist."""
+    print(f"mkdir -p {path}", flush=True)
+    path.mkdir(parents=True, exist_ok=True)
 
 
-def run(command, **kwargs):
-    print(command, flush=True)
-    subprocess.run(command, shell=True, check=True, env=os.environ, **kwargs)
+def copy_file(source: Path, target: Path) -> None:
+    """Copy a file to a target path or directory."""
+    print(f"cp {source} {target}", flush=True)
+    shutil.copy(source, target)
 
 
-def cp(source, target):
-    source = normpath(source)
-    target = normpath(target)
-    print("cp {0} {1}".format(source, target), flush=True)
-    shutil.copy(source, target)
+def copy_glob(pattern: str, target: Path) -> None:
+    """Copy files matching a glob pattern to a target directory."""
+    for source in ROOT.glob(pattern):
+        copy_file(source, target)
 
 
-def normpath(path):
-    """Normalize UNIX path to a native path."""
-    normalized = os.path.join(*path.split("/"))
-    if os.path.isabs(path):
-        return os.path.abspath("/") + normalized
-    else:
-        return normalized
+def cmake_config(options: argparse.Namespace) -> dict[str, str]:
+    """Create CMake configuration from CLI options."""
+    config = DEFAULT_CONFIG.copy()
+    config["USE_OPENMP"] = options.use_openmp
+    config["USE_NVTX"] = options.use_nvtx
+    config["PLUGIN_RMM"] = options.plugin_rmm
 
+    if options.log_capi_invocation == "ON":
+        config["LOG_CAPI_INVOCATION"] = "ON"
+    if options.use_debug == "ON":
+        config["CMAKE_BUILD_TYPE"] = "Debug"
+    if options.use_cuda == "ON":
+        config["USE_CUDA"] = "ON"
+        config["USE_NCCL"] = "ON"
+        config["USE_DLOPEN_NCCL"] = "OFF"
 
-def native_build(cli_args: argparse.Namespace) -> None:
-    CONFIG["USE_OPENMP"] = cli_args.use_openmp
-    if sys.platform == "darwin":
-        os.environ["JAVA_HOME"] = (
-            subprocess.check_output("/usr/libexec/java_home").strip().decode()
-        )
-    if cli_args.use_debug == "ON":
-        CONFIG["CMAKE_BUILD_TYPE"] = "Debug"
-    CONFIG["USE_NVTX"] = cli_args.use_nvtx
-    CONFIG["PLUGIN_RMM"] = cli_args.plugin_rmm
+    return config
 
-    print("building Java wrapper", flush=True)
-    with cd(".."):
-        build_dir = "build-gpu" if cli_args.use_cuda == "ON" else "build"
-        maybe_makedirs(build_dir)
-
-        if sys.platform == "linux":
-            maybe_parallel_build = " -- -j $(nproc)"
-        elif sys.platform == "win32":
-            maybe_parallel_build = ' -- /m /nodeReuse:false "/consoleloggerparameters:ShowCommandLine;Verbosity=minimal"'
-        else:
-            maybe_parallel_build = ""
-
-        if cli_args.log_capi_invocation == "ON":
-            CONFIG["LOG_CAPI_INVOCATION"] = "ON"
-
-        if cli_args.use_cuda == "ON":
-            CONFIG["USE_CUDA"] = "ON"
-            CONFIG["USE_NCCL"] = "ON"
-            CONFIG["USE_DLOPEN_NCCL"] = "OFF"
-
-        args = ["-D{0}:BOOL={1}".format(k, v) for k, v in CONFIG.items()]
-        if sys.platform != "win32":
+
+def cmake_args(config: dict[str, str]) -> list[str]:
+    """Create CMake command line arguments."""
+    args = [
+        f"-D{k}:BOOL={v}" if v in ("ON", "OFF") else f"-D{k}:STRING={v}"
+        for k, v in config.items()
+    ]
+
+    if sys.platform != "win32" and shutil.which("ninja"):
+        args.append("-GNinja")
+
+    # Set GPU_ARCH_FLAG to override the CUDA architectures.
+    if gpu_arch_flag := os.getenv("GPU_ARCH_FLAG"):
+        args.append(f"-DCMAKE_CUDA_ARCHITECTURES={gpu_arch_flag}")
+
+    return args
+
+
+def windows_generators() -> tuple[list[str], ...]:
+    """Return CMake generator arguments to try on Windows."""
+    return (
+        [],  # Let CMake decide.
+        ["-G", "Visual Studio 18 2026", "-A", "x64"],
+        ["-G", "Visual Studio 17 2022", "-A", "x64"],
+        ["-G", "Visual Studio 16 2019", "-A", "x64"],
+        ["-G", "Visual Studio 15 2017", "-A", "x64"],
+    )
+
+
+def configure(config_args: list[str], build_dir: Path) -> None:
+    """Configure the CMake build."""
+    if sys.platform == "win32":
+        for generator in windows_generators():
             try:
-                subprocess.check_call(["ninja", "--version"])
-                args.append("-GNinja")
-            except FileNotFoundError:
-                pass
-
-        # if enviorment set GPU_ARCH_FLAG
-        gpu_arch_flag = os.getenv("GPU_ARCH_FLAG", None)
-        if gpu_arch_flag is not None:
-            args.append("-DCMAKE_CUDA_ARCHITECTURES=%s" % gpu_arch_flag)
-
-        with cd(build_dir):
-            lib_dir = os.path.join(os.pardir, "lib")
-            if os.path.exists(lib_dir):
-                shutil.rmtree(lib_dir)
-
-            # Same trick as Python build, just test all possible generators.
-            if sys.platform == "win32":
-                supported_generators = (
-                    "",  # empty, decided by cmake
-                    '-G"Visual Studio 17 2022" -A x64',
-                    '-G"Visual Studio 16 2019" -A x64',
-                    '-G"Visual Studio 15 2017" -A x64',
+                run(["cmake", str(ROOT), *config_args, *generator], cwd=build_dir)
+                return
+            except subprocess.CalledProcessError as err:
+                print(
+                    f"Failed to build with generator: {shlex.join(generator)}",
+                    err,
+                    flush=True,
                 )
-                for generator in supported_generators:
-                    try:
-                        run("cmake .. " + " ".join(args + [generator]))
-                        break
-                    except subprocess.CalledProcessError as e:
-                        print(f"Failed to build with generator: {generator}", e, flush=True)
-                        with cd(os.path.pardir):
-                            shutil.rmtree(build_dir)
-                            maybe_makedirs(build_dir)
-            else:
-                run("cmake .. " + " ".join(args))
-            run("cmake --build . --config Release" + maybe_parallel_build)
+                shutil.rmtree(build_dir)
+                mkdir(build_dir)
+        raise RuntimeError("None of the supported CMake generators worked.")
+
+    run(["cmake", str(ROOT), *config_args], cwd=build_dir)
+
+
+def build(config: dict[str, str], build_dir: Path) -> None:
+    """Build the native library."""
+    if (lib_dir := ROOT / "lib").exists():
+        shutil.rmtree(lib_dir)
+
+    configure(cmake_args(config), build_dir)
+
+    build_args = ["cmake", "--build", ".", "--config", "Release"]
+    if sys.platform == "linux":
+        build_args.extend(["--", "-j", str(os.cpu_count() or 1)])
+    elif sys.platform == "win32":
+        build_args.extend(
+            [
+                "--",
+                "/m",
+                "/nodeReuse:false",
+                "/consoleloggerparameters:ShowCommandLine;Verbosity=minimal",
+            ]
+        )
+    run(build_args, cwd=build_dir)
 
 
-    print("copying native library", flush=True)
+def copy_native_library() -> None:
+    """Copy the native library into the JVM package resources."""
     library_name, os_folder = {
         "Windows": ("xgboost4j.dll", "windows"),
         "Darwin": ("libxgboost4j.dylib", "macos"),
@@ -154,35 +154,59 @@ def native_build(cli_args: argparse.Namespace) -> None:
         "arm64": "aarch64",  # on macOS & Windows ARM 64-bit
         "aarch64": "aarch64",
     }[platform.machine().lower()]
-    output_folder = "xgboost4j/src/main/resources/lib/{}/{}".format(
-        os_folder, arch_folder
+
+    output_folder = (
+        JVM_PACKAGES / "xgboost4j/src/main/resources/lib" / os_folder / arch_folder
     )
-    maybe_makedirs(output_folder)
-    cp("../lib/" + library_name, output_folder)
+    mkdir(output_folder)
+    copy_file(ROOT / "lib" / library_name, output_folder)
+
+
+def copy_test_resources(*, use_cuda: bool) -> None:
+    """Copy training data used by JVM package tests."""
+    xgboost4j_resources = JVM_PACKAGES / "xgboost4j/src/test/resources"
+    mkdir(xgboost4j_resources)
+    copy_glob("demo/data/agaricus.*", xgboost4j_resources)
+
+    xgboost4j_spark_resources = JVM_PACKAGES / "xgboost4j-spark/src/test/resources"
+    mkdir(xgboost4j_spark_resources)
+
+    regression_dir = ROOT / "demo/data/regression"
+    run([sys.executable, "mapfeat.py"], cwd=regression_dir)
+    run([sys.executable, "mknfold.py", "machine.txt", "1"], cwd=regression_dir)
+
+    copy_glob("demo/data/regression/machine.txt.t*", xgboost4j_spark_resources)
+    copy_glob("demo/data/agaricus.*", xgboost4j_spark_resources)
+
+    if use_cuda:
+        xgboost4j_spark_gpu_resources = (
+            JVM_PACKAGES / "xgboost4j-spark-gpu/src/test/resources"
+        )
+        mkdir(xgboost4j_spark_gpu_resources)
+        copy_glob("demo/data/veterans_lung_cancer.csv", xgboost4j_spark_gpu_resources)
+        copy_file(
+            xgboost4j_spark_resources / "rank.train.csv",
+            xgboost4j_spark_gpu_resources,
+        )
 
-    print("copying train/test files", flush=True)
 
-    # for xgboost4j
-    maybe_makedirs("xgboost4j/src/test/resources")
-    for file in glob.glob("../demo/data/agaricus.*"):
-        cp(file, "xgboost4j/src/test/resources")
-
-    # for xgboost4j-spark
-    maybe_makedirs("xgboost4j-spark/src/test/resources")
-    with cd("../demo/data/regression"):
-        run(f'"{sys.executable}" mapfeat.py')
-        run(f'"{sys.executable}" mknfold.py machine.txt 1')
-    for file in glob.glob("../demo/data/regression/machine.txt.t*"):
-        cp(file, "xgboost4j-spark/src/test/resources")
-    for file in glob.glob("../demo/data/agaricus.*"):
-        cp(file, "xgboost4j-spark/src/test/resources")
-
-    # for xgboost4j-spark-gpu
-    if cli_args.use_cuda == "ON":
-        maybe_makedirs("xgboost4j-spark-gpu/src/test/resources")
-        for file in glob.glob("../demo/data/veterans_lung_cancer.csv"):
-            cp(file, "xgboost4j-spark-gpu/src/test/resources")
-        cp("xgboost4j-spark/src/test/resources/rank.train.csv", "xgboost4j-spark-gpu/src/test/resources")
+def native_build(options: argparse.Namespace) -> None:
+    """Build and copy the native JNI library and its test resources."""
+    if sys.platform == "darwin":
+        os.environ["JAVA_HOME"] = (
+            subprocess.check_output(["/usr/libexec/java_home"]).strip().decode()
+        )
+
+    print("building Java wrapper", flush=True)
+    build_dir = ROOT / ("build-gpu" if options.use_cuda == "ON" else "build")
+    mkdir(build_dir)
+    build(cmake_config(options), build_dir)
+
+    print("copying native library", flush=True)
+    copy_native_library()
+
+    print("copying train/test files", flush=True)
+    copy_test_resources(use_cuda=options.use_cuda == "ON")
 
 
 if __name__ == "__main__":
@@ -195,5 +219,5 @@ def native_build(cli_args: argparse.Namespace) -> None:
     parser.add_argument("--use-debug", type=str, choices=["ON", "OFF"], default="OFF")
     parser.add_argument("--use-nvtx", type=str, choices=["ON", "OFF"], default="OFF")
     parser.add_argument("--plugin-rmm", type=str, choices=["ON", "OFF"], default="OFF")
-    cli_args = parser.parse_args()
-    native_build(cli_args)
+    parsed_args = parser.parse_args()
+    native_build(parsed_args)

python-package/xgboost/dask/utils.py

@@ -17,8 +17,8 @@
 
 def get_n_threads(local_param: Dict[str, Any], worker: "distributed.Worker") -> int:
     """Get the number of threads from a worker and the user-supplied parameters."""
-    # dask worker nthreads, "state" is available in 2022.6.1
-    dwnt = worker.state.nthreads if hasattr(worker, "state") else worker.nthreads
+    # dask worker nthreads
+    dwnt = worker.state.nthreads
     n_threads = None
     for p in ["nthread", "n_jobs"]:
         if local_param.get(p, None) is not None and local_param.get(p, dwnt) != dwnt: