Document for device ordinal. (#9398)

- Rewrite GPU demos. notebook is converted to script to avoid committing additional png plots.
- Add GPU demos into the sphinx gallery.
- Add RMM demos into the sphinx gallery.
- Test for firing threads with different device ordinals.

demo/c-api/basic/c-api-demo.c

@@ -53,15 +53,7 @@ int main() {
   // configure the training
   // available parameters are described here:
   //   https://xgboost.readthedocs.io/en/latest/parameter.html
-  safe_xgboost(XGBoosterSetParam(booster, "tree_method", use_gpu ? "gpu_hist" : "hist"));
-  if (use_gpu) {
-    // set the GPU to use;
-    // this is not necessary, but provided here as an illustration
-    safe_xgboost(XGBoosterSetParam(booster, "gpu_id", "0"));
-  } else {
-    // avoid evaluating objective and metric on a GPU
-    safe_xgboost(XGBoosterSetParam(booster, "gpu_id", "-1"));
-  }
+  safe_xgboost(XGBoosterSetParam(booster, "device", use_gpu ? "cuda" : "cpu"));
 
   safe_xgboost(XGBoosterSetParam(booster, "objective", "binary:logistic"));
   safe_xgboost(XGBoosterSetParam(booster, "min_child_weight", "1"));

demo/gpu_acceleration/README.md

@@ -1,5 +0,0 @@
-# GPU Acceleration Demo
-
-`cover_type.py` shows how to train a model on the [forest cover type](https://archive.ics.uci.edu/ml/datasets/covertype) dataset using GPU acceleration. The forest cover type dataset has 581,012 rows and 54 features, making it time consuming to process. We compare the run-time and accuracy of the GPU and CPU histogram algorithms.
-
-`shap.ipynb` demonstrates using GPU acceleration to compute SHAP values for feature importance.

demo/gpu_acceleration/README.rst

@@ -0,0 +1,8 @@
+:orphan:
+
+GPU Acceleration Demo
+=====================
+
+This is a collection of demonstration scripts to showcase the basic usage of GPU. Please
+see :doc:`/gpu/index` for more info. There are other demonstrations for distributed GPU
+training using dask or spark.

demo/gpu_acceleration/cover_type.py

@@ -1,41 +1,49 @@
+"""
+Using xgboost on GPU devices
+============================
+
+Shows how to train a model on the `forest cover type
+<https://archive.ics.uci.edu/ml/datasets/covertype>`_ dataset using GPU
+acceleration. The forest cover type dataset has 581,012 rows and 54 features, making it
+time consuming to process. We compare the run-time and accuracy of the GPU and CPU
+histogram algorithms.
+
+In addition, The demo showcases using GPU with other GPU-related libraries including
+cupy and cuml. These libraries are not strictly required.
+
+"""
 import time
 
+import cupy as cp
+from cuml.model_selection import train_test_split
 from sklearn.datasets import fetch_covtype
-from sklearn.model_selection import train_test_split
 
 import xgboost as xgb
 
 # Fetch dataset using sklearn
-cov = fetch_covtype()
-X = cov.data
-y = cov.target
+X, y = fetch_covtype(return_X_y=True)
+X = cp.array(X)
+y = cp.array(y)
+y -= y.min()
 
 # Create 0.75/0.25 train/test split
-X_train, X_test, y_train, y_test = train_test_split(X, y, test_size=0.25, train_size=0.75,
-                                                    random_state=42)
+X_train, X_test, y_train, y_test = train_test_split(
+    X, y, test_size=0.25, train_size=0.75, random_state=42
+)
 
 # Specify sufficient boosting iterations to reach a minimum
 num_round = 3000
 
 # Leave most parameters as default
-param = {'objective': 'multi:softmax', # Specify multiclass classification
-         'num_class': 8, # Number of possible output classes
-         'tree_method': 'gpu_hist' # Use GPU accelerated algorithm
-         }
-
-# Convert input data from numpy to XGBoost format
-dtrain = xgb.DMatrix(X_train, label=y_train)
-dtest = xgb.DMatrix(X_test, label=y_test)
-
-gpu_res = {} # Store accuracy result
-tmp = time.time()
+clf = xgb.XGBClassifier(device="cuda", n_estimators=num_round)
 # Train model
-xgb.train(param, dtrain, num_round, evals=[(dtest, 'test')], evals_result=gpu_res)
-print("GPU Training Time: %s seconds" % (str(time.time() - tmp)))
+start = time.time()
+clf.fit(X_train, y_train, eval_set=[(X_test, y_test)])
+gpu_res = clf.evals_result()
+print("GPU Training Time: %s seconds" % (str(time.time() - start)))
 
 # Repeat for CPU algorithm
-tmp = time.time()
-param['tree_method'] = 'hist'
-cpu_res = {}
-xgb.train(param, dtrain, num_round, evals=[(dtest, 'test')], evals_result=cpu_res)
-print("CPU Training Time: %s seconds" % (str(time.time() - tmp)))
+clf = xgb.XGBClassifier(device="cpu", n_estimators=num_round)
+start = time.time()
+cpu_res = clf.evals_result()
+print("CPU Training Time: %s seconds" % (str(time.time() - start)))

demo/gpu_acceleration/tree_shap.py

@@ -0,0 +1,55 @@
+"""
+Use GPU to speedup SHAP value computation
+=========================================
+
+Demonstrates using GPU acceleration to compute SHAP values for feature importance.
+
+"""
+import shap
+from sklearn.datasets import fetch_california_housing
+
+import xgboost as xgb
+
+# Fetch dataset using sklearn
+data = fetch_california_housing()
+print(data.DESCR)
+X = data.data
+y = data.target
+
+num_round = 500
+
+param = {
+    "eta": 0.05,
+    "max_depth": 10,
+    "tree_method": "hist",
+    "device": "cuda",
+}
+
+# GPU accelerated training
+dtrain = xgb.DMatrix(X, label=y, feature_names=data.feature_names)
+model = xgb.train(param, dtrain, num_round)
+
+# Compute shap values using GPU with xgboost
+model.set_param({"device": "cuda"})
+shap_values = model.predict(dtrain, pred_contribs=True)
+
+# Compute shap interaction values using GPU
+shap_interaction_values = model.predict(dtrain, pred_interactions=True)
+
+
+# shap will call the GPU accelerated version as long as the device parameter is set to
+# "cuda"
+explainer = shap.TreeExplainer(model)
+shap_values = explainer.shap_values(X)
+
+# visualize the first prediction's explanation
+shap.force_plot(
+    explainer.expected_value,
+    shap_values[0, :],
+    X[0, :],
+    feature_names=data.feature_names,
+    matplotlib=True,
+)
+
+# Show a summary of feature importance
+shap.summary_plot(shap_values, X, plot_type="bar", feature_names=data.feature_names)

demo/nvflare/horizontal/custom/trainer.py

@@ -70,8 +70,7 @@ class XGBoostTrainer(Executor):
             param = {'max_depth': 2, 'eta': 1, 'objective': 'binary:logistic'}
             if self._use_gpus:
                 self.log_info(fl_ctx, f'Training with GPU {rank}')
-                param['tree_method'] = 'gpu_hist'
-                param['gpu_id'] = rank
+                param['device'] = f"cuda:{rank}"
 
             # Specify validations set to watch performance
             watchlist = [(dtest, 'eval'), (dtrain, 'train')]

demo/rmm_plugin/README.md

@@ -1,47 +0,0 @@
-Using XGBoost with RAPIDS Memory Manager (RMM) plugin (EXPERIMENTAL)
-====================================================================
-[RAPIDS Memory Manager (RMM)](https://github.com/rapidsai/rmm) library provides a collection of
-efficient memory allocators for NVIDIA GPUs. It is now possible to use XGBoost with memory
-allocators provided by RMM, by enabling the RMM integration plugin.
-
-The demos in this directory highlights one RMM allocator in particular: **the pool sub-allocator**.
-This allocator addresses the slow speed of `cudaMalloc()` by allocating a large chunk of memory
-upfront. Subsequent allocations will draw from the pool of already allocated memory and thus avoid
-the overhead of calling `cudaMalloc()` directly. See
-[this GTC talk slides](https://on-demand.gputechconf.com/gtc/2015/presentation/S5530-Stephen-Jones.pdf)
-for more details.
-
-Before running the demos, ensure that XGBoost is compiled with the RMM plugin enabled. To do this,
-run CMake with option `-DPLUGIN_RMM=ON` (`-DUSE_CUDA=ON` also required):
-```
-cmake .. -DUSE_CUDA=ON -DUSE_NCCL=ON -DPLUGIN_RMM=ON
-make -j4
-```
-CMake will attempt to locate the RMM library in your build environment. You may choose to build
-RMM from the source, or install it using the Conda package manager. If CMake cannot find RMM, you
-should specify the location of RMM with the CMake prefix:
-```
-# If using Conda:
-cmake .. -DUSE_CUDA=ON -DUSE_NCCL=ON -DPLUGIN_RMM=ON -DCMAKE_PREFIX_PATH=$CONDA_PREFIX
-# If using RMM installed with a custom location
-cmake .. -DUSE_CUDA=ON -DUSE_NCCL=ON -DPLUGIN_RMM=ON -DCMAKE_PREFIX_PATH=/path/to/rmm
-```
-
-# Informing XGBoost about RMM pool
-
-When XGBoost is compiled with RMM, most of the large size allocation will go through RMM
-allocators, but some small allocations in performance critical areas are using a different
-caching allocator so that we can have better control over memory allocation behavior.
-Users can override this behavior and force the use of rmm for all allocations by setting
-the global configuration ``use_rmm``:
-
-``` python
-with xgb.config_context(use_rmm=True):
-    clf = xgb.XGBClassifier(tree_method="gpu_hist")
-```
-
-Depending on the choice of memory pool size or type of allocator, this may have negative
-performance impact.
-
-* [Using RMM with a single GPU](./rmm_singlegpu.py)
-* [Using RMM with a local Dask cluster consisting of multiple GPUs](./rmm_mgpu_with_dask.py)

demo/rmm_plugin/README.rst

@@ -0,0 +1,51 @@
+Using XGBoost with RAPIDS Memory Manager (RMM) plugin (EXPERIMENTAL)
+====================================================================
+
+`RAPIDS Memory Manager (RMM) <https://github.com/rapidsai/rmm>`__ library provides a
+collection of efficient memory allocators for NVIDIA GPUs. It is now possible to use
+XGBoost with memory allocators provided by RMM, by enabling the RMM integration plugin.
+
+The demos in this directory highlights one RMM allocator in particular: **the pool
+sub-allocator**.  This allocator addresses the slow speed of ``cudaMalloc()`` by
+allocating a large chunk of memory upfront. Subsequent allocations will draw from the pool
+of already allocated memory and thus avoid the overhead of calling ``cudaMalloc()``
+directly. See `this GTC talk slides
+<https://on-demand.gputechconf.com/gtc/2015/presentation/S5530-Stephen-Jones.pdf>`_ for
+more details.
+
+Before running the demos, ensure that XGBoost is compiled with the RMM plugin enabled. To do this,
+run CMake with option ``-DPLUGIN_RMM=ON`` (``-DUSE_CUDA=ON`` also required):
+
+.. code-block:: sh
+
+  cmake .. -DUSE_CUDA=ON -DUSE_NCCL=ON -DPLUGIN_RMM=ON
+  make -j$(nproc)
+
+CMake will attempt to locate the RMM library in your build environment. You may choose to build
+RMM from the source, or install it using the Conda package manager. If CMake cannot find RMM, you
+should specify the location of RMM with the CMake prefix:
+
+.. code-block:: sh
+
+  # If using Conda:
+  cmake .. -DUSE_CUDA=ON -DUSE_NCCL=ON -DPLUGIN_RMM=ON -DCMAKE_PREFIX_PATH=$CONDA_PREFIX
+  # If using RMM installed with a custom location
+  cmake .. -DUSE_CUDA=ON -DUSE_NCCL=ON -DPLUGIN_RMM=ON -DCMAKE_PREFIX_PATH=/path/to/rmm
+
+********************************
+Informing XGBoost about RMM pool
+********************************
+
+When XGBoost is compiled with RMM, most of the large size allocation will go through RMM
+allocators, but some small allocations in performance critical areas are using a different
+caching allocator so that we can have better control over memory allocation behavior.
+Users can override this behavior and force the use of rmm for all allocations by setting
+the global configuration ``use_rmm``:
+
+.. code-block:: python
+
+  with xgb.config_context(use_rmm=True):
+    clf = xgb.XGBClassifier(tree_method="hist", device="cuda")
+
+Depending on the choice of memory pool size or type of allocator, this may have negative
+performance impact.

demo/rmm_plugin/rmm_mgpu_with_dask.py

@@ -1,3 +1,7 @@
+"""
+Using rmm with Dask
+===================
+"""
 import dask
 from dask.distributed import Client
 from dask_cuda import LocalCUDACluster
@@ -11,25 +15,33 @@ def main(client):
     # xgb.set_config(use_rmm=True)
 
     X, y = make_classification(n_samples=10000, n_informative=5, n_classes=3)
-    # In pratice one should prefer loading the data with dask collections instead of using
-    # `from_array`.
+    # In pratice one should prefer loading the data with dask collections instead of
+    # using `from_array`.
     X = dask.array.from_array(X)
     y = dask.array.from_array(y)
     dtrain = xgb.dask.DaskDMatrix(client, X, label=y)
 
-    params = {'max_depth': 8, 'eta': 0.01, 'objective': 'multi:softprob', 'num_class': 3,
-              'tree_method': 'gpu_hist', 'eval_metric': 'merror'}
-    output = xgb.dask.train(client, params, dtrain, num_boost_round=100,
-                            evals=[(dtrain, 'train')])
-    bst = output['booster']
-    history = output['history']
-    for i, e in enumerate(history['train']['merror']):
-        print(f'[{i}] train-merror: {e}')
+    params = {
+        "max_depth": 8,
+        "eta": 0.01,
+        "objective": "multi:softprob",
+        "num_class": 3,
+        "tree_method": "hist",
+        "eval_metric": "merror",
+        "device": "cuda",
+    }
+    output = xgb.dask.train(
+        client, params, dtrain, num_boost_round=100, evals=[(dtrain, "train")]
+    )
+    bst = output["booster"]
+    history = output["history"]
+    for i, e in enumerate(history["train"]["merror"]):
+        print(f"[{i}] train-merror: {e}")
 
 
-if __name__ == '__main__':
-    # To use RMM pool allocator with a GPU Dask cluster, just add rmm_pool_size option to
-    # LocalCUDACluster constructor.
-    with LocalCUDACluster(rmm_pool_size='2GB') as cluster:
+if __name__ == "__main__":
+    # To use RMM pool allocator with a GPU Dask cluster, just add rmm_pool_size option
+    # to LocalCUDACluster constructor.
+    with LocalCUDACluster(rmm_pool_size="2GB") as cluster:
         with Client(cluster) as client:
             main(client)

demo/rmm_plugin/rmm_singlegpu.py

@@ -1,3 +1,7 @@
+"""
+Using rmm on a single node device
+=================================
+"""
 import rmm
 from sklearn.datasets import make_classification
 
@@ -16,7 +20,8 @@ params = {
     "eta": 0.01,
     "objective": "multi:softprob",
     "num_class": 3,
-    "tree_method": "gpu_hist",
+    "tree_method": "hist",
+    "device": "cuda",
 }
 # XGBoost will automatically use the RMM pool allocator
 bst = xgb.train(params, dtrain, num_boost_round=100, evals=[(dtrain, "train")])

doc/.gitignore

@@ -6,3 +6,5 @@ doxygen
 parser.py
 *.pyc
 web-data
+# generated by doxygen
+tmp

doc/conf.py

@@ -19,7 +19,6 @@ import sys
 import tarfile
 import urllib.request
 import warnings
-from subprocess import call
 from urllib.error import HTTPError
 
 from sh.contrib import git
@@ -148,12 +147,20 @@ extensions = [
 
 sphinx_gallery_conf = {
     # path to your example scripts
-    "examples_dirs": ["../demo/guide-python", "../demo/dask", "../demo/aft_survival"],
+    "examples_dirs": [
+        "../demo/guide-python",
+        "../demo/dask",
+        "../demo/aft_survival",
+        "../demo/gpu_acceleration",
+        "../demo/rmm_plugin"
+    ],
     # path to where to save gallery generated output
     "gallery_dirs": [
         "python/examples",
         "python/dask-examples",
         "python/survival-examples",
+        "python/gpu-examples",
+        "python/rmm-examples",
     ],
     "matplotlib_animations": True,
 }

doc/gpu/index.rst

@@ -23,20 +23,19 @@ The GPU algorithms currently work with CLI, Python, R, and JVM packages. See :do
   :caption: Python example
 
   params = dict()
-  params["device"] = "cuda:0"
+  params["device"] = "cuda"
   params["tree_method"] = "hist"
   Xy = xgboost.QuantileDMatrix(X, y)
   xgboost.train(params, Xy)
 
 .. code-block:: python
-  :caption: With Scikit-Learn interface
+  :caption: With the Scikit-Learn interface
 
   XGBRegressor(tree_method="hist", device="cuda")
 
-
 GPU-Accelerated SHAP values
 =============================
-XGBoost makes use of `GPUTreeShap <https://github.com/rapidsai/gputreeshap>`_ as a backend for computing shap values when the GPU predictor is selected.
+XGBoost makes use of `GPUTreeShap <https://github.com/rapidsai/gputreeshap>`_ as a backend for computing shap values when the GPU is used.
 
 .. code-block:: python
 
@@ -44,12 +43,12 @@ XGBoost makes use of `GPUTreeShap <https://github.com/rapidsai/gputreeshap>`_ as
   shap_values = booster.predict(dtrain, pred_contribs=True)
   shap_interaction_values = model.predict(dtrain, pred_interactions=True)
 
-See examples `here <https://github.com/dmlc/xgboost/tree/master/demo/gpu_acceleration>`__.
+See :ref:`sphx_glr_python_gpu-examples_tree_shap.py` for a worked example.
 
 Multi-node Multi-GPU Training
 =============================
 
-XGBoost supports fully distributed GPU training using `Dask <https://dask.org/>`_, ``Spark`` and ``PySpark``. For getting started with Dask see our tutorial :doc:`/tutorials/dask` and worked examples `here <https://github.com/dmlc/xgboost/tree/master/demo/dask>`__, also Python documentation :ref:`dask_api` for complete reference. For usage with ``Spark`` using Scala see :doc:`/jvm/xgboost4j_spark_gpu_tutorial`. Lastly for distributed GPU training with ``PySpark``, see :doc:`/tutorials/spark_estimator`.
+XGBoost supports fully distributed GPU training using `Dask <https://dask.org/>`_, ``Spark`` and ``PySpark``. For getting started with Dask see our tutorial :doc:`/tutorials/dask` and worked examples :doc:`/python/dask-examples/index`, also Python documentation :ref:`dask_api` for complete reference. For usage with ``Spark`` using Scala see :doc:`/jvm/xgboost4j_spark_gpu_tutorial`. Lastly for distributed GPU training with ``PySpark``, see :doc:`/tutorials/spark_estimator`.
 
 
 Memory usage
@@ -67,7 +66,8 @@ If you are getting out-of-memory errors on a big dataset, try the or :py:class:`
 
 CPU-GPU Interoperability
 ========================
-XGBoost models trained on GPUs can be used on CPU-only systems to generate predictions. For information about how to save and load an XGBoost model, see :doc:`/tutorials/saving_model`.
+
+The model can be used on any device regardless of the one used to train it. For instance, a model trained using GPU can still work on a CPU-only machine and vice versa. For more information about model serialization, see :doc:`/tutorials/saving_model`.
 
 
 Developer notes

doc/install.rst

@@ -189,7 +189,7 @@ This will check out the latest stable version from the Maven Central.
 
 For the latest release version number, please check `release page <https://github.com/dmlc/xgboost/releases>`_.
 
-To enable the GPU algorithm (``tree_method='gpu_hist'``), use artifacts ``xgboost4j-gpu_2.12`` and ``xgboost4j-spark-gpu_2.12`` instead (note the ``gpu`` suffix).
+To enable the GPU algorithm (``device='cuda'``), use artifacts ``xgboost4j-gpu_2.12`` and ``xgboost4j-spark-gpu_2.12`` instead (note the ``gpu`` suffix).
 
 
 .. note:: Windows not supported in the JVM package
@@ -325,4 +325,4 @@ The SNAPSHOT JARs are hosted by the XGBoost project. Every commit in the ``maste
 
 You can browse the file listing of the Maven repository at https://s3-us-west-2.amazonaws.com/xgboost-maven-repo/list.html.
 
-To enable the GPU algorithm (``tree_method='gpu_hist'``), use artifacts ``xgboost4j-gpu_2.12`` and ``xgboost4j-spark-gpu_2.12`` instead (note the ``gpu`` suffix).
+To enable the GPU algorithm (``device='cuda'``), use artifacts ``xgboost4j-gpu_2.12`` and ``xgboost4j-spark-gpu_2.12`` instead (note the ``gpu`` suffix).

doc/parameter.rst

@@ -34,27 +34,6 @@ General Parameters
 
   - Which booster to use. Can be ``gbtree``, ``gblinear`` or ``dart``; ``gbtree`` and ``dart`` use tree based models while ``gblinear`` uses linear functions.
 
-* ``verbosity`` [default=1]
-
-  - Verbosity of printing messages.  Valid values are 0 (silent), 1 (warning), 2 (info), 3
-    (debug).  Sometimes XGBoost tries to change configurations based on heuristics, which
-    is displayed as warning message.  If there's unexpected behaviour, please try to
-    increase value of verbosity.
-
-* ``validate_parameters`` [default to ``false``, except for Python, R and CLI interface]
-
-  - When set to True, XGBoost will perform validation of input parameters to check whether
-    a parameter is used or not.
-
-* ``nthread`` [default to maximum number of threads available if not set]
-
-  - Number of parallel threads used to run XGBoost.  When choosing it, please keep thread
-    contention and hyperthreading in mind.
-
-* ``disable_default_eval_metric`` [default= ``false``]
-
-  - Flag to disable default metric. Set to 1 or ``true`` to disable.
-
 * ``device`` [default= ``cpu``]
 
   .. versionadded:: 2.0.0
@@ -67,6 +46,29 @@ General Parameters
     + ``gpu``: Default GPU device selection from the list of available and supported devices. Only ``cuda`` devices are supported currently.
     + ``gpu:<ordinal>``: Default GPU device selection from the list of available and supported devices. Only ``cuda`` devices are supported currently.
 
+    For more information about GPU acceleration, see :doc:`/gpu/index`.
+
+* ``verbosity`` [default=1]
+
+  - Verbosity of printing messages.  Valid values are 0 (silent), 1 (warning), 2 (info), 3
+    (debug).  Sometimes XGBoost tries to change configurations based on heuristics, which
+    is displayed as warning message.  If there's unexpected behaviour, please try to
+    increase value of verbosity.
+
+* ``validate_parameters`` [default to ``false``, except for Python, R and CLI interface]
+
+  - When set to True, XGBoost will perform validation of input parameters to check whether
+    a parameter is used or not. A warning is emitted when there's unknown parameter.
+
+* ``nthread`` [default to maximum number of threads available if not set]
+
+  - Number of parallel threads used to run XGBoost.  When choosing it, please keep thread
+    contention and hyperthreading in mind.
+
+* ``disable_default_eval_metric`` [default= ``false``]
+
+  - Flag to disable default metric. Set to 1 or ``true`` to disable.
+
 Parameters for Tree Booster
 ===========================
 * ``eta`` [default=0.3, alias: ``learning_rate``]
@@ -160,7 +162,7 @@ Parameters for Tree Booster
     - ``grow_colmaker``: non-distributed column-based construction of trees.
     - ``grow_histmaker``: distributed tree construction with row-based data splitting based on global proposal of histogram counting.
     - ``grow_quantile_histmaker``: Grow tree using quantized histogram.
-    - ``grow_gpu_hist``: Grow tree with GPU. Same as setting ``tree_method`` to ``hist`` and use ``device=cuda``.
+    - ``grow_gpu_hist``: Grow tree with GPU. Enabled when ``tree_method`` is set to ``hist`` along with ``device=cuda``.
     - ``sync``: synchronizes trees in all distributed nodes.
     - ``refresh``: refreshes tree's statistics and/or leaf values based on the current data. Note that no random subsampling of data rows is performed.
     - ``prune``: prunes the splits where loss < min_split_loss (or gamma) and nodes that have depth greater than ``max_depth``.

doc/python/.gitignore

@@ -1,3 +1,5 @@
 examples
 dask-examples
-survival-examples
+survival-examples
+gpu-examples
+rmm-examples

doc/python/index.rst

@@ -17,3 +17,5 @@ Contents
   examples/index
   dask-examples/index
   survival-examples/index
+  gpu-examples/index
+  rmm-examples/index

doc/treemethod.rst

@@ -124,7 +124,7 @@ Following table summarizes some differences in supported features between 4 tree
 `T` means supported while `F` means unsupported.
 
 +------------------+-----------+---------------------+---------------------+------------------------+
-|                  | Exact     | Approx              | Hist                | GPU Hist               |
+|                  | Exact     | Approx              | Hist                | Hist (GPU)             |
 +==================+===========+=====================+=====================+========================+
 | grow_policy      | Depthwise | depthwise/lossguide | depthwise/lossguide | depthwise/lossguide    |
 +------------------+-----------+---------------------+---------------------+------------------------+
@@ -141,5 +141,5 @@ Following table summarizes some differences in supported features between 4 tree
 
 Features/parameters that are not mentioned here are universally supported for all 4 tree
 methods (for instance, column sampling and constraints).  The `P` in external memory means
-partially supported.  Please note that both categorical data and external memory are
+special handling.  Please note that both categorical data and external memory are
 experimental.

doc/tutorials/categorical.rst

@@ -35,8 +35,8 @@ parameter ``enable_categorical``:
 
 .. code:: python
 
-  # Supported tree methods are `gpu_hist`, `approx`, and `hist`.
-  clf = xgb.XGBClassifier(tree_method="gpu_hist", enable_categorical=True)
+  # Supported tree methods are `approx` and `hist`.
+  clf = xgb.XGBClassifier(tree_method="hist", enable_categorical=True, device="cuda")
   # X is the dataframe we created in previous snippet
   clf.fit(X, y)
   # Must use JSON/UBJSON for serialization, otherwise the information is lost.

doc/tutorials/external_memory.rst

@@ -81,7 +81,7 @@ constructor.
   it = Iterator(["file_0.svm", "file_1.svm", "file_2.svm"])
   Xy = xgboost.DMatrix(it)
 
-  # Other tree methods including ``hist`` and ``gpu_hist`` also work, but has some caveats
+  # The ``approx`` also work, but with low performance. GPU implementation is different from CPU.
   # as noted in following sections.
   booster = xgboost.train({"tree_method": "hist"}, Xy)
 
@@ -118,15 +118,15 @@ to reduce the overhead of file reading.
 GPU Version (GPU Hist tree method)
 **********************************
 
-External memory is supported by GPU algorithms (i.e. when ``tree_method`` is set to
-``gpu_hist``). However, the algorithm used for GPU is different from the one used for
+External memory is supported by GPU algorithms (i.e. when ``device`` is set to
+``cuda``). However, the algorithm used for GPU is different from the one used for
 CPU. When training on a CPU, the tree method iterates through all batches from external
 memory for each step of the tree construction algorithm. On the other hand, the GPU
 algorithm uses a hybrid approach. It iterates through the data during the beginning of
-each iteration and concatenates all batches into one in GPU memory. To reduce overall
-memory usage, users can utilize subsampling. The GPU hist tree method supports
-`gradient-based sampling`, enabling users to set a low sampling rate without compromising
-accuracy.
+each iteration and concatenates all batches into one in GPU memory for performance
+reasons. To reduce overall memory usage, users can utilize subsampling. The GPU hist tree
+method supports `gradient-based sampling`, enabling users to set a low sampling rate
+without compromising accuracy.
 
 .. code-block:: python
 

doc/tutorials/monotonic.rst

@@ -83,13 +83,14 @@ Some other examples:
 - ``(0,-1)``: No constraint on the first predictor and a decreasing constraint on the second.
 
 
-**Note for the 'hist' tree construction algorithm**.
-If ``tree_method`` is set to either ``hist``, ``approx`` or ``gpu_hist``, enabling
-monotonic constraints may produce unnecessarily shallow trees. This is because the
-``hist`` method reduces the number of candidate splits to be considered at each
-split. Monotonic constraints may wipe out all available split candidates, in which case no
-split is made. To reduce the effect, you may want to increase the ``max_bin`` parameter to
-consider more split candidates.
+.. note::
+
+   **Note for the 'hist' tree construction algorithm**.  If ``tree_method`` is set to
+   either ``hist`` or ``approx``, enabling monotonic constraints may produce unnecessarily
+   shallow trees. This is because the ``hist`` method reduces the number of candidate
+   splits to be considered at each split. Monotonic constraints may wipe out all available
+   split candidates, in which case no split is made. To reduce the effect, you may want to
+   increase the ``max_bin`` parameter to consider more split candidates.
 
 
 *******************

doc/tutorials/param_tuning.rst

@@ -38,10 +38,6 @@ There are in general two ways that you can control overfitting in XGBoost:
   - This includes ``subsample`` and ``colsample_bytree``.
   - You can also reduce stepsize ``eta``. Remember to increase ``num_round`` when you do so.
 
-***************************
-Faster training performance
-***************************
-There's a parameter called ``tree_method``, set it to ``hist`` or ``gpu_hist`` for faster computation.
 
 *************************
 Handle Imbalanced Dataset

doc/tutorials/rf.rst

@@ -50,13 +50,14 @@ Here is a sample parameter dictionary for training a random forest on a GPU usin
 xgboost::
 
   params = {
-    'colsample_bynode': 0.8,
-    'learning_rate': 1,
-    'max_depth': 5,
-    'num_parallel_tree': 100,
-    'objective': 'binary:logistic',
-    'subsample': 0.8,
-    'tree_method': 'gpu_hist'
+    "colsample_bynode": 0.8,
+    "learning_rate": 1,
+    "max_depth": 5,
+    "num_parallel_tree": 100,
+    "objective": "binary:logistic",
+    "subsample": 0.8,
+    "tree_method": "hist",
+    "device": "cuda",
   }
 
 A random forest model can then be trained as follows::

doc/tutorials/saving_model.rst

@@ -174,7 +174,7 @@ Will print out something similar to (not actual output as it's too long for demo
           "gbtree_train_param": {
             "num_parallel_tree": "1",
             "process_type": "default",
-            "tree_method": "gpu_hist",
+            "tree_method": "hist",
             "updater": "grow_gpu_hist",
             "updater_seq": "grow_gpu_hist"
           },

python-package/xgboost/sklearn.py

@@ -278,9 +278,15 @@ __model_doc = f"""
           without bias.
 
     device : Optional[str]
-        Device ordinal.
+
+        .. versionadded:: 2.0.0
+
+        Device ordinal, available options are `cpu`, `cuda`, and `gpu`.
+
     validate_parameters : Optional[bool]
+
         Give warnings for unknown parameter.
+
     enable_categorical : bool
 
         .. versionadded:: 1.5.0

python-package/xgboost/spark/estimator.py

@@ -144,8 +144,13 @@ class SparkXGBRegressor(_SparkXGBEstimator):
         .. deprecated:: 2.0.0
 
         Use `device` instead.
+
     device:
+
+        .. versionadded:: 2.0.0
+
         Device for XGBoost workers, available options are `cpu`, `cuda`, and `gpu`.
+
     force_repartition:
         Boolean value to specify if forcing the input dataset to be repartitioned
         before XGBoost training.
@@ -319,8 +324,13 @@ class SparkXGBClassifier(_SparkXGBEstimator, HasProbabilityCol, HasRawPrediction
         .. deprecated:: 2.0.0
 
         Use `device` instead.
+
     device:
+
+        .. versionadded:: 2.0.0
+
         Device for XGBoost workers, available options are `cpu`, `cuda`, and `gpu`.
+
     force_repartition:
         Boolean value to specify if forcing the input dataset to be repartitioned
         before XGBoost training.
@@ -497,8 +507,13 @@ class SparkXGBRanker(_SparkXGBEstimator):
         .. deprecated:: 2.0.0
 
         Use `device` instead.
+
     device:
+
+        .. versionadded:: 2.0.0
+
         Device for XGBoost workers, available options are `cpu`, `cuda`, and `gpu`.
+
     force_repartition:
         Boolean value to specify if forcing the input dataset to be repartitioned
         before XGBoost training.

src/data/data.cc

@@ -724,11 +724,15 @@ void MetaInfo::SynchronizeNumberOfColumns() {
 namespace {
 template <typename T>
 void CheckDevice(std::int32_t device, HostDeviceVector<T> const& v) {
-  CHECK(v.DeviceIdx() == Context::kCpuId || device == Context::kCpuId || v.DeviceIdx() == device)
-      << "Data is resided on a different device than `gpu_id`. "
-      << "Device that data is on: " << v.DeviceIdx() << ", "
-      << "`gpu_id` for XGBoost: " << device;
+  bool valid =
+      v.DeviceIdx() == Context::kCpuId || device == Context::kCpuId || v.DeviceIdx() == device;
+  if (!valid) {
+    LOG(FATAL) << "Invalid device ordinal. Data is associated with a different device ordinal than "
+                  "the booster. The device ordinal of the data is: "
+               << v.DeviceIdx() << "; the device ordinal of the Booster is: " << device;
+  }
 }
+
 template <typename T, std::int32_t D>
 void CheckDevice(std::int32_t device, linalg::Tensor<T, D> const& v) {
   CheckDevice(device, *v.Data());

src/gbm/gbtree.cc

@@ -42,22 +42,22 @@ DMLC_REGISTRY_FILE_TAG(gbtree);
 
 namespace {
 /** @brief Map the `tree_method` parameter to the `updater` parameter. */
-std::string MapTreeMethodToUpdaters(Context const* ctx_, TreeMethod tree_method) {
+std::string MapTreeMethodToUpdaters(Context const* ctx, TreeMethod tree_method) {
   // Choose updaters according to tree_method parameters
+  if (ctx->IsCUDA()) {
+    common::AssertGPUSupport();
+  }
   switch (tree_method) {
     case TreeMethod::kAuto:  // Use hist as default in 2.0
     case TreeMethod::kHist: {
-      return ctx_->DispatchDevice([] { return "grow_quantile_histmaker"; },
-                                  [] {
-                                    common::AssertGPUSupport();
-                                    return "grow_gpu_hist";
-                                  });
+      return ctx->DispatchDevice([] { return "grow_quantile_histmaker"; },
+                                 [] { return "grow_gpu_hist"; });
     }
     case TreeMethod::kApprox:
-      CHECK(ctx_->IsCPU()) << "The `approx` tree method is not supported on GPU.";
+      CHECK(ctx->IsCPU()) << "The `approx` tree method is not supported on GPU.";
       return "grow_histmaker";
     case TreeMethod::kExact:
-      CHECK(ctx_->IsCPU()) << "The `exact` tree method is not supported on GPU.";
+      CHECK(ctx->IsCPU()) << "The `exact` tree method is not supported on GPU.";
       return "grow_colmaker,prune";
     case TreeMethod::kGPUHist: {
       common::AssertGPUSupport();
@@ -150,6 +150,7 @@ void GBTree::Configure(Args const& cfg) {
     CHECK(tparam_.tree_method == TreeMethod::kHist || tparam_.tree_method == TreeMethod::kAuto)
         << "Only the hist tree method is supported for building multi-target trees with vector "
            "leaf.";
+    CHECK(ctx_->IsCPU()) << "GPU is not yet supported for vector leaf.";
   }
 
   LOG(DEBUG) << "Using tree method: " << static_cast<int>(tparam_.tree_method);

tests/ci_build/lint_python.py

@@ -29,10 +29,12 @@ class LintersPaths:
         "tests/python-gpu/load_pickle.py",
         "tests/python-gpu/test_gpu_pickling.py",
         "tests/python-gpu/test_gpu_eval_metrics.py",
+        "tests/python-gpu/test_gpu_with_sklearn.py",
         "tests/test_distributed/test_with_spark/",
         "tests/test_distributed/test_gpu_with_spark/",
         # demo
         "demo/dask/",
+        "demo/rmm_plugin",
         "demo/json-model/json_parser.py",
         "demo/guide-python/cat_in_the_dat.py",
         "demo/guide-python/categorical.py",

tests/python-gpu/test_from_cupy.py

@@ -234,7 +234,7 @@ Arrow specification.'''
         cp.cuda.runtime.setDevice(0)
         dtrain = dmatrix_from_cupy(np.float32, xgb.QuantileDMatrix, np.nan)
         with pytest.raises(
-            xgb.core.XGBoostError, match="Data is resided on a different device"
+            xgb.core.XGBoostError, match="Invalid device ordinal"
         ):
             xgb.train(
                 {'tree_method': 'gpu_hist', 'gpu_id': 1}, dtrain, num_boost_round=10

tests/python-gpu/test_gpu_with_sklearn.py

@@ -2,6 +2,7 @@ import json
 import os
 import sys
 import tempfile
+from concurrent.futures import ThreadPoolExecutor
 
 import numpy as np
 import pytest
@@ -23,18 +24,19 @@ def test_gpu_binary_classification():
     from sklearn.model_selection import KFold
 
     digits = load_digits(n_class=2)
-    y = digits['target']
-    X = digits['data']
+    y = digits["target"]
+    X = digits["data"]
     kf = KFold(n_splits=2, shuffle=True, random_state=rng)
     for cls in (xgb.XGBClassifier, xgb.XGBRFClassifier):
         for train_index, test_index in kf.split(X, y):
             xgb_model = cls(
-                random_state=42, tree_method='gpu_hist',
-                n_estimators=4, gpu_id='0').fit(X[train_index], y[train_index])
+                random_state=42, tree_method="gpu_hist", n_estimators=4, gpu_id="0"
+            ).fit(X[train_index], y[train_index])
             preds = xgb_model.predict(X[test_index])
             labels = y[test_index]
-            err = sum(1 for i in range(len(preds))
-                      if int(preds[i] > 0.5) != labels[i]) / float(len(preds))
+            err = sum(
+                1 for i in range(len(preds)) if int(preds[i] > 0.5) != labels[i]
+            ) / float(len(preds))
             assert err < 0.1
 
 
@@ -133,7 +135,7 @@ def test_classififer():
     X, y = load_digits(return_X_y=True)
     y *= 10
 
-    clf = xgb.XGBClassifier(tree_method="gpu_hist", n_estimators=1)
+    clf = xgb.XGBClassifier(tree_method="hist", n_estimators=1, device="cuda")
 
     # numpy
     with pytest.raises(ValueError, match=r"Invalid classes.*"):
@@ -161,3 +163,46 @@ def test_ranking_qid_df():
     import cudf
 
     run_ranking_qid_df(cudf, "gpu_hist")
+
+
+@pytest.mark.skipif(**tm.no_cupy())
+@pytest.mark.mgpu
+def test_device_ordinal() -> None:
+    import cupy as cp
+
+    n_devices = 2
+
+    def worker(ordinal: int, correct_ordinal: bool) -> None:
+        if correct_ordinal:
+            cp.cuda.runtime.setDevice(ordinal)
+        else:
+            cp.cuda.runtime.setDevice((ordinal + 1) % n_devices)
+
+        X, y, w = tm.make_regression(4096, 12, use_cupy=True)
+        reg = xgb.XGBRegressor(device=f"cuda:{ordinal}", tree_method="hist")
+
+        if correct_ordinal:
+            reg.fit(
+                X, y, sample_weight=w, eval_set=[(X, y)], sample_weight_eval_set=[w]
+            )
+            assert tm.non_increasing(reg.evals_result()["validation_0"]["rmse"])
+            return
+
+        with pytest.raises(ValueError, match="Invalid device ordinal"):
+            reg.fit(
+                X, y, sample_weight=w, eval_set=[(X, y)], sample_weight_eval_set=[w]
+            )
+
+    with ThreadPoolExecutor(max_workers=os.cpu_count()) as executor:
+        futures = []
+        n_trials = 32
+        for i in range(n_trials):
+            fut = executor.submit(
+                worker, ordinal=i % n_devices, correct_ordinal=i % 3 != 0
+            )
+            futures.append(fut)
+
+        for fut in futures:
+            fut.result()
+
+    cp.cuda.runtime.setDevice(0)