[backport] [pyspark] support gpu transform (#9542) (#9559)

---------

Co-authored-by: Bobby Wang <[email protected]>

python-package/xgboost/compat.py

@@ -88,6 +88,18 @@ def is_cudf_available() -> bool:
         return False
 
 
+def is_cupy_available() -> bool:
+    """Check cupy package available or not"""
+    if importlib.util.find_spec("cupy") is None:
+        return False
+    try:
+        import cupy
+
+        return True
+    except ImportError:
+        return False
+
+
 try:
     import scipy.sparse as scipy_sparse
     from scipy.sparse import csr_matrix as scipy_csr

python-package/xgboost/spark/core.py

@@ -59,7 +59,7 @@
 
 import xgboost
 from xgboost import XGBClassifier
-from xgboost.compat import is_cudf_available
+from xgboost.compat import is_cudf_available, is_cupy_available
 from xgboost.core import Booster, _check_distributed_params
 from xgboost.sklearn import DEFAULT_N_ESTIMATORS, XGBModel, _can_use_qdm
 from xgboost.training import train as worker_train
@@ -242,6 +242,13 @@ class _SparkXGBParams(
         TypeConverters.toList,
     )
 
+    def set_device(self, value: str) -> "_SparkXGBParams":
+        """Set device, optional value: cpu, cuda, gpu"""
+        _check_distributed_params({"device": value})
+        assert value in ("cpu", "cuda", "gpu")
+        self.set(self.device, value)
+        return self
+
     @classmethod
     def _xgb_cls(cls) -> Type[XGBModel]:
         """
@@ -1193,6 +1200,31 @@ def _post_transform(self, dataset: DataFrame, pred_col: Column) -> DataFrame:
             dataset = dataset.drop(pred_struct_col)
         return dataset
 
+    def _gpu_transform(self) -> bool:
+        """If gpu is used to do the prediction, true to gpu prediction"""
+
+        if _is_local(_get_spark_session().sparkContext):
+            # if it's local model, we just use the internal "device"
+            return use_cuda(self.getOrDefault(self.device))
+
+        gpu_per_task = (
+            _get_spark_session()
+            .sparkContext.getConf()
+            .get("spark.task.resource.gpu.amount")
+        )
+
+        # User don't set gpu configurations, just use cpu
+        if gpu_per_task is None:
+            if use_cuda(self.getOrDefault(self.device)):
+                get_logger("XGBoost-PySpark").warning(
+                    "Do the prediction on the CPUs since "
+                    "no gpu configurations are set"
+                )
+            return False
+
+        # User already sets the gpu configurations, we just use the internal "device".
+        return use_cuda(self.getOrDefault(self.device))
+
     def _transform(self, dataset: DataFrame) -> DataFrame:
         # pylint: disable=too-many-statements, too-many-locals
         # Save xgb_sklearn_model and predict_params to be local variable
@@ -1216,21 +1248,77 @@ def _transform(self, dataset: DataFrame) -> DataFrame:
 
         _, schema = self._out_schema()
 
+        is_local = _is_local(_get_spark_session().sparkContext)
+        run_on_gpu = self._gpu_transform()
+
         @pandas_udf(schema)  # type: ignore
         def predict_udf(iterator: Iterator[pd.DataFrame]) -> Iterator[pd.Series]:
             assert xgb_sklearn_model is not None
             model = xgb_sklearn_model
+
+            from pyspark import TaskContext
+
+            context = TaskContext.get()
+            assert context is not None
+
+            dev_ordinal = -1
+
+            if is_cudf_available():
+                if is_local:
+                    if run_on_gpu and is_cupy_available():
+                        import cupy as cp  # pylint: disable=import-error
+
+                        total_gpus = cp.cuda.runtime.getDeviceCount()
+                        if total_gpus > 0:
+                            partition_id = context.partitionId()
+                            # For transform local mode, default the dev_ordinal to
+                            # (partition id) % gpus.
+                            dev_ordinal = partition_id % total_gpus
+                elif run_on_gpu:
+                    dev_ordinal = _get_gpu_id(context)
+
+                if dev_ordinal >= 0:
+                    device = "cuda:" + str(dev_ordinal)
+                    get_logger("XGBoost-PySpark").info(
+                        "Do the inference with device: %s", device
+                    )
+                    model.set_params(device=device)
+                else:
+                    get_logger("XGBoost-PySpark").info("Do the inference on the CPUs")
+            else:
+                msg = (
+                    "CUDF is unavailable, fallback the inference on the CPUs"
+                    if run_on_gpu
+                    else "Do the inference on the CPUs"
+                )
+                get_logger("XGBoost-PySpark").info(msg)
+
+            def to_gpu_if_possible(data: ArrayLike) -> ArrayLike:
+                """Move the data to gpu if possible"""
+                if dev_ordinal >= 0:
+                    import cudf  # pylint: disable=import-error
+                    import cupy as cp  # pylint: disable=import-error
+
+                    # We must set the device after import cudf, which will change the device id to 0
+                    # See https://github.com/rapidsai/cudf/issues/11386
+                    cp.cuda.runtime.setDevice(dev_ordinal)  # pylint: disable=I1101
+                    df = cudf.DataFrame(data)
+                    del data
+                    return df
+                return data
+
             for data in iterator:
                 if enable_sparse_data_optim:
                     X = _read_csr_matrix_from_unwrapped_spark_vec(data)
                 else:
                     if feature_col_names is not None:
-                        X = data[feature_col_names]
+                        tmp = data[feature_col_names]
                     else:
-                        X = stack_series(data[alias.data])
+                        tmp = stack_series(data[alias.data])
+                    X = to_gpu_if_possible(tmp)
 
                 if has_base_margin:
-                    base_margin = data[alias.margin].to_numpy()
+                    base_margin = to_gpu_if_possible(data[alias.margin])
                 else:
                     base_margin = None
 

python-package/xgboost/spark/utils.py

@@ -10,7 +10,7 @@
 from typing import Any, Callable, Dict, Optional, Set, Type
 
 import pyspark
-from pyspark import BarrierTaskContext, SparkContext, SparkFiles
+from pyspark import BarrierTaskContext, SparkContext, SparkFiles, TaskContext
 from pyspark.sql.session import SparkSession
 
 from xgboost import Booster, XGBModel, collective
@@ -129,7 +129,7 @@ def _is_local(spark_context: SparkContext) -> bool:
     return spark_context._jsc.sc().isLocal()
 
 
-def _get_gpu_id(task_context: BarrierTaskContext) -> int:
+def _get_gpu_id(task_context: TaskContext) -> int:
     """Get the gpu id from the task resources"""
     if task_context is None:
         # This is a safety check.

tests/test_distributed/test_gpu_with_spark/test_gpu_spark.py

@@ -2,6 +2,7 @@
 import logging
 import subprocess
 
+import numpy as np
 import pytest
 import sklearn
 
@@ -13,7 +14,7 @@
 from pyspark.ml.tuning import CrossValidator, ParamGridBuilder
 from pyspark.sql import SparkSession
 
-from xgboost.spark import SparkXGBClassifier, SparkXGBRegressor
+from xgboost.spark import SparkXGBClassifier, SparkXGBRegressor, SparkXGBRegressorModel
 
 gpu_discovery_script_path = "tests/test_distributed/test_gpu_with_spark/discover_gpu.sh"
 
@@ -242,3 +243,33 @@ def test_sparkxgb_regressor_feature_cols_with_gpu(spark_diabetes_dataset_feature
     evaluator = RegressionEvaluator(metricName="rmse")
     rmse = evaluator.evaluate(pred_result_df)
     assert rmse <= 65.0
+
+
+def test_gpu_transform(spark_diabetes_dataset) -> None:
+    regressor = SparkXGBRegressor(device="cuda", num_workers=num_workers)
+    train_df, test_df = spark_diabetes_dataset
+    model: SparkXGBRegressorModel = regressor.fit(train_df)
+
+    # The model trained with GPUs, and transform with GPU configurations.
+    assert model._gpu_transform()
+
+    model.set_device("cpu")
+    assert not model._gpu_transform()
+    # without error
+    cpu_rows = model.transform(test_df).select("prediction").collect()
+
+    regressor = SparkXGBRegressor(device="cpu", num_workers=num_workers)
+    model = regressor.fit(train_df)
+
+    # The model trained with CPUs. Even with GPU configurations,
+    # still prefer transforming with CPUs
+    assert not model._gpu_transform()
+
+    # Set gpu transform explicitly.
+    model.set_device("cuda")
+    assert model._gpu_transform()
+    # without error
+    gpu_rows = model.transform(test_df).select("prediction").collect()
+
+    for cpu, gpu in zip(cpu_rows, gpu_rows):
+        np.testing.assert_allclose(cpu.prediction, gpu.prediction, atol=1e-3)

tests/test_distributed/test_with_spark/test_spark_local.py

@@ -888,6 +888,34 @@ def test_device_param(self, reg_data: RegData, clf_data: ClfData) -> None:
         clf = SparkXGBClassifier(device="cuda")
         clf._validate_params()
 
+    def test_gpu_transform(self, clf_data: ClfData) -> None:
+        """local mode"""
+        classifier = SparkXGBClassifier(device="cpu")
+        model: SparkXGBClassifierModel = classifier.fit(clf_data.cls_df_train)
+
+        with tempfile.TemporaryDirectory() as tmpdir:
+            path = "file:" + tmpdir
+            model.write().overwrite().save(path)
+
+            # The model trained with CPU, transform defaults to cpu
+            assert not model._gpu_transform()
+
+            # without error
+            model.transform(clf_data.cls_df_test).collect()
+
+            model.set_device("cuda")
+            assert model._gpu_transform()
+
+            model_loaded = SparkXGBClassifierModel.load(path)
+
+            # The model trained with CPU, transform defaults to cpu
+            assert not model_loaded._gpu_transform()
+            # without error
+            model_loaded.transform(clf_data.cls_df_test).collect()
+
+            model_loaded.set_device("cuda")
+            assert model_loaded._gpu_transform()
+
 
 class XgboostLocalTest(SparkTestCase):
     def setUp(self):