Export Python Interface for external memory. (#7070)

* Add Python iterator interface.
* Add tests.
* Add demo.
* Add documents.
* Handle empty dataset.

tests/python-gpu/test_gpu_data_iterator.py

@@ -0,0 +1,32 @@
+import numpy as np
+import xgboost as xgb
+from hypothesis import given, strategies, settings
+import pytest
+import sys
+
+sys.path.append("tests/python")
+from test_data_iterator import SingleBatch, make_batches
+from test_data_iterator import test_single_batch as cpu_single_batch
+from test_data_iterator import run_data_iterator
+from testing import IteratorForTest, no_cupy
+
+
+def test_gpu_single_batch() -> None:
+    cpu_single_batch("gpu_hist")
+
+
+@pytest.mark.skipif(**no_cupy())
+@given(
+    strategies.integers(0, 1024), strategies.integers(1, 7), strategies.integers(0, 13)
+)
+@settings(deadline=None)
+def test_gpu_data_iterator(
+    n_samples_per_batch: int, n_features: int, n_batches: int
+) -> None:
+    run_data_iterator(n_samples_per_batch, n_features, n_batches, "gpu_hist", True)
+    run_data_iterator(n_samples_per_batch, n_features, n_batches, "gpu_hist", False)
+
+
+def test_cpu_data_iterator() -> None:
+    """Make sure CPU algorithm can handle GPU inputs"""
+    run_data_iterator(1024, 2, 3, "approx", True)

tests/python-gpu/test_gpu_demos.py

@@ -9,7 +9,7 @@ import test_demos as td         # noqa
 
 @pytest.mark.skipif(**tm.no_cupy())
 def test_data_iterator():
-    script = os.path.join(td.PYTHON_DEMO_DIR, 'data_iterator.py')
+    script = os.path.join(td.PYTHON_DEMO_DIR, 'quantile_data_iterator.py')
     cmd = ['python', script]
     subprocess.check_call(cmd)
 

tests/python-gpu/test_gpu_updaters.py

@@ -112,7 +112,6 @@ class TestGPUUpdaters:
            tm.dataset_strategy)
     @settings(deadline=None)
     def test_external_memory(self, param, num_rounds, dataset):
-        pytest.xfail(reason='TestGPUUpdaters::test_external_memory is flaky')
         # We cannot handle empty dataset yet
         assume(len(dataset.y) > 0)
         param['tree_method'] = 'gpu_hist'