Add support for dlpack, expose python docs for DeviceQuantileDMatrix (#5465)

doc/python/python_api.rst

@@ -14,6 +14,9 @@ Core Data Structure
     :members:
     :show-inheritance:
 
+.. autoclass:: xgboost.DeviceQuantileDMatrix
+    :show-inheritance:
+
 .. autoclass:: xgboost.Booster
     :members:
     :show-inheritance:

python-package/xgboost/core.py

@@ -381,6 +381,17 @@ def _maybe_dt_data(data, feature_names, feature_types,
 
     return data, feature_names, feature_types
 
+def _is_dlpack(x):
+    return 'PyCapsule' in str(type(x)) and "dltensor" in str(x)
+
+# Just convert dlpack into cupy (zero copy)
+def _maybe_dlpack_data(data, feature_names, feature_types):
+    if not _is_dlpack(data):
+        return data, feature_names, feature_types
+    from cupy import fromDlpack # pylint: disable=E0401
+    data = fromDlpack(data)
+    return data, feature_names, feature_types
+
 
 def _convert_dataframes(data, feature_names, feature_types,
                         meta=None, meta_type=None):
@@ -399,6 +410,9 @@ def _convert_dataframes(data, feature_names, feature_types,
     data, feature_names, feature_types = _maybe_cudf_dataframe(
         data, feature_names, feature_types)
 
+    data, feature_names, feature_types = _maybe_dlpack_data(
+        data, feature_names, feature_types)
+
     return data, feature_names, feature_types
 
 
@@ -439,7 +453,7 @@ class DMatrix(object):
         """Parameters
         ----------
         data : os.PathLike/string/numpy.array/scipy.sparse/pd.DataFrame/
-               dt.Frame/cudf.DataFrame/cupy.array
+               dt.Frame/cudf.DataFrame/cupy.array/dlpack
             Data source of DMatrix.
             When data is string or os.PathLike type, it represents the path
             libsvm format txt file, csv file (by specifying uri parameter
@@ -1028,12 +1042,12 @@ class DMatrix(object):
 class DeviceQuantileDMatrix(DMatrix):
     """Device memory Data Matrix used in XGBoost for training with tree_method='gpu_hist'. Do not
     use this for test/validation tasks as some information may be lost in quantisation. This
-    DMatrix is primarily designed to save memory in training and avoids intermediate steps,
+    DMatrix is primarily designed to save memory in training from device memory inputs by
-    directly creating a compressed representation for training without allocating additional
+    avoiding intermediate storage. Implementation does not currently consider weights in
-    memory. Implementation does not currently consider weights in quantisation process(unlike
+    quantisation process(unlike DMatrix). Set max_bin to control the number of bins during
-    DMatrix).
+    quantisation.
 
-    You can construct DeviceDMatrix from cupy/cudf
+    You can construct DeviceQuantileDMatrix from cupy/cudf/dlpack.
     """
 
     def __init__(self, data, label=None, weight=None, base_margin=None,
@@ -1044,8 +1058,8 @@ class DeviceQuantileDMatrix(DMatrix):
                  nthread=None, max_bin=256):
         self.max_bin = max_bin
         if not (hasattr(data, "__cuda_array_interface__") or (
-                CUDF_INSTALLED and isinstance(data, CUDF_DataFrame))):
+                CUDF_INSTALLED and isinstance(data, CUDF_DataFrame)) or _is_dlpack(data)):
-            raise ValueError('Only cupy/cudf currently supported for DeviceDMatrix')
+            raise ValueError('Only cupy/cudf/dlpack currently supported for DeviceQuantileDMatrix')
 
         super().__init__(data, label=label, weight=weight, base_margin=base_margin,
                          missing=missing,

tests/python-gpu/test_from_cupy.py

@@ -95,7 +95,7 @@ def _test_cupy_metainfo(DMatrixT):
     assert np.array_equal(dmat.get_uint_info('group_ptr'), dmat_cupy.get_uint_info('group_ptr'))
 
 
-class TestFromArrayInterface:
+class TestFromCupy:
     '''Tests for constructing DMatrix from data structure conforming Apache
 Arrow specification.'''
 
@@ -122,3 +122,17 @@ Arrow specification.'''
     @pytest.mark.skipif(**tm.no_cupy())
     def test_cupy_metainfo_device_dmat(self):
         _test_cupy_metainfo(xgb.DeviceQuantileDMatrix)
+
+    @pytest.mark.skipif(**tm.no_cupy())
+    def test_dlpack_simple_dmat(self):
+        import cupy as cp
+        n = 100
+        X = cp.random.random((n, 2))
+        xgb.DMatrix(X.toDlpack())
+
+    @pytest.mark.skipif(**tm.no_cupy())
+    def test_dlpack_device_dmat(self):
+        import cupy as cp
+        n = 100
+        X = cp.random.random((n, 2))
+        xgb.DeviceQuantileDMatrix(X.toDlpack())