Support categorical data for dask functional interface and DQM. (#7043)

* Support categorical data for dask functional interface and DQM.

* Implement categorical data support for GPU GK-merge.
* Add support for dask functional interface.
* Add support for DQM.

* Get newer cupy.

python-package/xgboost/core.py

@@ -321,6 +321,7 @@ class DataIter:
     def __init__(self):
         self._handle = _ProxyDMatrix()
         self.exception = None
+        self.enable_categorical = False
 
     @property
     def proxy(self):
@@ -346,13 +347,12 @@ class DataIter:
             data,
             feature_names=None,
             feature_types=None,
-            enable_categorical=False,
             **kwargs
         ):
             from .data import dispatch_device_quantile_dmatrix_set_data
             from .data import _device_quantile_transform
             data, feature_names, feature_types = _device_quantile_transform(
-                data, feature_names, feature_types, enable_categorical,
+                data, feature_names, feature_types, self.enable_categorical,
             )
             dispatch_device_quantile_dmatrix_set_data(self.proxy, data)
             self.proxy.set_info(
@@ -1106,15 +1106,10 @@ class DeviceQuantileDMatrix(DMatrix):
             data = _transform_dlpack(data)
         if _is_iter(data):
             it = data
-            if enable_categorical:
-                raise NotImplementedError(
-                    "categorical support is not enabled on data iterator."
-                )
         else:
-            it = SingleBatchInternalIter(
-                data=data, enable_categorical=enable_categorical, **meta
-            )
+            it = SingleBatchInternalIter(data=data, **meta)
 
+        it.enable_categorical = enable_categorical
         reset_callback = ctypes.CFUNCTYPE(None, ctypes.c_void_p)(it.reset_wrapper)
         next_callback = ctypes.CFUNCTYPE(
             ctypes.c_int,

python-package/xgboost/dask.py

@@ -182,7 +182,7 @@ def concat(value: Any) -> Any:  # pylint: disable=too-many-return-statements
        lazy_isinstance(value[0], 'cudf.core.series', 'Series'):
         from cudf import concat as CUDF_concat  # pylint: disable=import-error
         return CUDF_concat(value, axis=0)
-    if lazy_isinstance(value[0], 'cupy.core.core', 'ndarray'):
+    if lazy_isinstance(value[0], 'cupy._core.core', 'ndarray'):
         import cupy
         # pylint: disable=c-extension-no-member,no-member
         d = cupy.cuda.runtime.getDevice()
@@ -258,6 +258,7 @@ class DaskDMatrix:
         self.feature_names = feature_names
         self.feature_types = feature_types
         self.missing = missing
+        self.enable_categorical = enable_categorical
 
         if qid is not None and weight is not None:
             raise NotImplementedError("per-group weight is not implemented.")
@@ -265,10 +266,6 @@ class DaskDMatrix:
             raise NotImplementedError(
                 "group structure is not implemented, use qid instead."
             )
-        if enable_categorical:
-            raise NotImplementedError(
-                "categorical support is not enabled on `DaskDMatrix`."
-            )
 
         if len(data.shape) != 2:
             raise ValueError(
@@ -311,7 +308,7 @@ class DaskDMatrix:
         qid: Optional[_DaskCollection] = None,
         feature_weights: Optional[_DaskCollection] = None,
         label_lower_bound: Optional[_DaskCollection] = None,
-        label_upper_bound: Optional[_DaskCollection] = None
+        label_upper_bound: Optional[_DaskCollection] = None,
     ) -> "DaskDMatrix":
         '''Obtain references to local data.'''
 
@@ -430,6 +427,7 @@ class DaskDMatrix:
                 'feature_weights': self.feature_weights,
                 'meta_names': self.meta_names,
                 'missing': self.missing,
+                'enable_categorical': self.enable_categorical,
                 'parts': self.worker_map.get(worker_addr, None),
                 'is_quantile': self.is_quantile}
 
@@ -668,6 +666,7 @@ def _create_device_quantile_dmatrix(
     missing: float,
     parts: Optional[_DataParts],
     max_bin: int,
+    enable_categorical: bool,
 ) -> DeviceQuantileDMatrix:
     worker = distributed.get_worker()
     if parts is None:
@@ -680,6 +679,7 @@ def _create_device_quantile_dmatrix(
             feature_names=feature_names,
             feature_types=feature_types,
             max_bin=max_bin,
+            enable_categorical=enable_categorical,
         )
         return d
 
@@ -709,6 +709,7 @@ def _create_device_quantile_dmatrix(
         feature_types=feature_types,
         nthread=worker.nthreads,
         max_bin=max_bin,
+        enable_categorical=enable_categorical,
     )
     dmatrix.set_info(feature_weights=feature_weights)
     return dmatrix
@@ -720,6 +721,7 @@ def _create_dmatrix(
     feature_weights: Optional[Any],
     meta_names: List[str],
     missing: float,
+    enable_categorical: bool,
     parts: Optional[_DataParts]
 ) -> DMatrix:
     '''Get data that local to worker from DaskDMatrix.
@@ -734,9 +736,12 @@ def _create_dmatrix(
     if list_of_parts is None:
         msg = 'worker {address} has an empty DMatrix.  '.format(address=worker.address)
         LOGGER.warning(msg)
-        d = DMatrix(numpy.empty((0, 0)),
-                    feature_names=feature_names,
-                    feature_types=feature_types)
+        d = DMatrix(
+            numpy.empty((0, 0)),
+            feature_names=feature_names,
+            feature_types=feature_types,
+            enable_categorical=enable_categorical,
+        )
         return d
 
     T = TypeVar('T')
@@ -764,6 +769,7 @@ def _create_dmatrix(
         feature_names=feature_names,
         feature_types=feature_types,
         nthread=worker.nthreads,
+        enable_categorical=enable_categorical,
     )
     dmatrix.set_info(
         base_margin=_base_margin,