Support CPU input for device QuantileDMatrix. (#8136)

- Copy `GHistIndexMatrix` to `Ellpack` when needed.

tests/python-gpu/test_device_quantile_dmatrix.py

@@ -31,6 +31,34 @@ class TestDeviceQuantileDMatrix:
         data = cp.random.randn(5, 5)
         xgb.DeviceQuantileDMatrix(data, cp.ones(5, dtype=np.float64))
 
+    @pytest.mark.skipif(**tm.no_cupy())
+    def test_from_host(self) -> None:
+        import cupy as cp
+        n_samples = 64
+        n_features = 3
+        X, y, w = tm.make_batches(
+            n_samples, n_features=n_features, n_batches=1, use_cupy=False
+        )
+        Xy = xgb.QuantileDMatrix(X[0], y[0], weight=w[0])
+        booster_0 = xgb.train({"tree_method": "gpu_hist"}, Xy, num_boost_round=4)
+
+        X[0] = cp.array(X[0])
+        y[0] = cp.array(y[0])
+        w[0] = cp.array(w[0])
+
+        Xy = xgb.QuantileDMatrix(X[0], y[0], weight=w[0])
+        booster_1 = xgb.train({"tree_method": "gpu_hist"}, Xy, num_boost_round=4)
+        cp.testing.assert_allclose(
+            booster_0.inplace_predict(X[0]), booster_1.inplace_predict(X[0])
+        )
+
+        with pytest.raises(ValueError, match="not initialized with CPU"):
+            # Training on CPU with GPU data is not supported.
+            xgb.train({"tree_method": "hist"}, Xy, num_boost_round=4)
+
+        with pytest.raises(ValueError, match=r"Only.*hist.*"):
+            xgb.train({"tree_method": "approx"}, Xy, num_boost_round=4)
+
     @pytest.mark.skipif(**tm.no_cupy())
     def test_metainfo(self) -> None:
         import cupy as cp