diff --git a/python/cuml/test/test_metrics.py b/python/cuml/test/test_metrics.py
index 22fc041a05..800b2c9771 100644
--- a/python/cuml/test/test_metrics.py
+++ b/python/cuml/test/test_metrics.py
@@ -1373,13 +1373,13 @@ def test_hinge_loss(nrows, ncols, n_info, input_type, n_classes):
                              stress_param(500000000)
                          ])
 @pytest.mark.parametrize("input_type", ["cudf", "cupy"])
-@pytest.mark.parametrize("dtype", [cp.float32, cp.float64])
-def test_kl_divergence(nfeatures, input_type, dtype):
+@pytest.mark.parametrize("dtypeP", [cp.float32, cp.float64])
+@pytest.mark.parametrize("dtypeQ", [cp.float32, cp.float64])
+def test_kl_divergence(nfeatures, input_type, dtypeP, dtypeQ):
     if not has_scipy():
-        pytest.skip('Skipping test_entropy_random because Scipy is missing')
+        pytest.skip('Skipping test_kl_divergence because Scipy is missing')
 
     from scipy.stats import entropy as sp_entropy
-    # Test larger sizes to sklearn
     rng = np.random.RandomState(5)
 
     P = rng.random_sample((nfeatures))
@@ -1390,11 +1390,17 @@ def test_kl_divergence(nfeatures, input_type, dtype):
     sk_res = sp_entropy(P, Q)
 
     if input_type == "cudf":
-        P = cudf.DataFrame(P, dtype=dtype)
-        Q = cudf.DataFrame(Q, dtype=dtype)
+        P = cudf.DataFrame(P, dtype=dtypeP)
+        Q = cudf.DataFrame(Q, dtype=dtypeQ)
     elif input_type == "cupy":
-        P = cp.asarray(P, dtype=dtype)
-        Q = cp.asarray(Q, dtype=dtype)
-    cu_res = cu_kl_divergence(P, Q)
+        P = cp.asarray(P, dtype=dtypeP)
+        Q = cp.asarray(Q, dtype=dtypeQ)
+
+    if dtypeP != dtypeQ:
+        with pytest.raises(TypeError):
+            cu_kl_divergence(P, Q, convert_dtype=False)
+        cu_res = cu_kl_divergence(P, Q)
+    else:
+        cu_res = cu_kl_divergence(P, Q, convert_dtype=False)
 
     cp.testing.assert_array_almost_equal(cu_res, sk_res)