[breaking] Remove the predictor param, allow fallback to prediction using DMatrix. (#9129)

- A `DeviceOrd` struct is implemented to indicate the device. It will eventually replace the `gpu_id` parameter.
- The `predictor` parameter is removed.
- Fallback to `DMatrix` when `inplace_predict` is not available.
- The heuristic for choosing a predictor is only used during training.

tests/python/test_predict.py

@@ -28,7 +28,7 @@ def run_threaded_predict(X, rows, predict_func):
         assert f.result()
 
 
-def run_predict_leaf(predictor):
+def run_predict_leaf(gpu_id: int) -> np.ndarray:
     rows = 100
     cols = 4
     classes = 5
@@ -42,13 +42,13 @@ def run_predict_leaf(predictor):
         {
             "num_parallel_tree": num_parallel_tree,
             "num_class": classes,
-            "predictor": predictor,
             "tree_method": "hist",
         },
         m,
         num_boost_round=num_boost_round,
     )
 
+    booster = tm.set_ordinal(gpu_id, booster)
     empty = xgb.DMatrix(np.ones(shape=(0, cols)))
     empty_leaf = booster.predict(empty, pred_leaf=True)
     assert empty_leaf.shape[0] == 0
@@ -74,13 +74,14 @@ def run_predict_leaf(predictor):
 
     # When there's only 1 tree, the output is a 1 dim vector
     booster = xgb.train({"tree_method": "hist"}, num_boost_round=1, dtrain=m)
+    booster = tm.set_ordinal(gpu_id, booster)
     assert booster.predict(m, pred_leaf=True).shape == (rows,)
 
     return leaf
 
 
-def test_predict_leaf():
-    run_predict_leaf("cpu_predictor")
+def test_predict_leaf() -> None:
+    run_predict_leaf(-1)
 
 
 def test_predict_shape():

tests/python/test_updaters.py

@@ -274,7 +274,7 @@ class TestTreeMethod:
     ) -> None:
         parameters: Dict[str, Any] = {"tree_method": tree_method}
         cat, label = tm.make_categorical(
-            n_samples=rows, n_features=cols, n_categories=cats, onehot=False, sparsity=0.5
+            rows, n_features=cols, n_categories=cats, onehot=False, sparsity=0.5
         )
         Xy = xgb.DMatrix(cat, label, enable_categorical=True)
 
@@ -294,7 +294,9 @@ class TestTreeMethod:
             y_predt = booster.predict(Xy)
 
             rmse = tm.root_mean_square(label, y_predt)
-            np.testing.assert_allclose(rmse, evals_result["Train"]["rmse"][-1])
+            np.testing.assert_allclose(
+                rmse, evals_result["Train"]["rmse"][-1], rtol=2e-5
+            )
 
         # Test with OHE split
         run(self.USE_ONEHOT)
@@ -311,10 +313,8 @@ class TestTreeMethod:
         by_etl_results: Dict[str, Dict[str, List[float]]] = {}
         by_builtin_results: Dict[str, Dict[str, List[float]]] = {}
 
-        predictor = "gpu_predictor" if tree_method == "gpu_hist" else None
         parameters: Dict[str, Any] = {
             "tree_method": tree_method,
-            "predictor": predictor,
             # Use one-hot exclusively
             "max_cat_to_onehot": self.USE_ONEHOT
         }

tests/python/test_with_sklearn.py

@@ -1418,23 +1418,6 @@ def test_categorical():
     np.testing.assert_allclose(predt_cat, predt_enc)
 
 
-def test_prediction_config():
-    reg = xgb.XGBRegressor()
-    assert reg._can_use_inplace_predict() is True
-
-    reg.set_params(predictor="cpu_predictor")
-    assert reg._can_use_inplace_predict() is False
-
-    reg.set_params(predictor="auto")
-    assert reg._can_use_inplace_predict() is True
-
-    reg.set_params(predictor=None)
-    assert reg._can_use_inplace_predict() is True
-
-    reg.set_params(booster="gblinear")
-    assert reg._can_use_inplace_predict() is False
-
-
 def test_evaluation_metric():
     from sklearn.datasets import load_diabetes, load_digits
     from sklearn.metrics import mean_absolute_error