[air] Tuner should use run_config from Trainer per default

python/ray/tune/impl/tuner_internal.py

@@ -43,7 +43,8 @@ class TunerInternal:
         tune_config: Tuning algorithm specific configs.
             Refer to ray.tune.tune_config.TuneConfig for more info.
         run_config: Runtime configuration that is specific to individual trials.
-            Refer to ray.ml.config.RunConfig for more info.
+            If passed, this will overwrite the run config passed to the Trainer,
+            if applicable. Refer to ray.ml.config.RunConfig for more info.
     """
 
     def __init__(
@@ -82,6 +83,11 @@ def __init__(
         if not trainable:
             raise TuneError("You need to provide a trainable to tune.")
 
+        # If no run config was passed to Tuner directly, use the one from the Trainer,
+        # if available
+        if not run_config and isinstance(trainable, Trainer):
+            run_config = trainable.run_config
+
         self._is_restored = False
         self._trainable = trainable
         self._tune_config = tune_config or TuneConfig()

python/ray/tune/tests/test_tuner.py

@@ -21,6 +21,20 @@
 from ray.tune.tuner import Tuner
 
 
+class DummyTrainer(Trainer):
+    _scaling_config_allowed_keys = [
+        "num_workers",
+        "num_cpus_per_worker",
+        "num_gpus_per_worker",
+        "additional_resources_per_worker",
+        "use_gpu",
+        "trainer_resources",
+    ]
+
+    def training_loop(self) -> None:
+        raise RuntimeError("There is an error in trainer!")
+
+
 class TestDatasource(Datasource):
     def __init__(self, do_shuffle: bool):
         self._shuffle = do_shuffle
@@ -174,19 +188,6 @@ def on_step_end(self, iteration, trials, **kwargs):
         assert len(results) == 2
 
     def test_tuner_trainer_fail(self):
-        class DummyTrainer(Trainer):
-            _scaling_config_allowed_keys = [
-                "num_workers",
-                "num_cpus_per_worker",
-                "num_gpus_per_worker",
-                "additional_resources_per_worker",
-                "use_gpu",
-                "trainer_resources",
-            ]
-
-            def training_loop(self) -> None:
-                raise RuntimeError("There is an error in trainer!")
-
         trainer = DummyTrainer()
         param_space = {
             "scaling_config": {
@@ -245,6 +246,12 @@ def test_tuner_with_torch_trainer(self):
         results = tuner.fit()
         assert len(results) == 8
 
+    def test_tuner_run_config_override(self):
+        trainer = DummyTrainer(run_config=RunConfig(stop={"metric": 4}))
+        tuner = Tuner(trainer)
+
+        assert tuner._local_tuner._run_config.stop == {"metric": 4}
+
 
 if __name__ == "__main__":
     import pytest

python/ray/tune/tuner.py

@@ -35,7 +35,8 @@ class Tuner:
         tune_config: Tuning algorithm specific configs.
             Refer to ray.tune.tune_config.TuneConfig for more info.
         run_config: Runtime configuration that is specific to individual trials.
-            Refer to ray.ml.config.RunConfig for more info.
+            If passed, this will overwrite the run config passed to the Trainer,
+            if applicable. Refer to ray.ml.config.RunConfig for more info.
 
     Usage pattern: