[serve] Handle None in ReplicaConfig's resource_dict (#23851)

This change sets `"memory"`'s default to `0` in the `resource_dict` but keeps the default as `None` in `ray_actor_options`. It adds logic to both problematic lines to handle `None` in case of future settings updates. It also adds unit tests to prevent regressions.

python/ray/serve/BUILD

@@ -252,7 +252,7 @@ py_test(
 
 py_test(
     name = "test_standalone2",
-    size = "medium",
+    size = "large",
     srcs = serve_tests_srcs,
     tags = ["exclusive", "team:serve"],
     deps = [":serve_lib"],

python/ray/serve/config.py

@@ -297,7 +297,7 @@ class ReplicaConfig:
 
         # TODO(suquark): reuse options validation of remote function/actor.
         # Ray defaults to zero CPUs for placement, we default to one here.
-        if self.ray_actor_options.get("num_cpus", None) is None:
+        if self.ray_actor_options.get("num_cpus") is None:
             self.ray_actor_options["num_cpus"] = 1
         num_cpus = self.ray_actor_options["num_cpus"]
         if not isinstance(num_cpus, (int, float)):
@@ -306,7 +306,7 @@ class ReplicaConfig:
             raise ValueError("num_cpus in ray_actor_options must be >= 0.")
         self.resource_dict["CPU"] = num_cpus
 
-        if self.ray_actor_options.get("num_gpus", None) is None:
+        if self.ray_actor_options.get("num_gpus") is None:
             self.ray_actor_options["num_gpus"] = 0
         num_gpus = self.ray_actor_options["num_gpus"]
         if not isinstance(num_gpus, (int, float)):
@@ -315,6 +315,7 @@ class ReplicaConfig:
             raise ValueError("num_gpus in ray_actor_options must be >= 0.")
         self.resource_dict["GPU"] = num_gpus
 
+        # Serve deployments use Ray's default for actor memory.
         self.ray_actor_options.setdefault("memory", None)
         memory = self.ray_actor_options["memory"]
         if memory is not None and not isinstance(memory, (int, float)):
@@ -335,7 +336,7 @@ class ReplicaConfig:
             raise ValueError("object_store_memory in ray_actor_options must be >= 0.")
         self.resource_dict["object_store_memory"] = object_store_memory
 
-        if self.ray_actor_options.get("resources", None) is None:
+        if self.ray_actor_options.get("resources") is None:
             self.ray_actor_options["resources"] = {}
         custom_resources = self.ray_actor_options["resources"]
         if not isinstance(custom_resources, dict):

python/ray/serve/deployment_state.py

@@ -267,9 +267,11 @@ class ActorReplicaWrapper:
         )
 
         self._actor_resources = deployment_info.replica_config.resource_dict
-        # it is currently not possiible to create a placement group
+        # it is currently not possible to create a placement group
         # with no resources (https://github.com/ray-project/ray/issues/20401)
-        has_resources_assigned = all((r > 0 for r in self._actor_resources.values()))
+        has_resources_assigned = all(
+            (r is not None and r > 0 for r in self._actor_resources.values())
+        )
         if USE_PLACEMENT_GROUP and has_resources_assigned:
             self._placement_group = self.create_placement_group(
                 self._placement_group_name, self._actor_resources
@@ -752,7 +754,11 @@ class DeploymentReplica(VersionedReplica):
         if self._actor.actor_resources is None:
             return "UNKNOWN", "UNKNOWN"
 
-        required = {k: v for k, v in self._actor.actor_resources.items() if v > 0}
+        required = {
+            k: v
+            for k, v in self._actor.actor_resources.items()
+            if v is not None and v > 0
+        }
         available = {
             k: v for k, v in self._actor.available_resources.items() if k in required
         }

python/ray/serve/tests/test_config.py

@@ -184,7 +184,12 @@ def test_replica_config_validation():
 
 @pytest.mark.parametrize(
     "memory_omitted_options",
-    [None, {}, {"CPU": 1, "GPU": 3}, {"CPU": 1, "GPU": 3, "memory": None}],
+    [
+        None,
+        {},
+        {"num_cpus": 1, "num_gpus": 3},
+        {"num_cpus": 1, "num_gpus": 3, "memory": None},
+    ],
 )
 def test_replica_config_default_memory_none(memory_omitted_options):
     """Checks that ReplicaConfig's default memory is None."""

python/ray/serve/tests/test_deployment_state.py

@@ -21,6 +21,7 @@ from ray.serve.deployment_state import (
     DeploymentState,
     DeploymentStateManager,
     DeploymentVersion,
+    DeploymentReplica,
     ReplicaStartupStatus,
     ReplicaState,
     ReplicaStateContainer,
@@ -2086,5 +2087,21 @@ def test_stopping_replicas_ranking():
     compare([2, 2, 3, 3], [2, 2, 3, 3])  # if equal, ordering should be kept
 
 
+def test_resource_requirements_none(mock_deployment_state):
+    """Ensure resource_requirements doesn't break if a requirement is None"""
+
+    class FakeActor:
+
+        actor_resources = {"num_cpus": 2, "fake": None}
+        available_resources = {}
+
+    # Make a DeploymentReplica just to accesss its resource_requirement function
+    replica = DeploymentReplica(None, None, None, None, None)
+    replica._actor = FakeActor()
+
+    # resource_requirements() should not error
+    replica.resource_requirements()
+
+
 if __name__ == "__main__":
     sys.exit(pytest.main(["-v", "-s", __file__]))

python/ray/serve/tests/test_standalone2.py

@@ -52,6 +52,20 @@ def test_standalone_actor_outside_serve():
     ray.shutdown()
 
 
+def test_memory_omitted_option(ray_shutdown):
+    """Ensure that omitting memory doesn't break the deployment."""
+
+    @serve.deployment(ray_actor_options={"num_cpus": 1, "num_gpus": 1})
+    def hello(*args, **kwargs):
+        return "world"
+
+    ray.init(num_gpus=3, namespace="serve")
+    serve.start()
+    hello.deploy()
+
+    assert ray.get(hello.get_handle().remote()) == "world"
+
+
 @pytest.mark.parametrize("detached", [True, False])
 def test_override_namespace(shutdown_ray, detached):
     """Test the _override_controller_namespace flag in serve.start()."""