fix: Model Garden deploy_publisher_model now honors dedicated_endpoint_disabled and spot without a machine type

agentplatform/_genai/model_garden.py

@@ -1573,9 +1573,11 @@ def _prepare_deploy_request(
         endpoint_config = types.DeployRequestEndpointConfig(
             endpoint_display_name=config.endpoint_display_name,
         )
-        disabled = config.dedicated_endpoint_disabled
-        if disabled is not None:
-            endpoint_config.dedicated_endpoint_enabled = not disabled
+        # Send the current field, as vertexai.model_garden does. The deprecated
+        # dedicated_endpoint_enabled cannot express "off": False is its proto3
+        # default, so the backend would keep its default (dedicated endpoint on).
+        if config.dedicated_endpoint_disabled:
+            endpoint_config.dedicated_endpoint_disabled = True
         if config.enable_private_service_connect:
             endpoint_config.private_service_connect_config = (
                 types.PrivateServiceConnectConfig(
@@ -1598,6 +1600,10 @@ def _prepare_deploy_request(
                 max_replica_count=config.max_replica_count,
                 spot=config.spot,
             )
+        elif config.spot:
+            # Without a machine shape the backend deploys the model's default
+            # verified config and applies spot to it, as with vertexai.model_garden.
+            deploy_config.dedicated_resources = types.DedicatedResources(spot=True)
 
         return model_config, endpoint_config, deploy_config
 

agentplatform/_genai/types/common.py

@@ -32128,7 +32128,11 @@ class DeployPublisherModelConfig(_common.BaseModel):
     accelerator_count: Optional[int] = Field(
         default=None, description="""Number of accelerators per replica."""
     )
-    spot: Optional[bool] = Field(default=None, description="""Schedule on Spot VMs.""")
+    spot: Optional[bool] = Field(
+        default=None,
+        description="""Schedule on Spot VMs. Without ``machine_type``, applies to the
+      model's default machine.""",
+    )
     dedicated_endpoint_disabled: Optional[bool] = Field(
         default=None,
         description="""Set True to serve predictions via the shared endpoint DNS
@@ -32217,7 +32221,8 @@ class DeployPublisherModelConfigDict(TypedDict, total=False):
     """Number of accelerators per replica."""
 
     spot: Optional[bool]
-    """Schedule on Spot VMs."""
+    """Schedule on Spot VMs. Without ``machine_type``, applies to the
+      model's default machine."""
 
     dedicated_endpoint_disabled: Optional[bool]
     """Set True to serve predictions via the shared endpoint DNS

tests/unit/agentplatform/genai/test_genai_model_garden.py

@@ -2120,6 +2120,7 @@ def test_deploy_publisher_model_default_config_no_optional_blocks(mock_client):
     assert kwargs["deploy_config"].dedicated_resources is None
     assert kwargs["model_config_val"].container_spec is None
     assert kwargs["endpoint_config"].dedicated_endpoint_enabled is None
+    assert kwargs["endpoint_config"].dedicated_endpoint_disabled is None
     assert kwargs["endpoint_config"].private_service_connect_config is None
 
 
@@ -2215,10 +2216,38 @@ def test_deploy_publisher_model_accelerator_only_triggers_dedicated_resources(
     assert dr.machine_spec.accelerator_type == "NVIDIA_L4"
 
 
-def test_deploy_publisher_model_dedicated_endpoint_disabled_inverts(
+def test_deploy_publisher_model_spot_without_machine_shape_forwarded(
     mock_client,
 ):
-  """dedicated_endpoint_disabled=True maps to dedicated_endpoint_enabled=False on wire."""
+  """spot alone is sent without a machine spec; the backend applies it to the default machine."""
+  with (
+      _patch_deploy_rpc(mock_client) as m_rpc,
+      mock.patch.object(
+          model_garden._operations_utils,
+          "await_operation",
+          return_value=_done_deploy_operation(),
+      ),
+  ):
+    mock_client.deploy_publisher_model(
+        model=_OPEN_MODEL,
+        config=types.DeployPublisherModelConfig(spot=True),
+    )
+    dr = m_rpc.call_args.kwargs["deploy_config"].dedicated_resources
+    assert dr is not None
+    assert dr.spot is True
+    assert dr.machine_spec is None
+    assert dr.min_replica_count is None
+    assert dr.max_replica_count is None
+
+
+def test_deploy_publisher_model_dedicated_endpoint_disabled_forwarded(
+    mock_client,
+):
+  """dedicated_endpoint_disabled=True is sent as-is, not as the deprecated field.
+
+  dedicated_endpoint_enabled=False is the proto3 default, so the backend
+  cannot tell it from unset and would keep the dedicated endpoint on.
+  """
   with (
       _patch_deploy_rpc(mock_client) as m_rpc,
       mock.patch.object(
@@ -2233,16 +2262,15 @@ def test_deploy_publisher_model_dedicated_endpoint_disabled_inverts(
             dedicated_endpoint_disabled=True
         ),
     )
-    assert (
-        m_rpc.call_args.kwargs["endpoint_config"].dedicated_endpoint_enabled
-        is False
-    )
+    endpoint_config = m_rpc.call_args.kwargs["endpoint_config"]
+    assert endpoint_config.dedicated_endpoint_disabled is True
+    assert endpoint_config.dedicated_endpoint_enabled is None
 
 
-def test_deploy_publisher_model_dedicated_endpoint_disabled_false_maps_to_true(
+def test_deploy_publisher_model_dedicated_endpoint_disabled_false_not_sent(
     mock_client,
 ):
-  """dedicated_endpoint_disabled=False maps to dedicated_endpoint_enabled=True."""
+  """dedicated_endpoint_disabled=False sends neither field (server default)."""
   with (
       _patch_deploy_rpc(mock_client) as m_rpc,
       mock.patch.object(
@@ -2257,10 +2285,9 @@ def test_deploy_publisher_model_dedicated_endpoint_disabled_false_maps_to_true(
             dedicated_endpoint_disabled=False
         ),
     )
-    assert (
-        m_rpc.call_args.kwargs["endpoint_config"].dedicated_endpoint_enabled
-        is True
-    )
+    endpoint_config = m_rpc.call_args.kwargs["endpoint_config"]
+    assert endpoint_config.dedicated_endpoint_disabled is None
+    assert endpoint_config.dedicated_endpoint_enabled is None
 
 
 def test_deploy_publisher_model_private_service_connect(mock_client):