feat: Add ThinkingConfig support to GenerationConfig

Expose thinking_level, thinking_budget, and include_thoughts in the
high-level vertexai.generative_models.GenerationConfig wrapper by
aliasing the v1 GAPIC ThinkingConfig type and converting to v1beta1
via binary serialization to preserve thinking_level through the
v1beta1 → v1 round-trip in GA GenerativeModel.

Co-Authored-By: Claude Opus 4.6 (1M context) <noreply@anthropic.com>

Author: caizifen

tests/unit/vertexai/test_generative_models.py

@@ -1537,6 +1537,73 @@ def test_generation_config_response_schema_dict_renaming(self):
         )
         assert config.to_dict()["response_schema"] == _RENAMING_EXPECTED_SCHEMA
 
+    def test_generation_config_with_thinking_budget(self):
+        config = generative_models.GenerationConfig(
+            thinking_config=generative_models.GenerationConfig.ThinkingConfig(
+                thinking_budget=1024,
+            ),
+        )
+        config_dict = config.to_dict()
+        assert config_dict["thinking_config"]["thinking_budget"] == 1024
+
+    def test_generation_config_with_include_thoughts(self):
+        config = generative_models.GenerationConfig(
+            thinking_config=generative_models.GenerationConfig.ThinkingConfig(
+                include_thoughts=True,
+            ),
+        )
+        config_dict = config.to_dict()
+        assert config_dict["thinking_config"]["include_thoughts"] is True
+
+    def test_generation_config_with_thinking_level(self):
+        ThinkingConfig = generative_models.GenerationConfig.ThinkingConfig
+        config = generative_models.GenerationConfig(
+            thinking_config=ThinkingConfig(
+                thinking_level=ThinkingConfig.ThinkingLevel.HIGH,
+            ),
+        )
+        # thinking_level is a v1-only field; verify it survives the
+        # v1 → v1beta1 → v1 binary serialization round-trip.
+        raw = config._raw_generation_config
+        serialized = type(raw).serialize(raw)
+        deserialized_v1 = types_v1.GenerationConfig.deserialize(serialized)
+        assert deserialized_v1.thinking_config.thinking_level == (
+            types_v1.GenerationConfig.ThinkingConfig.ThinkingLevel.HIGH
+        )
+
+    def test_generation_config_with_thinking_config_combined(self):
+        ThinkingConfig = generative_models.GenerationConfig.ThinkingConfig
+        config = generative_models.GenerationConfig(
+            thinking_config=ThinkingConfig(
+                thinking_budget=4096,
+                include_thoughts=True,
+                thinking_level=ThinkingConfig.ThinkingLevel.MEDIUM,
+            ),
+        )
+        config_dict = config.to_dict()
+        assert config_dict["thinking_config"]["thinking_budget"] == 4096
+        assert config_dict["thinking_config"]["include_thoughts"] is True
+        # Verify thinking_level survives binary round-trip
+        raw = config._raw_generation_config
+        serialized = type(raw).serialize(raw)
+        deserialized_v1 = types_v1.GenerationConfig.deserialize(serialized)
+        assert deserialized_v1.thinking_config.thinking_level == (
+            types_v1.GenerationConfig.ThinkingConfig.ThinkingLevel.MEDIUM
+        )
+
+    def test_generation_config_thinking_config_from_dict(self):
+        config = generative_models.GenerationConfig.from_dict(
+            {
+                "thinking_config": {
+                    "thinking_budget": 2048,
+                    "include_thoughts": True,
+                },
+            }
+        )
+        config_dict = config.to_dict()
+        assert config_dict["thinking_config"]["thinking_budget"] == 2048
+        assert config_dict["thinking_config"]["include_thoughts"] is True
+
     def test_tool_schema_dict_renaming(self):
         # The `Tool` constructor does not take a dict so we don't test it here.
         tool = generative_models.Tool.from_dict(

vertexai/generative_models/_generative_models.py

@@ -1764,6 +1764,7 @@ class GenerationConfig:
 
     Modality = gapic_content_types.GenerationConfig.Modality
     ModelConfig = gapic_content_types.GenerationConfig.ModelConfig
+    ThinkingConfig = types_v1.GenerationConfig.ThinkingConfig
 
     def __init__(
         self,
@@ -1785,6 +1786,7 @@ def __init__(
         response_logprobs: Optional[bool] = None,
         response_modalities: Optional[List["GenerationConfig.Modality"]] = None,
         model_config: Optional["GenerationConfig.ModelConfig"] = None,
+        thinking_config: Optional["GenerationConfig.ThinkingConfig"] = None,
     ):
         r"""Constructs a GenerationConfig object.
 
@@ -1817,6 +1819,11 @@ def __init__(
             logprobs: Logit probabilities.
             reponse_logprobs: If true, export the logprobs results in response.
             model_config: Sets cost vs quality preference for model routing requests.
+            thinking_config: Configuration for thinking features
+                (thinking_level, thinking_budget, include_thoughts). Use
+                ``GenerationConfig.ThinkingConfig`` to construct. Note:
+                ``thinking_level`` is not preserved by ``to_dict()``/``from_dict()``
+                due to v1beta1 proto limitations.
 
         Usage:
 
@@ -1864,6 +1871,15 @@ def __init__(
             self._raw_generation_config.routing_config = (
                 routing_config._gapic_routing_config
             )
+        if thinking_config is not None:
+            # Convert v1 ThinkingConfig to v1beta1 via binary serialization.
+            # This preserves thinking_level (field 4) as an unknown field in
+            # v1beta1, which survives the v1beta1 → v1 conversion in GA model.
+            self._raw_generation_config.thinking_config = (
+                gapic_content_types.GenerationConfig.ThinkingConfig.deserialize(
+                    type(thinking_config).serialize(thinking_config)
+                )
+            )
 
     @classmethod
     def _from_gapic(