feat: LLM - Enable tuning eval TensorBoard without evaluation data

PiperOrigin-RevId: 572460871

Author: Ark-kun

tests/unit/aiplatform/test_language_models.py

@@ -1638,6 +1638,56 @@ def test_tune_text_generation_model_ga(
                 == test_constants.EndpointConstants._TEST_ENDPOINT_NAME
             )
 
+    @pytest.mark.parametrize(
+        "job_spec",
+        [_TEST_PIPELINE_SPEC_JSON],
+    )
+    @pytest.mark.parametrize(
+        "mock_request_urlopen",
+        ["https://us-central1-kfp.pkg.dev/proj/repo/pack/latest"],
+        indirect=True,
+    )
+    def test_tune_text_generation_model_evaluation_with_only_tensorboard(
+        self,
+        mock_pipeline_service_create,
+        mock_pipeline_job_get,
+        mock_pipeline_bucket_exists,
+        job_spec,
+        mock_load_yaml_and_json,
+        mock_gcs_from_string,
+        mock_gcs_upload,
+        mock_request_urlopen,
+        mock_get_tuned_model,
+    ):
+        """Tests tuning the text generation model."""
+        with mock.patch.object(
+            target=model_garden_service_client.ModelGardenServiceClient,
+            attribute="get_publisher_model",
+            return_value=gca_publisher_model.PublisherModel(
+                _TEXT_BISON_PUBLISHER_MODEL_DICT
+            ),
+        ):
+            model = language_models.TextGenerationModel.from_pretrained(
+                "text-bison@001"
+            )
+
+            tuning_job_location = "europe-west4"
+            tensorboard_name = f"projects/{_TEST_PROJECT}/locations/{tuning_job_location}/tensorboards/123"
+
+            model.tune_model(
+                training_data=_TEST_TEXT_BISON_TRAINING_DF,
+                tuning_job_location=tuning_job_location,
+                tuned_model_location="us-central1",
+                tuning_evaluation_spec=preview_language_models.TuningEvaluationSpec(
+                    tensorboard=tensorboard_name,
+                ),
+            )
+            call_kwargs = mock_pipeline_service_create.call_args[1]
+            pipeline_arguments = call_kwargs[
+                "pipeline_job"
+            ].runtime_config.parameter_values
+            assert pipeline_arguments["tensorboard_resource_id"] == tensorboard_name
+
     @pytest.mark.parametrize(
         "job_spec",
         [_TEST_PIPELINE_SPEC_JSON],

vertexai/language_models/_language_models.py

@@ -214,13 +214,14 @@ def tune_model(
             tuning_parameters["learning_rate_multiplier"] = learning_rate_multiplier
         eval_spec = tuning_evaluation_spec
         if eval_spec is not None:
-            if isinstance(eval_spec.evaluation_data, str):
-                if eval_spec.evaluation_data.startswith("gs://"):
-                    tuning_parameters["evaluation_data_uri"] = eval_spec.evaluation_data
+            if eval_spec.evaluation_data:
+                if isinstance(eval_spec.evaluation_data, str):
+                    if eval_spec.evaluation_data.startswith("gs://"):
+                        tuning_parameters["evaluation_data_uri"] = eval_spec.evaluation_data
+                    else:
+                        raise ValueError("evaluation_data should be a GCS URI")
                 else:
-                    raise ValueError("evaluation_data should be a GCS URI")
-            else:
-                raise TypeError("evaluation_data should be a URI string")
+                    raise TypeError("evaluation_data should be a URI string")
             if eval_spec.evaluation_interval is not None:
                 tuning_parameters["evaluation_interval"] = eval_spec.evaluation_interval
             if eval_spec.enable_early_stopping is not None:
@@ -648,7 +649,7 @@ class TuningEvaluationSpec:
 
     __module__ = "vertexai.language_models"
 
-    evaluation_data: str
+    evaluation_data: Optional[str] = None
     evaluation_interval: Optional[int] = None
     enable_early_stopping: Optional[bool] = None
     tensorboard: Optional[Union[aiplatform.Tensorboard, str]] = None