feat: Add UnifiedMetric support to Vertex Tuning evaluation config

google/genai/_transformers.py

@@ -1295,6 +1295,19 @@ def t_metrics(
     metrics_payload = []
 
     for metric in metrics:
+
+      if isinstance(metric, dict) and types.UnifiedMetric.model_validate(metric):
+        metric = types.UnifiedMetric.model_validate(metric)
+
+      if isinstance(metric, types.UnifiedMetric):
+        unified_metric_payload: dict[str, Any] = metric.model_dump()
+        unified_metric_payload['aggregation_metrics'] = [
+            'AVERAGE',
+            'STANDARD_DEVIATION',
+        ]
+        metrics_payload.append(unified_metric_payload)
+        continue
+
       metric_payload_item: dict[str, Any] = {}
       metric_payload_item['aggregation_metrics'] = [
           'AVERAGE',

google/genai/tests/tunings/test_tune.py

@@ -344,6 +344,56 @@ def test_eval_config_with_metrics(client):
     assert tuning_job.state == genai_types.JobState.JOB_STATE_PENDING
 
 
+@pytest.mark.skipif(
+    "config.getoption('--private')",
+    reason="Skipping in pre-public tests"
+)
+def test_eval_config_with_unified_metrics(client):
+  """Tests tuning with eval config metrics."""
+  if client._api_client.vertexai:
+    evaluation_config=genai_types.EvaluationConfig(
+        metrics=[
+            genai_types.Metric(
+                name="prompt-relevance",
+                prompt_template="How well does the response address the prompt?: PROMPT: {request}\n RESPONSE: {response}\n",
+                return_raw_output=True,
+                judge_model_system_instruction="You are a cat. Make all evaluations from this perspective.",
+            ),
+            genai_types.Metric(name='bleu'),
+            genai_types.Metric(name='rouge_1'),
+            genai_types.UnifiedMetric(
+                bleu_spec=genai_types.BleuSpec(
+                    use_effective_order=True
+                )
+            ),
+        ],
+        output_config=genai_types.OutputConfig(
+            gcs_destination=genai_types.GcsDestination(
+                output_uri_prefix="gs://sararob_test/"
+            )
+        ),
+        autorater_config=genai_types.AutoraterConfig(
+            sampling_count=1,
+            autorater_model="test-model",
+        ),
+    )
+    tuning_job = client.tunings.tune(
+      base_model="gemini-2.5-flash",
+      training_dataset=genai_types.TuningDataset(gcs_uri="gs://cloud-samples-data/ai-platform/generative_ai/gemini-2_0/text/sft_train_data.jsonl"),
+      config=genai_types.CreateTuningJobConfig(
+          tuned_model_display_name="tuning job with eval config",
+          epoch_count=1,
+          learning_rate_multiplier=1.0,
+          adapter_size="ADAPTER_SIZE_ONE",
+          validation_dataset=genai_types.TuningValidationDataset(
+              gcs_uri="gs://cloud-samples-data/ai-platform/generative_ai/gemini-2_0/text/sft_validation_data.jsonl"
+          ),
+          evaluation_config=evaluation_config,
+      ),
+    )
+    assert tuning_job.state == genai_types.JobState.JOB_STATE_PENDING
+
+
 @pytest.mark.skipif(
     "config.getoption('--private')",
     reason="Skipping in pre-public tests"