feat: LLM - Text Embedding - Added validation for text embedding tuning parameters.

PiperOrigin-RevId: 632301450

Author: vertex-sdk-bot

tests/unit/aiplatform/test_language_models.py

@@ -2407,6 +2407,58 @@ def test_tune_text_embedding_model(
                 == test_constants.EndpointConstants._TEST_ENDPOINT_NAME
             )
 
+    @pytest.mark.parametrize(
+        "optional_tune_args,error_regex",
+        [
+            (
+                dict(test_data="/tmp/bucket/test.tsv"),
+                "Each tuning dataset file must be a Google Cloud Storage URI starting with 'gs://'.",
+            ),
+            (
+                dict(output_dimensionality=-1),
+                "output_dimensionality must be an integer between 1 and 768",
+            ),
+            (
+                dict(learning_rate_multiplier=0),
+                "learning_rate_multiplier must be greater than 0",
+            ),
+            (
+                dict(train_steps=29),
+                "train_steps must be greater than or equal to 30",
+            ),
+            (
+                dict(batch_size=2048),
+                "batch_size must be between 1 and 1024",
+            ),
+        ],
+    )
+    def test_tune_text_embedding_model_invalid_values(
+        self, optional_tune_args, error_regex
+    ):
+        """Tests that certain embedding tuning values fail validation."""
+        aiplatform.init(
+            project=_TEST_PROJECT,
+            location=_TEST_LOCATION,
+            encryption_spec_key_name=_TEST_ENCRYPTION_KEY_NAME,
+        )
+        with mock.patch.object(
+            target=model_garden_service_client.ModelGardenServiceClient,
+            attribute="get_publisher_model",
+            return_value=gca_publisher_model.PublisherModel(
+                _TEXT_GECKO_PUBLISHER_MODEL_DICT
+            ),
+        ):
+            model = preview_language_models.TextEmbeddingModel.from_pretrained(
+                "text-multilingual-embedding-002"
+            )
+            with pytest.raises(ValueError, match=error_regex):
+                model.tune_model(
+                    training_data="gs://bucket/training.tsv",
+                    corpus_data="gs://bucket/corpus.jsonl",
+                    queries_data="gs://bucket/queries.jsonl",
+                    **optional_tune_args,
+                )
+
     @pytest.mark.parametrize(
         "job_spec",
         [_TEST_PIPELINE_SPEC_JSON, _TEST_PIPELINE_JOB],

vertexai/language_models/_language_models.py

@@ -2192,8 +2192,6 @@ async def get_embeddings_async(
 
 # TODO(b/625884109): Support Union[str, "pandas.core.frame.DataFrame"]
 # for corpus, queries, test and validation data.
-# TODO(b/625884109): Validate input args, batch_size >0 and train_steps >30, and
-# task_type must be 'DEFAULT' or None if _model_id is textembedding-gecko@001.
 class _PreviewTunableTextEmbeddingModelMixin(_TunableModelMixin):
     @classmethod
     def get_tuned_model(cls, *args, **kwargs):
@@ -2265,9 +2263,39 @@ def tune_model(
             Calling `job.result()` blocks until the tuning is complete and returns a `LanguageModel` object.
 
         Raises:
-            ValueError: If the "tuned_model_location" value is not supported
+            ValueError: If the provided parameter combinations or values are not
+                supported.
             RuntimeError: If the model does not support tuning
         """
+        if batch_size is not None and batch_size not in range(1, 1024):
+            raise ValueError(
+                f"batch_size must be between 1 and 1024. Given {batch_size}."
+            )
+        if train_steps is not None and train_steps < 30:
+            raise ValueError(
+                f"train_steps must be greater than or equal to 30. Given {train_steps}."
+            )
+        if learning_rate_multiplier is not None and learning_rate_multiplier <= 0:
+            raise ValueError(
+                f"learning_rate_multiplier must be greater than 0. Given {learning_rate_multiplier}."
+            )
+        if output_dimensionality is not None and output_dimensionality not in range(
+            1, 769
+        ):
+            raise ValueError(
+                f"output_dimensionality must be an integer between 1 and 768. Given {output_dimensionality}."
+            )
+        for dataset in [
+            training_data,
+            corpus_data,
+            queries_data,
+            test_data,
+            validation_data,
+        ]:
+            if dataset is not None and not dataset.startswith("gs://"):
+                raise ValueError(
+                    f"Each tuning dataset file must be a Google Cloud Storage URI starting with 'gs://'. Given {dataset}."
+                )
 
         return super().tune_model(
             training_data=training_data,