fix: Register TensorFlow models from Ray checkpoints for more recent TensorFlow version, addressing the deprecation of SavedModel format in keras 3

PiperOrigin-RevId: 628562509

Author: yinghsienwu

google/cloud/aiplatform/preview/vertex_ray/predict/tensorflow/register.py

@@ -24,6 +24,7 @@
 from google.cloud import aiplatform
 from google.cloud.aiplatform import initializer
 from google.cloud.aiplatform import utils
+from google.cloud.aiplatform.preview.vertex_ray.predict.util import constants
 from google.cloud.aiplatform.preview.vertex_ray.predict.util import (
     predict_utils,
 )
@@ -44,6 +45,7 @@ def register_tensorflow(
     artifact_uri: Optional[str] = None,
     _model: Optional[Union["tf.keras.Model", Callable[[], "tf.keras.Model"]]] = None,
     display_name: Optional[str] = None,
+    tensorflow_version: Optional[str] = None,
     **kwargs,
 ) -> aiplatform.Model:
     """Uploads a Ray Tensorflow Checkpoint as Tensorflow Model to Model Registry.
@@ -79,6 +81,11 @@ def create_model():
         display_name (str):
             Optional. The display name of the Model. The name can be up to 128
             characters long and can be consist of any UTF-8 characters.
+        tensorflow_version (str):
+                Optional. The version of the Tensorflow serving container.
+                Supported versions:
+                https://cloud.google.com/vertex-ai/docs/predictions/pre-built-containers
+                If the version is not specified, the latest version is used.
         **kwargs:
             Any kwargs will be passed to aiplatform.Model registration.
 
@@ -89,6 +96,9 @@ def create_model():
     Raises:
         ValueError: Invalid Argument.
     """
+
+    if tensorflow_version is None:
+        tensorflow_version = constants._TENSORFLOW_VERSION
     artifact_uri = artifact_uri or initializer.global_config.staging_bucket
     predict_utils.validate_artifact_uri(artifact_uri)
     prefix = "ray-on-vertex-registered-tensorflow-model"
@@ -99,10 +109,16 @@ def create_model():
     )
     tf_model = _get_tensorflow_model_from(checkpoint, model=_model)
     model_dir = os.path.join(artifact_uri, prefix)
-    tf_model.save(model_dir)
+    try:
+        import tensorflow as tf
+
+        tf.saved_model.save(tf_model, model_dir)
+    except ImportError:
+        logging.warning("TensorFlow must be installed to save the trained model.")
     return aiplatform.Model.upload_tensorflow_saved_model(
         saved_model_dir=model_dir,
         display_name=display_model_name,
+        tensorflow_version=tensorflow_version,
         **kwargs,
     )
 
@@ -139,13 +155,13 @@ def _get_tensorflow_model_from(
 
         return checkpoint.get_model(model)
 
-    # get_model() signature changed in future versions
     try:
-        from tensorflow import keras
+        import tensorflow as tf
 
         try:
-            return keras.models.load_model(checkpoint.path)
+            return tf.saved_model.load(checkpoint.path)
         except OSError:
-            return keras.models.load_model("gs://" + checkpoint.path)
+            return tf.saved_model.load("gs://" + checkpoint.path)
+
     except ImportError:
         logging.warning("TensorFlow must be installed to load the trained model.")

google/cloud/aiplatform/preview/vertex_ray/predict/util/constants.py

@@ -21,3 +21,8 @@
 _PICKLE_EXTENTION = ".pkl"
 
 _XGBOOST_VERSION = "1.6"
+# TensorFlow 2.13 requires typing_extensions<4.6 and will cause errors in Ray.
+# https://github.com/tensorflow/tensorflow/blob/v2.13.0/tensorflow/tools/pip_package/setup.py#L100
+# 2.13 is the latest supported version of Vertex prebuilt prediction container.
+# Set 2.12 as default here since 2.13 cause errors.
+_TENSORFLOW_VERSION = "2.12"

google/cloud/aiplatform/preview/vertex_ray/predict/xgboost/register.py

@@ -77,8 +77,9 @@ def register_xgboost(
             Optional. The display name of the Model. The name can be up to 128
             characters long and can be consist of any UTF-8 characters.
         xgboost_version (str): Optional. The version of the XGBoost serving container.
-                Supported versions: ["0.82", "0.90", "1.1", "1.2", "1.3", "1.4", "1.6", "1.7"].
-                If the version is not specified, the latest version is used.
+                Supported versions:
+                https://cloud.google.com/vertex-ai/docs/predictions/pre-built-containers
+                If the version is not specified, the version 1.6 is used.
         **kwargs:
             Any kwargs will be passed to aiplatform.Model registration.