Add possiblity to accomodate for training/inference version mismatch

Author: shantanutrip

src/sagemaker/tensorflow/model.py

@@ -185,6 +185,14 @@ def __init__(
             )
         self.framework_version = framework_version
 
+        ## Inference framework version is being introduced to accomodate the mismatch between tensorflow and tensorflow serving releases,
+        ## wherein the TF and TFS might have different patch versions, but end up hosting the model of same TF version. For eg., the
+        ## upstream TFS-2.12.0 release was a bad release and hence a new TFS-2.12.1 release was made to host models from TF-2.12.0.
+        training_inference_version_mismatch_dict = {"2.12.0": "2.12.1"}
+        self.inference_framework_version = training_inference_version_mismatch_dict.get(
+            framework_version, framework_version
+        )
+
         super(TensorFlowModel, self).__init__(
             model_data=model_data,
             role=role,
@@ -457,7 +465,7 @@ def _get_image_uri(
         return image_uris.retrieve(
             self._framework_name,
             region_name or self.sagemaker_session.boto_region_name,
-            version=self.framework_version,
+            version=self.inference_framework_version,
             instance_type=instance_type,
             accelerator_type=accelerator_type,
             image_scope="inference",