Fix issue with bad path mount in container when user supplies code

Author: siddvenk

src/sagemaker/djl_inference/model.py

@@ -10,7 +10,7 @@
 # distributed on an "AS IS" BASIS, WITHOUT WARRANTIES OR CONDITIONS OF
 # ANY KIND, either express or implied. See the License for the specific
 # language governing permissions and limitations under the License.
-
+"""Placeholder docstring"""
 from __future__ import absolute_import
 
 import json
@@ -19,8 +19,8 @@
 from enum import Enum
 from typing import Optional, Union, Dict
 
-from sagemaker.djl_inference import defaults
 import sagemaker
+from sagemaker.djl_inference import defaults
 from sagemaker import s3, Predictor, image_uris, fw_utils
 from sagemaker.deserializers import JSONDeserializer
 from sagemaker.model import FrameworkModel
@@ -32,8 +32,8 @@
 
 logger = logging.getLogger("sagemaker")
 
-# DJL Serving uses log4j, so we convert python logging name to log4j equivalent
-LOG_LEVEL_MAP = {
+# DJL Serving uses log4j, so we convert python logging level to log4j equivalent
+_LOG_LEVEL_MAP = {
     logging.INFO: "info",
     logging.DEBUG: "debug",
     logging.WARNING: "warn",
@@ -107,12 +107,9 @@ def __init__(
         predictor_cls: callable = DJLLargeModelPredictor,
         **kwargs,
     ):
-        if kwargs.get("model_data") is not None:
-            raise ValueError(
-                "DJLLargeModels do not support the model_data parameter. Please use"
-                "uncompressed_model_data and ensure the s3 uri points to a folder containing"
-                "all model artifacts, not a tar.gz file"
-            )
+        if kwargs.get("model_data"):
+            logger.warning("DJLLargeModels do not use model_data parameter. You only need to specify the"
+                           "uncompressed_model_data parameter.")
         super(DJLLargeModel, self).__init__(
             None, image_uri, role, entry_point, predictor_cls=predictor_cls, **kwargs
         )
@@ -189,7 +186,7 @@ def deploy(
         if instance_type is None and not self.inference_recommender_job_results:
             raise ValueError(
                 f"instance_type must be specified, or inference recommendation from right_size()"
-                "must be run to deploy the model. Supported instance type families are :"
+                "must be run to deploy the model. Supported instance type families are:"
                 f"{defaults.ALLOWED_INSTANCE_FAMILIES}"
             )
         if instance_type:
@@ -225,11 +222,6 @@ def prepare_container_def(
         accelerator_type=None,
         serverless_inference_config=None,
     ):
-        if serverless_inference_config is not None:
-            raise ValueError("DJLLargeModel does not support serverless deployment")
-        if accelerator_type is not None:
-            raise ValueError("DJLLargeModel does not support Elastic Inference accelerator")
-
         if not self.image_uri:
             region_name = self.sagemaker_session.boto_session.region_name
             self.image_uri = self.serving_image_uri(region_name)
@@ -243,8 +235,8 @@ def prepare_container_def(
             else self.sagemaker_session.settings.local_download_dir
         )
         with _tmpdir(directory=local_download_dir) as tmp:
-            logger.info(f"Using tmp dir {tmp}")
             if self.source_dir or self.entry_point:
+                # Below method downloads from s3, or moves local files to tmp/code
                 _create_or_update_code_dir(
                     tmp,
                     self.entry_point,
@@ -253,11 +245,14 @@ def prepare_container_def(
                     self.sagemaker_session,
                     tmp,
                 )
-            existing_serving_properties = _read_existing_serving_properties(tmp)
+            tmp_code_dir = os.path.join(tmp, "code")
+            existing_serving_properties = _read_existing_serving_properties(tmp_code_dir)
             kwargs_serving_properties = self.generate_serving_properties()
             existing_serving_properties.update(kwargs_serving_properties)
 
-            with open(os.path.join(tmp, "serving.properties"), "w+") as f:
+            if not os.path.exists(tmp_code_dir):
+                os.mkdir(tmp_code_dir)
+            with open(os.path.join(tmp_code_dir, "serving.properties"), "w+") as f:
                 for key, val in existing_serving_properties.items():
                     f.write(f"{key}={val}\n")
 
@@ -270,7 +265,7 @@ def prepare_container_def(
                 bucket,
                 deploy_key_prefix,
                 self.entry_point,
-                directory=tmp,
+                directory=tmp_code_dir,
                 dependencies=self.dependencies,
                 kms_key=self.model_kms_key,
             )
@@ -318,11 +313,11 @@ def _get_container_env(self):
         if not self.container_log_level:
             return self.env
 
-        if self.container_log_level not in LOG_LEVEL_MAP:
+        if self.container_log_level not in _LOG_LEVEL_MAP:
             logger.warning(f"Ignoring invalid container log level: {self.container_log_level}")
             return self.env
 
-        self.env["SERVING_OPTS"] = f'"-Dai.djl.logging.level={LOG_LEVEL_MAP[self.container_log_level]}"'
+        self.env["SERVING_OPTS"] = f'"-Dai.djl.logging.level={_LOG_LEVEL_MAP[self.container_log_level]}"'
         return self.env