[Feat] Support MLflow Model Format Through ModelBuilder (aws#4564)

* Initial commit for mlflow support in ModelBuilder

Fix some formating issues

Fix style issues and typo

Fix s3 downloading for mlflow

Fix logging in inference script for torchserve

Add validation for mlflow

fix unused param in s3 downloader

* Add mlflow tensorflow flavor load in inference script

[Fix] enhance error handling for mlflow model support

* Add unit tests for utils

fix black format

* mlflow integration integ tests initial commit

Fix schema builder and constant import

Fix typo

Add integ tests for mlflow xgboost flavor

fix style issues for integ tests

adding ut

fix import

fix import

add missed mock

add mock for path.isfile

add mock for open

try another way of patching

test patching

fix patching

fix patching

fix patching

fix patching

fix patching

fix patching

fix patching

fix patching

add ut for s3 input

add negative test case

fix indent

increase test coverage

fix patching module

fix session patch

fix test input

debug

remov debug messages

fix ut

prepare for initial pr

fix doc8

fix ut name collision

fix integ tests naming

fix ut run in PR

Add ignore path to doc8

* Move mlflow inputs under model_metadata instead

fix ut failure

fix assertion

* Fix local path input for mlflow

fix ut

fix ut

normalize path befor copying over

fix for ut

fix black format

* mark local container integ test as flaky as they all bind to the same host

* marg mlflow xgboost integ local container test as flaky

* skip copying files if src and dst are the same

* resolve pr comments

* keep resolving pr comments

* update docstrings so that mlflow support is mark as beta

* fix pylint failures

---------

Co-authored-by: Jacky Lee <[email protected]>

Author: 2 people

src/sagemaker/serve/builder/model_builder.py

@@ -20,6 +20,8 @@
 
 from pathlib import Path
 
+from sagemaker.s3 import S3Downloader
+
 from sagemaker import Session
 from sagemaker.model import Model
 from sagemaker.base_predictor import PredictorBase
@@ -37,6 +39,22 @@
 from sagemaker.serve.builder.jumpstart_builder import JumpStart
 from sagemaker.serve.builder.transformers_builder import Transformers
 from sagemaker.predictor import Predictor
+from sagemaker.serve.model_format.mlflow.constants import (
+    MLFLOW_MODEL_PATH,
+    MLFLOW_METADATA_FILE,
+    MLFLOW_PIP_DEPENDENCY_FILE,
+)
+from sagemaker.serve.model_format.mlflow.utils import (
+    _get_default_model_server_for_mlflow,
+    _mlflow_input_is_local_path,
+    _download_s3_artifacts,
+    _select_container_for_mlflow_model,
+    _generate_mlflow_artifact_path,
+    _get_all_flavor_metadata,
+    _get_deployment_flavor,
+    _validate_input_for_mlflow,
+    _copy_directory_contents,
+)
 from sagemaker.serve.save_retrive.version_1_0_0.metadata.metadata import Metadata
 from sagemaker.serve.spec.inference_spec import InferenceSpec
 from sagemaker.serve.utils import task
@@ -145,8 +163,11 @@ class ModelBuilder(Triton, DJL, JumpStart, TGI, Transformers):
             to the model server). Possible values for this argument are
             ``TORCHSERVE``, ``MMS``, ``TENSORFLOW_SERVING``, ``DJL_SERVING``,
             ``TRITON``, and``TGI``.
-        model_metadata (Optional[Dict[str, Any]): Dictionary used to override the HuggingFace
-            model metadata. Currently ``HF_TASK`` is overridable.
+        model_metadata (Optional[Dict[str, Any]): Dictionary used to override model metadata.
+            Currently, ``HF_TASK`` is overridable for HuggingFace model. ``MLFLOW_MODEL_PATH``
+            is available for providing local path or s3 path to MLflow artifacts. However,
+            ``MLFLOW_MODEL_PATH`` is experimental and is not intended for production use
+            at this moment.
     """
 
     model_path: Optional[str] = field(
@@ -245,7 +266,10 @@ class ModelBuilder(Triton, DJL, JumpStart, TGI, Transformers):
     )
     model_metadata: Optional[Dict[str, Any]] = field(
         default=None,
-        metadata={"help": "Define the model metadata to override, currently supports `HF_TASK`"},
+        metadata={
+            "help": "Define the model metadata to override, currently supports `HF_TASK`, "
+            "`MLFLOW_MODEL_PATH`"
+        },
     )
 
     def _build_validations(self):
@@ -297,6 +321,8 @@ def _save_model_inference_spec(self):
                 save_pkl(code_path, (self._framework, self.schema_builder))
             else:
                 save_pkl(code_path, (self.model, self.schema_builder))
+        elif self._is_mlflow_model:
+            save_pkl(code_path, self.schema_builder)
         else:
             raise ValueError("Cannot detect required model or inference spec")
 
@@ -577,6 +603,76 @@ def wrapper(*args, **kwargs):
 
         return wrapper
 
+    def _check_if_input_is_mlflow_model(self) -> bool:
+        """Checks whether an MLmodel file exists in the given directory.
+
+        Returns:
+            bool: True if the MLmodel file exists, False otherwise.
+        """
+        if self.inference_spec or self.model:
+            logger.info(
+                "Either inference spec or model is provided. "
+                "ModelBuilder is not handling MLflow model input"
+            )
+            return False
+
+        if not self.model_metadata:
+            logger.info(
+                "No ModelMetadata provided. ModelBuilder is not handling MLflow model input"
+            )
+            return False
+
+        path = self.model_metadata.get(MLFLOW_MODEL_PATH)
+        if not path:
+            logger.info(
+                "%s is not provided in ModelMetadata. ModelBuilder is not handling MLflow model "
+                "input",
+                MLFLOW_MODEL_PATH,
+            )
+            return False
+
+        # Check for S3 path
+        if path.startswith("s3://"):
+            s3_downloader = S3Downloader()
+            if not path.endswith("/"):
+                path += "/"
+            s3_uri_to_mlmodel_file = f"{path}{MLFLOW_METADATA_FILE}"
+            response = s3_downloader.list(s3_uri_to_mlmodel_file, self.sagemaker_session)
+            return len(response) > 0
+
+        file_path = os.path.join(path, MLFLOW_METADATA_FILE)
+        return os.path.isfile(file_path)
+
+    def _initialize_for_mlflow(self) -> None:
+        """Initialize mlflow model artifacts, image uri and model server."""
+        mlflow_path = self.model_metadata.get(MLFLOW_MODEL_PATH)
+        if not _mlflow_input_is_local_path(mlflow_path):
+            # TODO: extend to package arn, run id and etc.
+            _download_s3_artifacts(mlflow_path, self.model_path, self.sagemaker_session)
+        else:
+            _copy_directory_contents(mlflow_path, self.model_path)
+        mlflow_model_metadata_path = _generate_mlflow_artifact_path(
+            self.model_path, MLFLOW_METADATA_FILE
+        )
+        # TODO: add validation on MLmodel file
+        mlflow_model_dependency_path = _generate_mlflow_artifact_path(
+            self.model_path, MLFLOW_PIP_DEPENDENCY_FILE
+        )
+        flavor_metadata = _get_all_flavor_metadata(mlflow_model_metadata_path)
+        deployment_flavor = _get_deployment_flavor(flavor_metadata)
+
+        self.model_server = self.model_server or _get_default_model_server_for_mlflow(
+            deployment_flavor
+        )
+        self.image_uri = self.image_uri or _select_container_for_mlflow_model(
+            mlflow_model_src_path=self.model_path,
+            deployment_flavor=deployment_flavor,
+            region=self.sagemaker_session.boto_region_name,
+            instance_type=self.instance_type,
+        )
+        self.env_vars.update({"MLFLOW_MODEL_FLAVOR": f"{deployment_flavor}"})
+        self.dependencies.update({"requirements": mlflow_model_dependency_path})
+
     # Model Builder is a class to build the model for deployment.
     # It supports two modes of deployment
     # 1/ SageMaker Endpoint
@@ -620,6 +716,14 @@ def build(  # pylint: disable=R0911
         self.serve_settings = self._get_serve_setting()
 
         self._is_custom_image_uri = self.image_uri is not None
+        self._is_mlflow_model = self._check_if_input_is_mlflow_model()
+        if self._is_mlflow_model:
+            logger.warning(
+                "Support of MLflow format models is experimental and is not intended"
+                " for production at this moment."
+            )
+            self._initialize_for_mlflow()
+            _validate_input_for_mlflow(self.model_server)
 
         if isinstance(self.model, str):
             model_task = None

src/sagemaker/serve/model_format/mlflow/constants.py

@@ -0,0 +1,41 @@
+# Copyright Amazon.com, Inc. or its affiliates. All Rights Reserved.
+#
+# Licensed under the Apache License, Version 2.0 (the "License"). You
+# may not use this file except in compliance with the License. A copy of
+# the License is located at
+#
+#     http://aws.amazon.com/apache2.0/
+#
+# or in the "license" file accompanying this file. This file is
+# 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.
+"""Holds constants used for interpreting MLflow models."""
+from __future__ import absolute_import
+
+DEFAULT_FW_USED_FOR_DEFAULT_IMAGE = "pytorch"
+DEFAULT_PYTORCH_VERSION = {
+    "py38": "1.12.1",
+    "py39": "1.13.1",
+    "py310": "2.2.0",
+}
+MLFLOW_MODEL_PATH = "MLFLOW_MODEL_PATH"
+MLFLOW_METADATA_FILE = "MLmodel"
+MLFLOW_PIP_DEPENDENCY_FILE = "requirements.txt"
+MLFLOW_PYFUNC = "python_function"
+MLFLOW_FLAVOR_TO_PYTHON_PACKAGE_MAP = {
+    "sklearn": "scikit-learn",
+    "pytorch": "torch",
+    "tensorflow": "tensorflow",
+    "keras": "tensorflow",
+    "xgboost": "xgboost",
+    "lightgbm": "lightgbm",
+    "h2o": "h2o",
+    "spark": "pyspark",
+    "onnx": "onnxruntime",
+}
+FLAVORS_WITH_FRAMEWORK_SPECIFIC_DLC_SUPPORT = [  # will extend to keras and tf
+    "sklearn",
+    "pytorch",
+    "xgboost",
+]