MLflow#

MLflow is an open source framework for tracking ML experiments, packaging ML code for training pipelines, and capturing models logged from experiments. It enables data scientists to iterate quickly during model development while keeping their experiments and training pipelines reproducible.

BentoML, on the other hand, focuses on ML in production. By design, BentoML is agnostic to the experimentation platform and the model development environment.

Comparing to the MLflow model registry, BentoMLā€™s model format and model store is designed for managing model artifacts that will be used for building, testing, and deploying prediction services. It is best fitted to manage your ā€œfinalized modelā€, sets of models that yield the best outcomes from your periodic training pipelines and are meant for running in production.

BentoML integrates with MLflow natively. Users can not only port over models logged with MLflow Tracking to BentoML for high-performance model serving but also combine MLFlow projects and pipelines with BentoMLā€™s model deployment workflow in an efficient manner.

Compatibility#

BentoML supports MLflow 0.9 and above.

Examples#

Besides this documentation, also check out code samples demonstrating BentoML and MLflow integration at: bentoml/examples: MLflow Examples.

Import an MLflow model#

MLflow Model is a format for saving trained model artifacts in MLflow experiments and pipelines. BentoML supports importing MLflow model to its own format for model serving. For example:

mlflow.sklearn.save_model(model, "./my_model")
bentoml.mlflow.import_model("my_sklearn_model", model_uri="./my_model")

with mlflow.start_run():
    mlflow.pytorch.log_model(model, artifact_path="pytorch-model")

    model_uri = mlflow.get_artifact_uri("pytorch-model")
    bento_model = bentoml.mlflow.import_model(
        'mlflow_pytorch_mnist',
        model_uri,
        signatures={'predict': {'batchable': True}}
    )

The bentoml.mlflow.import_model API is similar to the other save_model APIs found in BentoML, where the first argument represent the model name in BentoML model store. A new version will be automatically generated when a new MLflow model is imported. Users can manage imported MLflow models same as models saved with other ML frameworks:

bentoml models list mlflow_pytorch_mnist

The second argument model_uri takes a URI to the MLflow model. It can be a local path, a 'runs:/' URI, or a remote storage URI (e.g., an 's3://' URI). Here are some example model_uri values commonly used in MLflow:

/Users/me/path/to/local/model
../relative/path/to/local/model
s3://my_bucket/path/to/model
runs:/<mlflow_run_id>/run-relative/path/to/model
models:/<model_name>/<model_version>
models:/<model_name>/<stage>

Running Imported Model#

MLflow models imported to BentoML can be loaded back for running inference in a various of ways.

Loading original model flavor#

For evaluation and testing purpose, sometimes itā€™s convenient to load the model in its native form

bento_model = bentoml.mlflow.get("mlflow_pytorch_mnist:latest")
mlflow_model_path = bento_model.path_of(bentoml.mlflow.MLFLOW_MODEL_FOLDER)

loaded_pytorch_model = mlflow.pytorch.load_model(mlflow_model_path)
loaded_pytorch_model.to(device)
loaded_pytorch_model.eval()
with torch.no_grad():
    input_tensor = torch.from_numpy(test_input_arr).to(device)
    predictions = loaded_pytorch_model(input_tensor)

Loading Pyfunc flavor#

By default, bentoml.mflow.load_model will load the imported MLflow model using the python_function flavor for best compatibility across all ML frameworks supported by MLflow.

pyfunc_model: mlflow.pyfunc.PyFuncModel = bentoml.mlflow.load_model("mlflow_pytorch_mnist:latest")
predictions = pyfunc_model.predict(test_input_arr)

Using Model Runner#

Imported MLflow models can be loaded as BentoML Runner for best performance in building prediction service with BentoML. To test out the runner API:

runner = bentoml.mlflow.get("mlflow_pytorch_mnist:latest").to_runner()
runner.init_local()
runner.predict.run(input_df)

Learn more about BentoML Runner at Using Runners.

Runner created from an MLflow model supports the following input types. Note that for some ML frameworks, only a subset of this list is supported.

MLflowRunnerInput = Union[pandas.DataFrame, np.ndarray, List[Any], Dict[str, Any]]
MLflowRunnerOutput = Union[pandas.DataFrame, pandas.Series, np.ndarray, list]

Note

To use adaptive batching with a MLflow Runner, make sure to set signatures={'predict': {'batchable': True}} when importing the model:

bento_model = bentoml.mlflow.import_model(
    'mlflow_pytorch_mnist',
    model_uri,
    signatures={'predict': {'batchable': True}}
)

Optimizations#

There are two major limitations of using MLflow Runner in BentoML:

  • Lack of support for GPU

  • Lack of support for multiple inference method

A common optimization we recommend, is to save trained model instance directly with BentoML, instead of importing MLflow pyfunc model. This makes it possible to run GPU inference and expose multiple inference signatures.

  1. Save model directly with bentoml

mlflow.sklearn.log_model(clf, "model")
bentoml.sklearn.save_model("iris_clf", clf)

  1. Load original flavor and save with BentoML

loaded_model = mlflow.sklearn.load_model(model_uri)
bentoml.sklearn.save_model("iris_clf", loaded_model)

This way, it goes back to a typically BentoML workflow, which allow users to use a Runner specifically built for the target ML framework, with GPU support and multiple signatures available.

Build Prediction Service#

Hereā€™s an example bentoml.Service built with a MLflow model:

import bentoml
import mlflow
import torch

mnist_runner = bentoml.mlflow.get('mlflow_pytorch_mnist:latest').to_runner()

svc = bentoml.Service('mlflow_pytorch_mnist', runners=[ mnist_runner ])

input_spec = bentoml.io.NumpyNdarray(
    dtype="float32",
    shape=[-1, 1, 28, 28],
    enforce_shape=True,
    enforce_dtype=True,
)

@svc.api(input=input_spec, output=bentoml.io.NumpyNdarray())
def predict(input_arr):
    return mnist_runner.predict.run(input_arr)

To try out the full example, visit bentoml/examples: MLflow Pytorch Example.

MLflow šŸ¤ BentoML Workflow#

There are numerous ways you can integrate BentoML with your MLflow workflow for model serving and deployment.

  1. Find model_uri from a MLflow model instance returned from log_model:

# https://github.com/bentoml/BentoML/tree/main/examples/mlflow/sklearn_logistic_regression
logged_model = mlflow.sklearn.log_model(lr, "model")
print("Model saved in run %s" % mlflow.active_run().info.run_uuid)

# Import logged mlflow model to BentoML model store for serving:
bento_model = bentoml.mlflow.import_model('logistic_regression_model', logged_model.model_uri)
print("Model imported to BentoML: %s" % bento_model)

  1. Find model artifact path inside current mlflow.run scope:

# https://github.com/bentoml/BentoML/tree/main/examples/mlflow/pytorch
with mlflow.start_run():
    ...
    mlflow.pytorch.log_model(model, artifact_path="pytorch-model")
    model_uri = mlflow.get_artifact_uri("pytorch-model")
    bento_model = bentoml.mlflow.import_model('mlflow_pytorch_mnist', model_uri)

  1. When using autolog, find model_uri by last active run_id:

import mlflow
import bentoml
from sklearn.linear_model import LinearRegression

# enable autologging
mlflow.sklearn.autolog()

# prepare training data
X = np.array([[1, 1], [1, 2], [2, 2], [2, 3]])
y = np.dot(X, np.array([1, 2])) + 3

# train a model
model = LinearRegression()
model.fit(X, y)

# import logged MLflow model to BentoML
run_id = mlflow.last_active_run().info.run_id
artifact_path = "model"
model_uri = f"runs:/{run_id}/{artifact_path}"
bento_model = bentoml.mlflow.import_model('logistic_regression_model', model_uri)
print(f"Model imported to BentoML: {bento_model}")

  1. Import a registered model on MLflow server

When using a MLflow tracking server, users can also import registered models directly to BentoML for serving.

# Import from a version:
model_name = "sk-learn-random-forest-reg-model"
model_version = 1
model_uri=f"models:/{model_name}/{model_version}"
bentoml.mlflow.import_model('my_mlflow_model', model_uri)

# Import from a stage:
model_name = "sk-learn-random-forest-reg-model"
stage = 'Staging'
model_uri=f"models:/{model_name}/{stage}"
bentoml.mlflow.import_model('my_mlflow_model', model_uri)

Additional Tips#

Use MLflow model dependencies config#

Most MLflow models bundles dependency information that is required for running framework model. If no additional dependencies are required in the Service definition code, users may pass through dependency requirements from within MLflow model to BentoML.

First, put the following in your bentofile.yaml build file:

python:
    requirements_txt: $BENTOML_MLFLOW_MODEL_PATH/mlflow_model/requirements.txt
    lock_packages: False

Alternatively, one can also use MLflow modelā€™s generated conda environment file:

conda:
    environment_yml: $BENTOML_MLFLOW_MODEL_PATH/mlflow_model/conda.yaml

This allows BentoML to dynamically find the given dependency file based on a user-defined environment variable. In this case, the bentoml get CLI returns the path to the target MLflow model folder and expose it to bentoml build via the environment variable BENTOML_MLFLOW_MODEL_PATH:

export BENTOML_MLFLOW_MODEL_PATH=$(bentoml models get my_mlflow_model:latest -o path)
bentoml build

Attach model params, metrics, and tags#

MLflow model format encapsulates lots of context information regarding the training metrics and parameters. The following code snippet demonstrates how to package metadata logged from a given MLflow model to the BentoML model store.

run_id = '0e4425ecbf3e4672ba0c1741651bb47a'
run = mlflow.get_run(run_id)
model_uri = f"{run.info.artifact_uri}/model"
bentoml.mlflow.import_model(
    "my_mlflow_model",
    model_uri,
    labels=run.data.tags,
    metadata={
        "metrics": run.data.metrics,
        "params": run.data.params,
    }
)