Multi-framework / High-performance / Easy to learn / Production ready
What does BentoML do?
- Package models trained with any ML framework and reproduce them for model serving in production
- Package once and deploy anywhere for real-time API serving or offline batch serving
- High-Performance API model server with adaptive micro-batching support
- Central storage hub with Web UI and APIs for managing and accessing packaged models
- Modular and flexible design allowing advanced users to easily customize
BentoML is a framework for serving, managing and deploying machine learning models. It is aiming to bridge the gap between Data Science and DevOps, and enable data science teams to continuesly deliver prediction services to production.
👉 Join the community: BentoML Slack Channel and BentoML Discussions.
BentoML provides abstractions for creating prediction service that's bundled with one or multiple trained models. User can define inference APIs with serving logic with Python code and specify the expected input/output data format:
import pandas as pd
from bentoml import env, artifacts, api, BentoService
from bentoml.adapters import DataframeInput
from bentoml.frameworks.sklearn import SklearnModelArtifact
from my_library import preprocess
@env(infer_pip_packages=True)
@artifacts([SklearnModelArtifact('my_model')])
class MyPredictionService(BentoService):
"""
A minimum prediction service exposing a Scikit-learn model
"""
@api(input=DataframeInput(orient="records"), batch=True)
def predict(self, df: pd.DataFrame):
"""
An inference API named `predict` with Dataframe input adapter, which codifies
how HTTP requests or CSV files are converted to a pandas Dataframe object as the
inference API function input
"""
model_input = preprocess(df)
return self.artifacts.my_model.predict(model_input)At the end of your model training pipeline, import your BentoML prediction service
class, pack it with your trained model, and persist the entire prediction service with
save call at the end:
from my_prediction_service import MyPredictionService
svc = MyPredictionService()
svc.pack('my_model', my_sklearn_model)
svc.save() # default saves to ~/bentoml/repository/MyPredictionService/{version}/This will save all the code, files, serialized models, and configs required for reproducing this prediction service for inference. BentoML automatically find all the pip package dependencies and local python code dependencies, and make sure all those are packaged and versioned with your code and model in one place.
With the saved prediction service, a user can easily start a local API server hosting it:
bentoml serve MyPredictionService:latestAnd create a docker container image for this API model server with just one command:
bentoml containerize my_prediction_service MyPredictionService:latest -t my_prediction_service
docker run -p 5000:5000 my_prediction_serviceBentoML will make sure the container has all the required dependencies installed. In addition to the model inference API, this containerized BentoML model server also comes with instrumentations, metrics/healthcheck endpoints, prediction logging, tracing and it is thus ready for your DevOps team to deploy in production.
If you are at a small team without DevOps support, BentoML also provides an one-click deployment option, which deploys the model server API to cloud platforms with minimum setup.
Read the Quickstart Guide to learn more about the basic functionalities of BentoML. You can also try it out here on Google Colab.
BentoML documentation: https://docs.bentoml.org/
- Quickstart Guide, try it out on Google Colab
- Core Concepts
- API References
- FAQ
- Example projects: bentoml/Gallery
Online serving with API model server:
- Containerized model server for production deployment with Docker, Kubernetes, OpenShift, AWS ECS, Azure, GCP GKE, etc
- Adaptive micro-batching for optimal online serving performance
- Discover and package all dependencies automatically, including PyPI, conda packages and local python modules
- Support multiple ML frameworks including PyTorch, TensorFlow, Scikit-Learn, XGBoost, and many more
- Serve compositions of multiple models
- Serve multiple endpoints in one model server
- Serve any Python code along with trained models
- Automatically generate HTTP API spec in Swagger/OpenAPI format
- Prediction logging and feedback logging endpoint
- Health check endpoint and Prometheus
/metricsendpoint for monitoring - Load and replay historical prediction request logs (roadmap)
- Model serving via gRPC endpoint (roadmap)
Advanced workflow for model serving and deployment:
- Central repository for managing all your team's packaged models via Web UI and API
- Launch inference run from CLI or Python, which enables CI/CD testing, programmatic access and batch offline inference job
- One-click deployment to cloud platforms including AWS Lambda, AWS SageMaker, and Azure Functions
- Distributed batch job or streaming job with Apache Spark (improved Spark support is on roadmap)
- Advanced model deployment workflows for Kubernetes, including auto-scaling, scale-to-zero, A/B testing, canary deployment, and multi-armed-bandit (roadmap)
- Deep integration with ML experimentation platforms including MLFlow, Kubeflow (roadmap)
- Scikit-Learn - Docs | Examples
- PyTorch - Docs | Examples
- Tensorflow 2 - Docs | Examples
- Tensorflow Keras - Docs | Examples
- XGBoost - Docs | Examples
- LightGBM - Docs | Examples
- FastText - Docs | Examples
- FastAI - Docs | Examples
- H2O - Docs | Examples
- ONNX - Docs | Examples
- Spacy - Docs | Examples
- Statsmodels - Docs | Examples
- CoreML - Docs
- Transformers - Docs
Be sure to check out deployment overview doc to understand which deployment option is best suited for your use case.
-
One-click deployment with BentoML:
-
Deploy with open-source platforms:
-
Deploy directly to cloud services:
Moving trained Machine Learning models to serving applications in production is hard. It is a sequential process across data science, engineering and DevOps teams: after a model is trained by the data science team, they hand it over to the engineering team to refine and optimize code and creates an API, before DevOps can deploy.
And most importantly, Data Science teams want to continously repeat this process, monitor the models deployed in production and ship new models quickly. It often takes months for an engineering team to build a model serving & deployment solution that allow data science teams to ship new models in a repeatable and reliable way.
BentoML is a framework designed to solve this problem. It provides high-level APIs for Data Science team to create prediction services, abstract away DevOps' infrastructure needs and performance optimizations in the process. This allows DevOps team to seamlessly work with data science side-by-side, deploy and operate their models packaged in BentoML format in production.
Check out Frequently Asked Questions page on how does BentoML compares to Tensorflow-serving, Clipper, AWS SageMaker, MLFlow, etc.
Have questions or feedback? Post a new github issue
or discuss in our Slack channel:
Want to help build BentoML? Check out our contributing guide and the development guide.
BentoML is under active development and is evolving rapidly. Currently it is a Beta release, we may change APIs in future releases.
Read more about the latest features and changes in BentoML from the releases page.
BentoML by default collects anonymous usage data using Amplitude. It only collects BentoML library's own actions and parameters, no user or model data will be collected. Here is the code that does it.
This helps BentoML team to understand how the community is using this tool and what to build next. You can easily opt-out of usage tracking by running the following command:
# From terminal:
bentoml config set usage_tracking=false# From python:
import bentoml
bentoml.config().set('core', 'usage_tracking', 'False')