djl-serving for Mac

DJL Serving

Overview

DJL Serving is a high performance universal stand-alone model serving solution powered by DJL. It takes a deep learning model, several models, or workflows and makes them available through an HTTP endpoint. It can serve the following model types out of the box: PyTorch TorchScript model TensorFlow SavedModel bundle ONNX model (CPU) Python script model

You can install extra extensions to enable the following models: XGBoost model LightGBM model Sentencepiece model fastText/BlazingText model

Key features Performance - DJL serving running multithreading inference in a single JVM. Our benchmark shows DJL serving has higher throughput than most C++ model servers on the market Ease of use - DJL serving can serve most models out of the box Easy to extend - DJL serving plugins make it easy to add custom extensions Auto-scale - DJL serving automatically scales up/down worker threads based on the load Dynamic batching - DJL serving supports dynamic batching to increase throughput Model versioning - DJL allows users to load different versions of a model on a single endpoint Multi-engine support - DJL allows users to serve models from different engines at the same time

Installation

For macOS

For Ubuntu

For Windows

We are considering to create a chocolatey package for Windows. For the time being, you can download djl-serving zip file from here.

Docker

You can also use docker to run DJL Serving:

Usage

DJL Serving can be started from the command line. To see examples, see the starting page.

More examples Serving a Python model Serving with docker

More command line options

See configuration for more details about defining models, model-store, and workflows.

REST API

DJL Serving uses a RESTful API for both inference and management calls.

When DJL Serving starts up, it has two web services:

Inference API - Used by clients to query the server and run models Management API - Used to add, remove, and scale models on the server

By default, DJL Serving listens on port 8080 and is only accessible from localhost. Please see DJL Serving Configuration for how to enable access from a remote host.

Architecture

Details about how DJL Serving is implemented can be found in the architecture docs.

Plugin management

DJL Serving supports plugins, user can implement their own plugins to enrich DJL Serving features. See DJL Plugin Management for how to install plugins to DJL Serving.

Logging you can set the logging level on the command-line adding a parameter for the JVM