Skip to content

Latest commit

 

History

History
219 lines (158 loc) · 6.58 KB

README.md

File metadata and controls

219 lines (158 loc) · 6.58 KB
Raw

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
  • Apache MXNet model
  • ONNX model (CPU)
  • TensorRT model
  • Python script model

You can install extra extensions to enable the following models:

  • PaddlePaddle model
  • TFLite model
  • Neo DLR (TVM) model
  • XGBoost 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

brew install djl-serving

# Start djl-serving as service:
brew services start djl-serving

# Stop djl-serving service
brew services stop djl-serving

For Ubuntu

curl -O https://publish.djl.ai/djl-serving/djl-serving_0.17.0-1_all.deb
sudo dpkg -i djl-serving_0.17.0-1_all.deb

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.

curl -O https://publish.djl.ai/djl-serving/serving-0.17.0.zip
unzip serving-0.17.0.zip
# start djl-serving
serving-0.17.0\bin\serving.bat

Docker

You can also use docker to run DJL Serving:

docker run -itd -p 8080:8080 deepjavalibrary/djl-serving

Usage

Sample Usage

Use the following command to start model server locally:

djl-serving

The model server will be listening on port 8080. You can also load a model for serving on start up:

djl-serving -m "https://resources.djl.ai/demo/mxnet/resnet18_v1.zip"

Open another terminal, and type the following command to test the inference REST API:

curl -O https://resources.djl.ai/images/kitten.jpg
curl -X POST http://localhost:8080/predictions/resnet18_v1 -T kitten.jpg

or:

curl -X POST http://localhost:8080/predictions/resnet18_v1 -F "data=@kitten.jpg"

[
  {
    "className": "n02123045 tabby, tabby cat",
    "probability": 0.4838452935218811
  },
  {
    "className": "n02123159 tiger cat",
    "probability": 0.20599420368671417
  },
  {
    "className": "n02124075 Egyptian cat",
    "probability": 0.18810515105724335
  },
  {
    "className": "n02123394 Persian cat",
    "probability": 0.06411745399236679
  },
  {
    "className": "n02127052 lynx, catamount",
    "probability": 0.010215568356215954
  }
]

Examples for loading models

# Load models from the DJL model zoo on startup
djl-serving -m "djl://ai.djl.pytorch/resnet"

# Load version v1 of a PyTorch model on GPU(0) from the local file system
djl-serving -m "resnet:v1:PyTorch:0=file:$HOME/models/pytorch/resnet18/"

# Load a TensorFlow model from TFHub
djl-serving -m "resnet=https://tfhub.dev/tensorflow/resnet_50/classification/1"

Examples for customizing data processing

# Use the default data processing for a well-known application
djl-serving -m "file:/resnet?application=CV/image_classification"

# Specify a custom data processing with a Translator
djl-serving -m "file:/resnet?translatorFactory=MyFactory"

## Pass parameters for data processing
djl-serving -m "djl://ai.djl.pytorch/resnet?applySoftmax=false"

Using DJL Extensions

# Load a model from an AWS S3 Bucket
djl-serving -m "s3://djl-ai/demo/resnet/resnet18.zip"

# Load a model from HDFS
djl-serving -m "hdfs://localhost:50070/models/pytorch/resnet18/"

# Use a HuggingFace tokenizer
djl-serving -m "file:/resnet?transaltorFactory=ai.djl.huggingface.BertQATranslator"

More examples

More command line options

djl-serving --help
usage: djl-serving [OPTIONS]
 -f,--config-file <CONFIG-FILE>    Path to the configuration properties file.
 -h,--help                         Print this help.
 -m,--models <MODELS>              Models to be loaded at startup.
 -s,--model-store <MODELS-STORE>   Model store location where models can be loaded.
 -w,--workflows <WORKFLOWS>   Workflows to be loaded at startup.

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

DJL serving is built on top of Deep Java Library. You can visit the DJL github repository to learn more.

It is also possible to leverage only the worker thread pool using the separate WorkLoadManager module. The separate WorkLoadManager can be used to take advantage of DJL serving's model batching and threading and integrate it into a custom Java service.

Architecture Diagram

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

-Dai.djl.logging.level={FATAL|ERROR|WARN|INFO|DEBUG|TRACE}