JAMS

This crate provides a CLI for interacting J.A.M.S - Just Another Model Server.

⚠️ DISCLAIMER: jams is reaching stable version but may not run properly on ARM chips. Future releases will fix this. For now use docker image or Linux x86_64 architecture. Only Pytorch 2.2.0 is supported for now due to dependencies on FFI bindings. Although you may be able to run models trained on version <= 2.2.0

---

Features

• Async
• Separate Rayon Threadpool for computing predictions
• Multiple Model Frameworks Supported
  • Tensorflow
  • Torch*
  • Catboost
  • LightGBM
• Multiple Model Store Backends Supported with Polling
  • Local File System
  • AWS S3
  • Azure Blob Storage
  • MinIO
• Thin & Fast API Layer
  • HTTP via Axum
  • gRPC via Tonic
• Ready to use clients
  • Python
  • Go
  • Rust
  • TypeScript 🚧
  • JAVA 🚧

The following features are in progress 🚧

• Support XGBoost framework
• ModelSpec - Single source of information about models. This will assist in input validations

---

Docker Setup

J.A.M.S is hosted on DockerHub.

docker pull gagansingh894/jams

Docker Compose

Ensure that you have docker compose installed on your system. Please follow instructions here

Note: If you are on Apple Silicon, please disable Use Rosetta for x86_64/amd64 emulation on Apple Silicon option under settings

To quickly get started running with J.A.M.S, please run the following commands

1. Execute bash script

The script creates a jams-playground directory with a subdirectory models, and then downloads a Docker Compose configuration file and several machine learning model files from a GitHub repository into the respective directories. It runs with superuser privileges using sudo to ensure necessary permissions for directory creation and file downloads.

sudo bash -c 'mkdir -p jams-playground/models && \
wget -q -O jams-playground/docker-compose-http-grpc-minio.yml https://raw.githubusercontent.com/gagansingh894/jams-rs/main/build/docker-compose-http-grpc-minio.yml && \
wget -q -O jams-playground/models/catboost-titanic_model.tar.gz https://github.com/gagansingh894/jams-rs/raw/main/jams-serve/tests/model_store/catboost-titanic_model.tar.gz && \
wget -q -O jams-playground/models/lightgbm-my_awesome_reg_model.tar.gz https://github.com/gagansingh894/jams-rs/raw/main/jams-serve/tests/model_store/lightgbm-my_awesome_reg_model.tar.gz && \
wget -q -O jams-playground/models/pytorch-my_awesome_californiahousing_model.tar.gz https://github.com/gagansingh894/jams-rs/raw/main/jams-serve/tests/model_store/pytorch-my_awesome_californiahousing_model.tar.gz && \
wget -q -O jams-playground/models/tensorflow-my_awesome_penguin_model.tar.gz https://github.com/gagansingh894/jams-rs/raw/main/jams-serve/tests/model_store/tensorflow-my_awesome_penguin_model.tar.gz'

2. Run docker compose

docker compose -f jams-playground/docker-compose-http-grpc-minio.yml up

If everything works fine, this should start a minio server with some preloaded models as model store, J.A.M.S http and J.A.M.S grpc server for making predictions. You can add new models by uploading them directly to minio via UI (http://0.0.0.0:9001). The models should be of supported types and follow the naming convention <model_framework>-model_name.tar.gz.

Use the curl commands to make predictions

curl --location '0.0.0.0:3000/api/predict' \
--header 'Content-Type: application/json' \
--header 'Accept: application/json' \
--data '{
  "model_name": "my_awesome_penguin_model",
  "input": "{\"island\":[2.0,2.0,2.0,2.0,2.0],\"bill_length_mm\":[39.1,39.5,40.3,36.7,39.3],\"bill_depth_mm\":[18.7,17.4,18.0,19.3,20.6],\"flipper_length_mm\":[181.0,186.0,195.0,193.0,190.0],\"body_mass_g\":[3750.0,3800.0,3250.0,3450.0,3650.0],\"sex\":[1.0,0.0,0.0,0.0,1.0]}"
}'

curl --location '0.0.0.0:3000/api/predict' \
--header 'Content-Type: application/json' \
--header 'Accept: application/json' \
--data '{
  "model_name": "titanic_model",
  "input": "{\"adult_male\":[\"True\",\"False\"],\"age\":[22.0,23.79929292929293],\"alone\":[\"True\",\"False\"],\"class\":[\"First\",\"Third\"],\"deck\":[\"Unknown\",\"Unknown\"],\"embark_town\":[\"Southampton\",\"Cherbourg\"],\"embarked\":[\"S\",\"C\"],\"fare\":[151.55,14.4542],\"parch\":[\"0\",\"0\"],\"pclass\":[\"1\",\"3\"],\"sex\":[\"male\",\"female\"],\"sibsp\":[\"0\",\"1\"],\"who\":[\"man\",\"woman\"]}"
}'

Alternatively, you can use Postman or equivalent.

Config File

The easiest way to start J.A.M.S is by providing a config TOML file

[config]
protocol = "http"                               # Specifies the protocol to be used by the server.
                                                # Allowed values: "http", "grpc"

port = 3000                                     # Defines the port number on which the server will listen.
                                                # This should be an integer between 1 and 65535.
                                                # Example: 3000 for HTTP, 443 for HTTPS

model_store = "local"                           # Indicates the type of model store being used.
                                                # Allowed values:
                                                # - "local": Use local storage.
                                                # - "aws": Use AWS S3 for model storage.
                                                # - "azure": Use Azure Blob Storage.

model_dir = "<absolute path>"                   # Specifies the directory path where models are stored locally.
                                                # If `model_store` is set to "local", this directory is used
                                                # to store or load models.

azure_storage_container_name = "jamsmodelstore" # Specifies the name of the Azure Blob Storage container
                                                # used for storing models when `model_store` is set to "azure".
                                                # This should be a valid container name in Azure.

s3_bucket_name = "jamsmodelstore"               # Indicates the name of the S3 bucket used for storing models
                                                # when `model_store` is set to "aws".
                                                # This should be a valid bucket name in AWS S3.

poll_interval = 600                             # Defines the time interval (in seconds) for polling the model store
                                                # to check for updates.
                                                # Example: 600 means the application will poll every 10 minutes.

num_workers = 4                                 # Sets the number of Rayon threadpool worker threads
                                                # Example: 4 threads

Then Run

docker run --rm -v /your/path/to/model_store:/model_store -p 3000:3000 gagansingh894/jams start -f config.toml

There are other ways to start J.A.M.S. Please follow the following commands to start the server inside docker.

If you want to disable polling, then do not pass --poll-interval

To run HTTP server, use

docker run --rm -v /your/path/to/model_store:/model_store -p 3000:3000 gagansingh894/jams start http --model-dir local --poll-interval 3600

To run gRPC server, use

docker run --rm -v /your/path/to/model_store:/model_store -p 4000:4000 gagansingh894/jams start grpc --model-store=local --poll-interval 3600

To run with a S3/MinIo backend

• Create a S3 bucket with some models in it. Please refer to the structure of model store here.
• Set the environment variables - AWS_REGION, AWS_ACCESS_KEY_ID, AWS_SECRET_ACCESS_KEY. Alternatively if you have multiple AWS profiles then just set the AWS_PROFILE-<profile_name> You also need to set the bucket name. This can either be set via S3_BUCKET_NAME env variable or passed via --s3-bucket-name flag
• If using minio, you will need to set these environment variables in addition to the above = MINIO_ACCESS_KEY_ID, MINIO_SECRET_ACCESS_KEY and MINIO_URL. The default values are minioadmin and http://0.0.0.0:9000 respectively
• Run the command to start HTTP server with S3 model store. It assumes that bucket name is already set via S3_BUCKET_NAME

docker run --rm -p 3000:3000 gagansingh894/jams start http --model-store=<aws|minio> --poll-interval 3600

• For gRPC server, use

docker run --rm -p 4000:4000 gagansingh894/jams start grpc --model-store=<aws|minio> --poll-interval 3600

• If you want to pass bucket name, use

docker run --rm -p 3000:3000 gagansingh894/jams start http --model-store=<aws|minio> --s3-bucket-name=<bucket_name>

To run with a Azure Blob Storage backend

• Create a Azure Storage container with some models in it. Please refer to the structure of model store here.
• Set the environment variables - STORAGE_ACCOUNT, STORAGE_ACCESS_KEY. You also need to set the azure container name. This can either be set via AZURE_STORAGE_CONTAINER_NAME env variable or passed via --azure-storage-container-name flag
• Run the command to start HTTP server with Azure model store. It assumes that container name is already set via AZURE_STORAGE_CONTAINER_NAME

docker run --rm -p 3000:3000 gagansingh894/jams start http --model-store=azure --poll-interval 3600

• For gRPC server, use

docker run --rm -p 4000:4000 gagansingh894/jams start grpc --model-store=azure --poll-interval 3600

• If you want to pass container name, use

docker run --rm -p 3000:3000 gagansingh894/jams start http --model-store=azure --azure-storage-container-name=<container_name> --poll-interval 3600

Please refer to OpenAPI Spec for API endpoints. Alternatively, you can also refer to the proto definition.

---

---

Local Setup

Ensure that Cargo and Rust compiler are installed. Follow instructions here if not installed

This project relies on a couple of shared libraries. To easily set up, please follow the steps below

Mac - Apple Silicon

1. Install Homebrew if not already installed
2. Install Rust if not already installed. The MSRV is 1.81.
3. Run the following command to install bazel, lightgbm, pytorch and tensorflow

brew install bazel lightgbm tensorflow

3. Download Pytorch 2.2.0 from Pytorch website and set it to path

sudo sh -c '
    echo "Downloading libtorch for macOS (ARM64)...";
    wget https://download.pytorch.org/libtorch/cpu/libtorch-macos-arm64-2.2.0.zip -O /usr/local/lib/libtorch-macos-arm64-2.2.0.zip;

    echo "Unzipping libtorch into /usr/local/lib/libtorch2_2_0...";
    mkdir -p /usr/local/lib/libtorch2_2_0;
    unzip /usr/local/lib/libtorch-macos-arm64-2.2.0.zip -d /usr/local/lib/libtorch2_2_0;

    echo "Cleaning up by deleting the zip file...";
    rm /usr/local/lib/libtorch-macos-arm64-2.2.0.zip;

    echo "Please add the following to your .bashrc/.zshrc and restart terminal...";
    COMMON_LIBS_PATH=/usr/local/lib
    LIBTORCH=$COMMON_LIBS_PATH/libtorch2_2_0/libtorch
    LIBTORCH_INCLUDE=$COMMON_LIBS_PATH/libtorch2_2_0/libtorch
    LIBTORCH_LIB=$COMMON_LIBS_PATH/libtorch2_2_0/libtorch
    DYLD_LIBRARY_PATH=$COMMON_LIBS_PATH:$COMMON_LIBS_PATH/libtorch2_2_0/libtorch

    echo "Libtorch installed in /usr/local/lib/libtorch2_2_0 and zip file deleted.";
'

4. Download catboost library(.dylib) directly from Github

sudo wget -q https://github.com/catboost/catboost/releases/download/v1.2.5/libcatboostmodel-darwin-universal2-1.2.5.dylib -O /usr/local/lib/libcatboostmodel.dylib

5. Copy lightGBM library(.dylib) and libomp library to usr/local/lib

sudo cp /opt/homebrew/Cellar/lightgbm/$(brew list --versions lightgbm | awk '{print $2}')/lib/lib_lightgbm.dylib /usr/local/lib/ && sudo cp /opt/homebrew/Cellar/libomp/$(brew list --versions libomp | awk '{print $2}')/lib/libomp.dylib /usr/local/lib/

6. Add the following environment variables to .bashrc/.zshrc

export COMMON_LIBS_PATH=/usr/local/lib
export LIBTORCH=$COMMON_LIBS_PATH/libtorch2_2_0/libtorch
export LIBTORCH_INCLUDE=$COMMON_LIBS_PATH/libtorch2_2_0/libtorch
export LIBTORCH_LIB=$COMMON_LIBS_PATH/libtorch2_2_0/libtorch
export DYLD_LIBRARY_PATH=$COMMON_LIBS_PATH:$COMMON_LIBS_PATH/libtorch2_2_0/libtorch/lib

6. Run the following command to install jams

cargo install jams

Linux

1. Use the bash script here based on your system architecture

2. Run the following commands or add them to shell profile

# add environment variables
export COMMON_LIBS_PATH=/usr/local/lib
export LIGHTGBM_LIB_DIR=$COMMON_LIBS_PATH
export LIBTORCH=$COMMON_LIBS_PATH/libtorch
export LIBTORCH_INCLUDE=$COMMON_LIBS_PATH/libtorch
export LIBTORCH_LIB=$COMMON_LIBS_PATH/libtorch
export LD_LIBRARY_PATH=$COMMON_LIBS_PATH:$COMMON_LIBS_PATH/libtorch/lib
export LIBRARY_PATH=$LIBRARY_PATH:$COMMON_LIBS_PATH/libtensorflow
export LD_LIBRARY_PATH=$LD_LIBRARY_PATH:$COMMON_LIBS_PATH/libtensorflow/lib

3. Run the following command to install J.A.M.S

cargo install jams

---

API Endpoints

Once J.A.M.S is up and running, these endpoints will help you interact with the server.

Please refer to OpenAPI Spec for details.

/healthcheck: Endpoint for health checks

/api/predict: Endpoint for making predictions

/api/models: Endpoint for managing models

Alternatively, you can refer also refer to the proto definition. It provides the following RPCs

• HealthCheck
• Predict
• GetModels
• AddModel
• UpdateModel
• DeleteModel

---

Usage

The CLI provides the following commands

- jams start
- jams predict

start

Use this command to start either the HTTP/gRPC model server on 0.0.0.0:3000/0.0.0.0:4000 with separate rayon threadpool for computing predictions

The server expects a model directory containing the models. This can be either passed using the --model-dir flag

To start HTTP server

jams start http --model-dir path/to/model_dir

To start gRPC server

jams start grpc --model-dir path/to/model_dir

Alternatively, you can set the MODEL_STORE_DIR env variable pointing to the model directory and run jams start http or jams start grpc

export MODEL_STORE_DIR=path/to/model_dir

By default, the server runs on port 3000 and 2 workers in the rayon threadpool.You can override using the --port and --num-workers flags respectively. The log level can also be changed to DEBUG level using --use-debug-level=true.

Model Store

Below is the expected structure of model stores.

• Notice the model naming convention <model_framework>-model_name.tar.gz.
• The server unpacks and loads the model files.
• The server will warn about the unsupported formats and continue to load other models

└── model_store
    ├── catboost-my_awesome_binary_model.tar.gz
    ├── catboost-my_awesome_multiclass_model.tar.gz
    ├── catboost-my_awesome_regressor_model.tar.gz
    ├── catboost-titanic_model.tar.gz
    ├── lightgbm-my_awesome_binary_model_2.tar.gz
    ├── lightgbm-my_awesome_reg_model.tar.gz
    ├── lightgbm-my_awesome_xen_binary_model.tar.gz
    ├── lightgbm-my_awesome_xen_prob_model.tar.gz
    ├── pytorch-my_awesome_californiahousing_model.tar.gz
    ├── tensorflow-my_awesome_autompg_model.tar.gz
    ├── tensorflow-my_awesome_penguin_model.tar.gz
    ├── tensorflow-my_awesome_sequential_model.tar.gz
    └── torch-my_awesome_penguin_model.tar.gz

predict

Use this command for making predictions via CLI for making predictions for the following models

• Tensorflow
• Torch
• Catboost
• LightGBM

Refer below for some examples of the predict command.

There are multiple python scripts in examples folder which would allow you to generate different models and their corresponding sample json input. Below are some examples

Tensorflow

1. Run tensorflow_penguin_multiclass_classification_model.py
2. This will create two files- a model file and input json file
3. Run the following command and pass in the path for model file and input file

jams predict tensorflow --model-path=tensorflow_penguin_functional --input-path=tensorflow_input.json

Torch

1. Run torch_penguin_multiclass_classification_model.py
2. This will create two files- a model file and input json file
3. Run the following command and pass in the path for model file and input file

jams predict torch --model-path=torch_penguin.pt --input-path=torch_input.json

Catboost

1. Run catboost_titanic_binary_classification_model.py
2. This will create two files- a model file and input json file
3. Run the following command and pass in the path for model file and input file

jams predict catboost --model-path=catboost_titanic --input-path=catboost_input.json

LightGBM

1. Run lightgbm_iris_binary_classification_model.py
2. This will create two files- a model file and input json file
3. Run the following command(example) and pass in the path for model file and input file

jams predict lightgbm --model-path=lightgbm_iris.txt --input-path=lightgbm_input.json