Overview

Ray 2.0, released on August 15, 2022, introduces Ray AIR (AI Runtime) for a unified ML API covering training, tuning, and serving.

Main Features

Ray AIR

Ray AIR unifies Ray libraries (Train, Tune, Serve, Data) under a consistent API to simplify end-to-end ML workflows.

python 
import ray
from ray import train
from ray.train.xgboost import XGBoostTrainer

ray.init()

trainer = XGBoostTrainer(
    label_column='target',
    params={'max_depth': 6, 'eta': 0.3},
    datasets={'train': train_dataset},
)
result = trainer.fit()
print(result.metrics)

Unified ML API

Predictors and Checkpoints provide a common interface for model deployment, regardless of the training framework used.

python 
import ray
from ray import serve

@serve.deployment
class ModelServing:
    def __init__(self):
        self.model = None  # load model

    async def __call__(self, request):
        data = await request.json()
        return {'prediction': 42}

# serve.run(ModelServing.bind())

Sources