README.rst

Ray provides a simple, universal API for building distributed applications.

Ray is packaged with the following libraries for accelerating machine learning workloads:

• Tune: Scalable Hyperparameter Tuning
• RLlib: Scalable Reinforcement Learning
• Train: Distributed Deep Learning (beta)
• Datasets: Distributed Data Loading and Compute

As well as libraries for taking ML and distributed apps to production:

• Serve: Scalable and Programmable Serving
• Workflows: Fast, Durable Application Flows (alpha)

There are also many community integrations with Ray, including Dask, MARS, Modin, Horovod, Hugging Face, Scikit-learn, and others. Check out the full list of Ray distributed libraries here.

Install Ray with: pip install ray. For nightly wheels, see the Installation page.

Quick Start

Execute Python functions in parallel.

import ray
ray.init()

@ray.remote
def f(x):
    return x * x

futures = [f.remote(i) for i in range(4)]
print(ray.get(futures))

To use Ray's actor model:

import ray
ray.init()

@ray.remote
class Counter(object):
    def __init__(self):
        self.n = 0

    def increment(self):
        self.n += 1

    def read(self):
        return self.n

counters = [Counter.remote() for i in range(4)]
[c.increment.remote() for c in counters]
futures = [c.read.remote() for c in counters]
print(ray.get(futures))

Ray programs can run on a single machine, and can also seamlessly scale to large clusters. To execute the above Ray script in the cloud, just download this configuration file, and run:

ray submit [CLUSTER.YAML] example.py --start

Read more about launching clusters.

Tune Quick Start

Tune is a library for hyperparameter tuning at any scale.

• Launch a multi-node distributed hyperparameter sweep in less than 10 lines of code.
• Supports any deep learning framework, including PyTorch, PyTorch Lightning, TensorFlow, and Keras.
• Visualize results with TensorBoard.
• Choose among scalable SOTA algorithms such as Population Based Training (PBT), Vizier's Median Stopping Rule, HyperBand/ASHA.
• Tune integrates with many optimization libraries such as Facebook Ax, HyperOpt, and Bayesian Optimization and enables you to scale them transparently.

To run this example, you will need to install the following:

$ pip install "ray[tune]"

This example runs a parallel grid search to optimize an example objective function.

from ray import tune


def objective(step, alpha, beta):
    return (0.1 + alpha * step / 100)**(-1) + beta * 0.1


def training_function(config):
    # Hyperparameters
    alpha, beta = config["alpha"], config["beta"]
    for step in range(10):
        # Iterative training function - can be any arbitrary training procedure.
        intermediate_score = objective(step, alpha, beta)
        # Feed the score back back to Tune.
        tune.report(mean_loss=intermediate_score)


analysis = tune.run(
    training_function,
    config={
        "alpha": tune.grid_search([0.001, 0.01, 0.1]),
        "beta": tune.choice([1, 2, 3])
    })

print("Best config: ", analysis.get_best_config(metric="mean_loss", mode="min"))

# Get a dataframe for analyzing trial results.
df = analysis.results_df

If TensorBoard is installed, automatically visualize all trial results:

tensorboard --logdir ~/ray_results

RLlib Quick Start

RLlib is an industry-grade library for reinforcement learning (RL), built on top of Ray. It offers high scalability and unified APIs for a variety of industry- and research applications.

$ pip install "ray[rllib]" tensorflow  # or torch

import gym
from ray.rllib.algorithms.ppo import PPO


# Define your problem using python and openAI's gym API:
class SimpleCorridor(gym.Env):
    """Corridor in which an agent must learn to move right to reach the exit.

    ---------------------
    | S | 1 | 2 | 3 | G |   S=start; G=goal; corridor_length=5
    ---------------------

    Possible actions to chose from are: 0=left; 1=right
    Observations are floats indicating the current field index, e.g. 0.0 for
    starting position, 1.0 for the field next to the starting position, etc..
    Rewards are -0.1 for all steps, except when reaching the goal (+1.0).
    """

    def __init__(self, config):
        self.end_pos = config["corridor_length"]
        self.cur_pos = 0
        self.action_space = gym.spaces.Discrete(2)  # left and right
        self.observation_space = gym.spaces.Box(0.0, self.end_pos, shape=(1,))

    def reset(self):
        """Resets the episode and returns the initial observation of the new one.
        """
        self.cur_pos = 0
        # Return initial observation.
        return [self.cur_pos]

    def step(self, action):
        """Takes a single step in the episode given `action`

        Returns:
            New observation, reward, done-flag, info-dict (empty).
        """
        # Walk left.
        if action == 0 and self.cur_pos > 0:
            self.cur_pos -= 1
        # Walk right.
        elif action == 1:
            self.cur_pos += 1
        # Set `done` flag when end of corridor (goal) reached.
        done = self.cur_pos >= self.end_pos
        # +1 when goal reached, otherwise -1.
        reward = 1.0 if done else -0.1
        return [self.cur_pos], reward, done, {}


# Create an RLlib Trainer instance.
trainer = PPO(
    config={
        # Env class to use (here: our gym.Env sub-class from above).
        "env": SimpleCorridor,
        # Config dict to be passed to our custom env's constructor.
        "env_config": {
            # Use corridor with 20 fields (including S and G).
            "corridor_length": 20
        },
        # Parallelize environment rollouts.
        "num_workers": 3,
    })

# Train for n iterations and report results (mean episode rewards).
# Since we have to move at least 19 times in the env to reach the goal and
# each move gives us -0.1 reward (except the last move at the end: +1.0),
# we can expect to reach an optimal episode reward of -0.1*18 + 1.0 = -0.8
for i in range(5):
    results = trainer.train()
    print(f"Iter: {i}; avg. reward={results['episode_reward_mean']}")

After training, you may want to perform action computations (inference) in your environment. Here is a minimal example on how to do this. Also check out our more detailed examples here (in particular for normal models, LSTMs, and attention nets).

# Perform inference (action computations) based on given env observations.
# Note that we are using a slightly different env here (len 10 instead of 20),
# however, this should still work as the agent has (hopefully) learned
# to "just always walk right!"
env = SimpleCorridor({"corridor_length": 10})
# Get the initial observation (should be: [0.0] for the starting position).
obs = env.reset()
done = False
total_reward = 0.0
# Play one episode.
while not done:
    # Compute a single action, given the current observation
    # from the environment.
    action = trainer.compute_single_action(obs)
    # Apply the computed action in the environment.
    obs, reward, done, info = env.step(action)
    # Sum up rewards for reporting purposes.
    total_reward += reward
# Report results.
print(f"Played 1 episode; total-reward={total_reward}")

Ray Serve Quick Start

Ray Serve is a scalable model-serving library built on Ray. It is:

• Framework Agnostic: Use the same toolkit to serve everything from deep learning models built with frameworks like PyTorch or Tensorflow & Keras to Scikit-Learn models or arbitrary business logic.
• Python First: Configure your model serving declaratively in pure Python, without needing YAMLs or JSON configs.
• Performance Oriented: Turn on batching, pipelining, and GPU acceleration to increase the throughput of your model.
• Composition Native: Allow you to create "model pipelines" by composing multiple models together to drive a single prediction.
• Horizontally Scalable: Serve can linearly scale as you add more machines. Enable your ML-powered service to handle growing traffic.

To run this example, you will need to install the following:

$ pip install scikit-learn
$ pip install "ray[serve]"

This example runs serves a scikit-learn gradient boosting classifier.

import pickle
import requests

from sklearn.datasets import load_iris
from sklearn.ensemble import GradientBoostingClassifier

from ray import serve

serve.start()

# Train model.
iris_dataset = load_iris()
model = GradientBoostingClassifier()
model.fit(iris_dataset["data"], iris_dataset["target"])

@serve.deployment(route_prefix="/iris")
class BoostingModel:
    def __init__(self, model):
        self.model = model
        self.label_list = iris_dataset["target_names"].tolist()

    async def __call__(self, request):
        payload = (await request.json())["vector"]
        print(f"Received flask request with data {payload}")

        prediction = self.model.predict([payload])[0]
        human_name = self.label_list[prediction]
        return {"result": human_name}


# Deploy model.
BoostingModel.deploy(model)

# Query it!
sample_request_input = {"vector": [1.2, 1.0, 1.1, 0.9]}
response = requests.get("http://localhost:8000/iris", json=sample_request_input)
print(response.text)
# Result:
# {
#  "result": "versicolor"
# }

More Information

• Documentation
• Tutorial
• Blog
• Ray 1.0 Architecture whitepaper (new)
• Exoshuffle: large-scale data shuffle in Ray (new)
• RLlib paper
• RLlib flow paper
• Tune paper

Older documents:

• Ray paper
• Ray HotOS paper

Getting Involved

Platform	Purpose	Estimated Response Time	Support Level
Discourse Forum	For discussions about development and questions about usage.	< 1 day	Community
GitHub Issues	For reporting bugs and filing feature requests.	< 2 days	Ray OSS Team
Slack	For collaborating with other Ray users.	< 2 days	Community
StackOverflow	For asking questions about how to use Ray.	3-5 days	Community
Meetup Group	For learning about Ray projects and best practices.	Monthly	Ray DevRel
Twitter	For staying up-to-date on new features.	Daily	Ray DevRel