User Guide & Configuring Tune

These pages will demonstrate the various features and configurations of Tune.


Before you continue, be sure to have read Key Concepts.

This document provides an overview of the core concepts as well as some of the configurations for running Tune.

Parallelism / GPUs


To run everything sequentially, use Ray Local Mode.

Parallelism is determined by resources_per_trial (defaulting to 1 CPU, 0 GPU per trial) and the resources available to Tune (ray.cluster_resources()).

Tune will allocate the specified GPU and CPU from resources_per_trial to each individual trial. A trial will not be scheduled unless at least that amount of resources is available, preventing the cluster from being overloaded.

By default, Tune automatically runs N concurrent trials, where N is the number of CPUs (cores) on your machine.

# If you have 4 CPUs on your machine, this will run 4 concurrent trials at a time., num_samples=10)

You can override this parallelism with resources_per_trial:

# If you have 4 CPUs on your machine, this will run 2 concurrent trials at a time., num_samples=10, resources_per_trial={"cpu": 2})

# If you have 4 CPUs on your machine, this will run 1 trial at a time., num_samples=10, resources_per_trial={"cpu": 4})

# Fractional values are also supported, (i.e., {"cpu": 0.5})., num_samples=10, resources_per_trial={"cpu": 0.5})

To leverage GPUs, you must set gpu in resources_per_trial. This will automatically set CUDA_VISIBLE_DEVICES for each trial.

# If you have 8 GPUs, this will run 8 trials at once., num_samples=10, resources_per_trial={"gpu": 1})

# If you have 4 CPUs on your machine and 1 GPU, this will run 1 trial at a time., num_samples=10, resources_per_trial={"cpu": 2, "gpu": 1})

You can find an example of this in the Keras MNIST example.


If ‘gpu’ is not set, CUDA_VISIBLE_DEVICES environment variable will be set as empty, disallowing GPU access.

To attach to a Ray cluster, simply run ray.init before See Starting Ray via the CLI (ray start) for more information about ray.init:

# Connect to an existing distributed Ray cluster
ray.init(address=<ray_address>), num_samples=100, resources_per_trial={"cpu": 2, "gpu": 1})

Search Space (Grid/Random)

You can specify a grid search or sampling distribution via the dict passed into

parameters = {
    "qux": tune.sample_from(lambda spec: 2 + 2),
    "bar": tune.grid_search([True, False]),
    "foo": tune.grid_search([1, 2, 3]),
    "baz": "asd",  # a constant value
}, config=parameters)

By default, each random variable and grid search point is sampled once. To take multiple random samples, add num_samples: N to the experiment config. If grid_search is provided as an argument, the grid will be repeated num_samples of times.

 # num_samples=10 repeats the 3x3 grid search 10 times, for a total of 90 trials
         "alpha": tune.uniform(100),
         "beta": tune.sample_from(lambda spec: spec.config.alpha * np.random.normal()),
         "nn_layers": [
             tune.grid_search([16, 64, 256]),
             tune.grid_search([16, 64, 256]),

Note that search spaces may not be interoperable across different search algorithms. For example, for many search algorithms, you will not be able to use a grid_search parameter. Read about this in the Search Space API page.

Reporting Metrics

You can log arbitrary values and metrics in both training APIs:

def trainable(config):
    num_epochs = 100
    for i in range(num_epochs):
        accuracy = model.train()
        metric_1 = f(model)
        metric_2 = model.get_loss(), metric_foo=random_metric_1, bar=metric_2)

class Trainable(tune.Trainable):

    def step(self):  # this is called iteratively
        accuracy = self.model.train()
        metric_1 = f(self.model)
        metric_2 = self.model.get_loss()
        # don't call report here!
        return dict(acc=accuracy, metric_foo=random_metric_1, bar=metric_2)

During training, Tune will automatically log the below metrics in addition to the user-provided values. All of these can be used as stopping conditions or passed as a parameter to Trial Schedulers/Search Algorithms.

# (Optional/Auto-filled) training is terminated. Filled only if not provided.
DONE = "done"

# (Optional) Enum for user controlled checkpoint
SHOULD_CHECKPOINT = "should_checkpoint"

# (Auto-filled) The hostname of the machine hosting the training process.
HOSTNAME = "hostname"

# (Auto-filled) The auto-assigned id of the trial.
TRIAL_ID = "trial_id"

# (Auto-filled) The auto-assigned id of the trial.
EXPERIMENT_TAG = "experiment_tag"

# (Auto-filled) The node ip of the machine hosting the training process.
NODE_IP = "node_ip"

# (Auto-filled) The pid of the training process.
PID = "pid"

# (Optional) Mean reward for current training iteration
EPISODE_REWARD_MEAN = "episode_reward_mean"

# (Optional) Mean loss for training iteration
MEAN_LOSS = "mean_loss"

# (Optional) Mean loss for training iteration
NEG_MEAN_LOSS = "neg_mean_loss"

# (Optional) Mean accuracy for training iteration
MEAN_ACCURACY = "mean_accuracy"

# Number of episodes in this iteration.
EPISODES_THIS_ITER = "episodes_this_iter"

# (Optional/Auto-filled) Accumulated number of episodes for this trial.
EPISODES_TOTAL = "episodes_total"

# Number of timesteps in this iteration.
TIMESTEPS_THIS_ITER = "timesteps_this_iter"

# (Auto-filled) Accumulated number of timesteps for this entire trial.
TIMESTEPS_TOTAL = "timesteps_total"

# (Auto-filled) Time in seconds this iteration took to run.
# This may be overriden to override the system-computed time difference.
TIME_THIS_ITER_S = "time_this_iter_s"

# (Auto-filled) Accumulated time in seconds for this entire trial.
TIME_TOTAL_S = "time_total_s"

# (Auto-filled) The index of this training iteration.
TRAINING_ITERATION = "training_iteration"


When running a hyperparameter search, Tune can automatically and periodically save/checkpoint your model. This allows you to:

  • save intermediate models throughout training

  • use pre-emptible machines (by automatically restoring from last checkpoint)

  • Pausing trials when using Trial Schedulers such as HyperBand and PBT.

To use Tune’s checkpointing features, you must expose a checkpoint_dir argument in the function signature, and call tune.checkpoint_dir:

import os
import time
from ray import tune

def train_func(config, checkpoint_dir=None):
    start = 0
    if checkpoint_dir:
        with open(os.path.join(checkpoint_dir, "checkpoint")) as f:
            state = json.loads(
            start = state["step"] + 1

    for iter in range(start, 100):

        # Obtain a checkpoint directory
        with tune.checkpoint_dir(step=step) as checkpoint_dir:
            path = os.path.join(checkpoint_dir, "checkpoint")
            with open(path, "w") as f:
                f.write(json.dumps({"step": start}))"world", ray="tune")

In this example, checkpoints will be saved by training iteration to local_dir/exp_name/trial_name/checkpoint_<step>.

You can restore a single trial checkpoint by using<checkpoint_dir>) By doing this, you can change whatever experiments’ configuration such as the experiment’s name:

# Restored previous trial from the given checkpoint
    name="RestoredExp", # The name can be different.
    stop={"training_iteration": 10}, # train 5 more iterations than previous
    config={"env": "CartPole-v0"},

Distributed Checkpointing

On a multinode cluster, Tune automatically creates a copy of all trial checkpoints on the head node. This requires the Ray cluster to be started with the cluster launcher and also requires rsync to be installed.

Note that you must use the tune.checkpoint_dir API to trigger syncing. Also, if running Tune on Kubernetes, be sure to use the KubernetesSyncer to transfer files between different pods.

If you do not use the cluster launcher, you should set up a NFS or global file system and disable cross-node syncing:

sync_config = tune.SyncConfig(sync_to_driver=False), sync_config=sync_config)

Handling Large Datasets

You often will want to compute a large object (e.g., training data, model weights) on the driver and use that object within each trial. Tune provides a pin_in_object_store utility function that can be used to broadcast such large objects. Objects pinned in this way will never be evicted from the Ray object store while the driver process is running, and can be efficiently retrieved from any task via get_pinned_object.

import ray
from ray import tune
from ray.tune.utils import pin_in_object_store, get_pinned_object

import numpy as np


# X_id can be referenced in closures
X_id = pin_in_object_store(np.random.random(size=100000000))

def f(config):
    X = get_pinned_object(X_id)
    # use X

Stopping Trials

You can control when trials are stopped early by passing the stop argument to This argument takes either a dictionary or a function.

If a dictionary is passed in, the keys may be any field in the return result of in the Function API or step() (including the results from step and auto-filled metrics).

In the example below, each trial will be stopped either when it completes 10 iterations OR when it reaches a mean accuracy of 0.98. These metrics are assumed to be increasing.

# training_iteration is an auto-filled metric by Tune.
    stop={"training_iteration": 10, "mean_accuracy": 0.98}

For more flexibility, you can pass in a function instead. If a function is passed in, it must take (trial_id, result) as arguments and return a boolean (True if trial should be stopped and False otherwise).

def stopper(trial_id, result):
    return result["mean_accuracy"] / result["training_iteration"] > 5, stop=stopper)

Finally, you can implement the Stopper abstract class for stopping entire experiments. For example, the following example stops all trials after the criteria is fulfilled by any individual trial, and prevents new ones from starting:

from ray.tune import Stopper

class CustomStopper(Stopper):
    def __init__(self):
        self.should_stop = False

    def __call__(self, trial_id, result):
        if not self.should_stop and result['foo'] > 10:
            self.should_stop = True
        return self.should_stop

    def stop_all(self):
        """Returns whether to stop trials and prevent new ones from starting."""
        return self.should_stop

stopper = CustomStopper(), stop=stopper)

Note that in the above example the currently running trials will not stop immediately but will do so once their current iterations are complete. See the Stopper (tune.Stopper) documentation.


Tune by default will log results for Tensorboard, CSV, and JSON formats. If you need to log something lower level like model weights or gradients, see Trainable Logging.

Learn more about logging and customizations here: Loggers (tune.logger).

Tune will log the results of each trial to a subfolder under a specified local dir, which defaults to ~/ray_results.

# This logs to 2 different trial folders:
# ~/ray_results/trainable_name/trial_name_1 and ~/ray_results/trainable_name/trial_name_2
# trainable_name and trial_name are autogenerated., num_samples=2)

You can specify the local_dir and trainable_name:

# This logs to 2 different trial folders:
# ./results/test_experiment/trial_name_1 and ./results/test_experiment/trial_name_2
# Only trial_name is autogenerated., num_samples=2, local_dir="./results", name="test_experiment")

To specify custom trial folder names, you can pass use the trial_name_creator argument to This takes a function with the following signature:

def trial_name_string(trial):
        trial (Trial): A generated trial object.

        trial_name (str): String representation of Trial.
    return str(trial)

See the documentation on Trials: Trial.

Tensorboard (Logging)

Tune automatically outputs Tensorboard files during To visualize learning in tensorboard, install tensorboardX:

$ pip install tensorboardX

Then, after you run an experiment, you can visualize your experiment with TensorBoard by specifying the output directory of your results.

$ tensorboard --logdir=~/ray_results/my_experiment

If you are running Ray on a remote multi-user cluster where you do not have sudo access, you can run the following commands to make sure tensorboard is able to write to the tmp directory:

$ export TMPDIR=/tmp/$USER; mkdir -p $TMPDIR; tensorboard --logdir=~/ray_results

If using TF2, Tune also automatically generates TensorBoard HParams output, as shown below:
        "lr": tune.grid_search([1e-5, 1e-4]),
        "momentum": tune.grid_search([0, 0.9])

Console Output

The following fields will automatically show up on the console output, if provided:

  1. episode_reward_mean

  2. mean_loss

  3. mean_accuracy

  4. timesteps_this_iter (aggregated into timesteps_total).

Below is an example of the console output:

== Status ==
Memory usage on this node: 11.4/16.0 GiB
Using FIFO scheduling algorithm.
Resources requested: 4/12 CPUs, 0/0 GPUs, 0.0/3.17 GiB heap, 0.0/1.07 GiB objects
Result logdir: /Users/foo/ray_results/myexp
Number of trials: 4 (4 RUNNING)
| Trial name           | status   | loc                 |    param1 | param2 |    acc | total time (s) |  iter |
| MyTrainable_a826033a | RUNNING  | | 0.303706  | 0.0761 | 0.1289 |        7.54952 |    15 |
| MyTrainable_a8263fc6 | RUNNING  | | 0.929276  | 0.158  | 0.4865 |        7.0501  |    14 |
| MyTrainable_a8267914 | RUNNING  | | 0.068426  | 0.0319 | 0.9585 |        7.0477  |    14 |
| MyTrainable_a826b7bc | RUNNING  | | 0.729127  | 0.0748 | 0.1797 |        7.05715 |    14 |

You can use a Reporter object to customize the console output.

Uploading Results

If an upload directory is provided, Tune will automatically sync results from the local_dir to the given directory, natively supporting standard S3/gsutil URIs.

You can customize this to specify arbitrary storages with the sync_to_cloud argument in tune.SyncConfig. This argument supports either strings with the same replacement fields OR arbitrary functions.

If a string is provided, then it must include replacement fields {source} and {target}, like s3 sync {source} {target}. Alternatively, a function can be provided with the following signature:

def custom_sync_func(source, target):
    # do arbitrary things inside
    sync_cmd = "s3 {source} {target}".format(
    sync_process = subprocess.Popen(sync_cmd, shell=True)

By default, syncing occurs every 300 seconds. To change the frequency of syncing, set the TUNE_CLOUD_SYNC_S environment variable in the driver to the desired syncing period.

Note that uploading only happens when global experiment state is collected, and the frequency of this is determined by the TUNE_GLOBAL_CHECKPOINT_S environment variable. So the true upload period is given by max(TUNE_CLOUD_SYNC_S, TUNE_GLOBAL_CHECKPOINT_S).

Using Tune with Kubernetes

Tune automatically syncs files and checkpoints between different remote nodes as needed. To make this work in your Kubernetes cluster, you will need to pass a KubernetesSyncer to the sync_to_driver argument of tune.SyncConfig. You have to specify your Kubernetes namespace explicitly:

from ray.tune.integration.kubernetes import NamespacedKubernetesSyncer
sync_config = tune.SyncConfig(
), sync_config=sync_config)

Redirecting stdout and stderr to files

The stdout and stderr streams are usually printed to the console. For remote actors, Ray collects these logs and prints them to the head process.

However, if you would like to collect the stream outputs in files for later analysis or troubleshooting, Tune offers an utility parameter, log_to_file, for this.

By passing log_to_file=True to, stdout and stderr will be logged to trial_logdir/stdout and trial_logdir/stderr, respectively:

If you would like to specify the output files, you can either pass one filename, where the combined output will be stored, or two filenames, for stdout and stderr, respectively:
    log_to_file=("my_stdout.log", "my_stderr.log"))

The file names are relative to the trial’s logdir. You can pass absolute paths, too.

If log_to_file is set, Tune will automatically register a new logging handler for Ray’s base logger and log the output to the specified stderr output file.


By default, Tune will run hyperparameter evaluations on multiple processes. However, if you need to debug your training process, it may be easier to do everything on a single process. You can force all Ray functions to occur on a single process with local_mode by calling the following before


Local mode with multiple configuration evaluations will interleave computation, so it is most naturally used when running a single configuration evaluation.

Note that local_mode has some known issues, so please read these tips for more info.

Stopping after the first failure

By default, will continue executing until all trials have terminated or errored. To stop the entire Tune run as soon as any trial errors:, fail_fast=True)

This is useful when you are trying to setup a large hyperparameter experiment.

Further Questions or Issues?

You can post questions or issues or feedback through the following channels:

  1. StackOverflow: For questions about how to use Ray.

  2. GitHub Issues: For bug reports and feature requests.