A distribution strategy for synchronous training on multiple workers.

Inherits From: Strategy

This strategy implements synchronous distributed training across multiple workers, each with potentially multiple GPUs. Similar to tf.distribute.MirroredStrategy, it replicates all variables and computations to each local device. The difference is that it uses a distributed collective implementation (e.g. all-reduce), so that multiple workers can work together.

You need to launch your program on each worker and configure cluster_resolver correctly. For example, if you are using tf.distribute.cluster_resolver.TFConfigClusterResolver, each worker needs to have its corresponding task_type and task_id set in the TF_CONFIG environment variable. An example TF_CONFIG on worker-0 of a two worker cluster is:

TF_CONFIG = '{"cluster": {"worker": ["localhost:12345", "localhost:23456"]}, "task": {"type": "worker", "index": 0} }'

Your program runs on each worker as-is. Note that collectives require each worker to participate. All tf.distribute and non tf.distribute API may use collectives internally, e.g. checkpointing and saving since reading a tf.Variable with tf.VariableSynchronization.ON_READ all-reduces the value. Therefore it's recommended to run exactly the same program on each worker. Dispatching based on task_type or task_id of the worker is error-prone.

cluster_resolver.num_accelerators() determines the number of GPUs the strategy uses. If it's zero, the strategy uses the CPU. All workers need to use the same number of devices, otherwise the behavior is undefined.

This strategy is not intended for TPU. Use tf.distribute.TPUStrategy instead.

After setting up TF_CONFIG, using this strategy is similar to using tf.distribute.MirroredStrategy and tf.distribute.TPUStrategy.

strategy = tf.distribute.MultiWorkerMirroredStrategy()

with strategy.scope():
  model = tf.keras.Sequential([
    tf.keras.layers.Dense(2, input_shape=(5,)),
  optimizer = tf.keras.optimizers.SGD(learning_rate=0.1)

def dataset_fn(ctx):
  x = np.random.random((2, 5)).astype(np.float32)
  y = np.random.randint(2, size=(2, 1))
  dataset =, y))
  return dataset.repeat().batch(1, drop_remainder=True)
dist_dataset = strategy.distribute_datasets_from_function(dataset_fn)


You can also write your own training loop:

def train_step(iterator):

  def step_fn(inputs):
    features, labels = inputs
    with tf.GradientTape() as tape:
      logits = model(features, training=True)
      loss = tf.keras.losses.sparse_categorical_crossentropy(
          labels, logits)

    grads = tape.gradient(loss, model.trainable_variables)
    optimizer.apply_gradients(zip(grads, model.trainable_variables)), args=(next(iterator),))

for _ in range(NUM_STEP):

See Multi-worker training with Keras for a detailed tutorial.


You need to save and checkpoint on all workers instead of just one. This is because variables whose synchronization=ON_READ triggers aggregation during saving. It's recommended to save to a different path on each worker to avoid race conditions. Each worker saves the same thing. See Multi-worker training with Keras tutorial for examples.

Known Issues

cluster_resolver Returns the cluster resolver associated with this strategy.

In general, when using a multi-worker tf.distribute strategy such as tf.distribute.experimental.MultiWorkerMirroredStrategy or tf.distribute.TPUStrategy(), there is a tf.distribute.cluster_resolver.ClusterResolver associated with the strategy used, and such an instance is returned by this property.

Strategies that intend to have an associated tf.distribute.cluster_resolver.ClusterResolver must set the relevant attribute, or override this property; otherwise, None is returned by default. Those strategies should also provide information regarding what is returned by this property.

Single-worker strategies usually do not have a tf.distribute.cluster_resolver.ClusterResolver, and in those cases this property will return None.

The tf.distribute.cluster_resolver.ClusterResolver may be useful when the user needs to access information such as the cluster spec, task type or task id. For example,

os.environ['TF_CONFIG'] = json.dumps({
  'cluster': {
      'worker': ["localhost:12345", "localhost:23456"],
      'ps': ["localhost:34567"]
  'task': {'type': 'worker', 'index': 0}

# This implicitly uses TF_CONFIG for the cluster and current task info.
strategy = tf.distribute.experimental.MultiWorkerMirroredStrategy()


if strategy.cluster_resolver.task_type == 'worker':
  # Perform something that's only applicable on workers. Since we set this
  # as a worker above, this block will run on this particular instance.
elif strategy.cluster_resolver.task_type == 'ps':
  # Perform something that's only applicable on parameter servers. Since we
  # set this as a worker above, this block will not run on this particular
  # instance.

For more information, please see tf.distribute.cluster_resolver.ClusterResolver's API docstring.

extended tf.distribute.StrategyExtended with additional methods.
num_replicas_in_sync Returns number of replicas over which gradients are aggregated.



View source

Distributes instances created by calls to dataset_fn.

The argument dataset_fn that users pass in is an input function that has a tf.distribute.InputContext argument and returns a instance. It is expected that the returned dataset from dataset_fn is already batched by per-replica batch size (i.e. global batch size divided by the number of replicas in sync) and sharded. tf.distribute.Strategy.distribute_datasets_from_function does not batch or shard the instance returned from the input function. dataset_fn will be called on the CPU device of each of the workers and each generates a dataset where every replica on that worker will dequeue one batch of inputs (i.e. if a worker has two replicas, two batches will be dequeued from the Dataset every step).

This method can be used for several purposes. First, it allows you to specify your own batching and sharding logic. (In contrast, tf.distribute.experimental_distribute_dataset does batching and sharding for you.) For example, where experimental_distribute_dataset is unable to shard the input files, this method might be used to manually shard the dataset (avoiding the slow fallback behavior in experimental_distribute_dataset). In cases where the dataset is infinite, this sharding can be done by creating dataset replicas that differ only in their random seed.

The dataset_fn should take an tf.distribute.InputContext instance where information about batching and input replication can be accessed.

You can use element_spec property of the tf.distribute.DistributedDataset returned by this API to query the tf.TypeSpec of the elements returned by the iterator. This can be used to set the input_signature property of a tf.function. Follow tf.distribute.DistributedDataset.element_spec to see an example.

For a tutorial on more usage and properties of this method, refer to the tutorial on distributed input). If you are interested in last partial batch handling, read this section.

dataset_fn A function taking a tf.distribute.InputContext instance and returning a
options tf.distribute.InputOptions used to control options on how this dataset is distributed.

A tf.distribute.DistributedDataset.


View source

Creates tf.distribute.DistributedDataset from

The returned tf.distribute.DistributedDataset can be iterated over similar to regular datasets. NOTE: The user cannot add any more transformations to a tf.distribute.DistributedDataset. You can only create an iterator or examine the tf.TypeSpec of the data generated by it. See API docs of tf.distribute.DistributedDataset to learn more.

The following is an example:

global_batch_size = 2
# Passing the devices is optional.
strategy = tf.distribute.MirroredStrategy(devices=["GPU:0", "GPU:1"])
# Create a dataset
dataset =
# Distribute that dataset
dist_dataset = strategy.experimental_distribute_dataset(dataset)
def replica_fn(input):
  return input*2
result = []
# Iterate over the `tf.distribute.DistributedDataset`
for x in dist_dataset:
  # process dataset elements
  result.append(, args=(x,)))
  0: <tf.Tensor: shape=(1,), dtype=int64, numpy=array([0])>,
  1: <tf.Tensor: shape=(1,), dtype=int64, numpy=array([2])>
}, PerReplica:{
  0: <tf.Tensor: shape=(1,), dtype=int64, numpy=array([4])>,
  1: <tf.Tensor: shape=(1,), dtype=int64, numpy=array([6])>

Three key actions happening under the hood of this method are batching, sharding, and prefetching.

In the code snippet above, dataset is batched by global_batch_size, and calling experimental_distribute_dataset on it rebatches dataset to a new batch size that is equal to the global batch size divided by the number of replicas in sync. We iterate through it using a Pythonic for loop. x is a tf.distribute.DistributedValues containing data for all replicas, and each replica gets data of the new batch size. will take care of feeding the right per-replica data in x to the right replica_fn executed on each replica.

Sharding contains autosharding across multiple workers and within every worker. First, in multi-worker distributed training (i.e. when you use tf.distribute.experimental.MultiWorkerMirroredStrategy or tf.distribute.TPUStrategy), autosharding a dataset over a set of workers means that each worker is assigned a subset of the entire dataset (if the right is set). This is to ensure that at each step, a global batch size of non-overlapping dataset elements will be processed by each worker. Autosharding has a couple of different options that can be specified using Then, sharding within each worker means the method will split the data among all the worker devices (if more than one a present). This will happen regardless of multi-worker autosharding.

By default, this method adds a prefetch transformation at the end of the user provided instance. The argument to the prefetch transformation which is buffer_size is equal to the number of replicas in sync.

If the above batch splitting and dataset sharding logic is undesirable, please use tf.distribute.Strategy.distribute_datasets_from_function instead, which does not do any automatic batching or sharding for you.

For a tutorial on more usage and properties of this method, refer to the tutorial on distributed input. If you are interested in last partial batch handling, read this section.

dataset that will be sharded across all replicas using the rules stated above.
options tf.distribute.InputOptions used to control options on how this dataset is distributed.

A tf.distribute.DistributedDataset.


View source

Returns the list of all local per-replica values contained in value.

value A value returned by experimental_run(), run(), or a variable created inscope`.

A tuple of values contained in value where ith element corresponds to ith replica. If value represents a single value, this returns (value,).


View source

Makes a for input provided via a numpy array.

This avoids adding numpy_input as a large constant in the graph, and copies the data to the machine or machines that will be processing the input.

Note that you will likely need to use tf.distribute.Strategy.experimental_distribute_dataset with the returned dataset to further distribute it with the strategy.


numpy_input = np.ones([10], dtype=np.float32)
dataset = strategy.experimental_make_numpy_dataset(numpy_input)
dist_dataset = strategy.experimental_distribute_dataset(dataset)

numpy_input A nest of NumPy input arrays that will be converted into a dataset. Note that lists of Numpy arrays are stacked, as that is normal behavior.
session (TensorFlow v1.x graph execution only) A session used for initialization.

A representing numpy_input.


View source

Runs ops in fn on each replica, with inputs from input_iterator. (deprecated)

When eager execution is enabled, executes ops specified by fn on each replica. Otherwise, builds a graph to execute the ops on each replica.

Each replica will take a single, different input from the inputs provided by one get_next call on the input iterator.

fn may call tf.distribute.get_replica_context() to access members such as replica_id_in_sync_group.

fn The function to run. The inputs to the function must match the outputs of input_iterator.get_next(). The output must be a tf.nest of Tensors.
input_iterator (Optional) input iterator from which the inputs are taken.

Merged return value of fn across replicas. The structure of the return value is the same as the return value from fn. Each element in the structure can either be PerReplica (if the values are unsynchronized), Mirrored (if the values are kept in sync), or Tensor (if running on a single replica).


View source

Makes an iterator for input provided via dataset.

Data from the given dataset will be distributed evenly across all the compute replicas. We will assume that the input dataset is batched by the global batch size. With this assumption, we will make a best effort to divide each batch across all the replicas (one or more workers). If this effort fails, an error will be thrown, and the user should instead use make_input_fn_iterator which provides more control to the user, and does not try to divide a batch across replicas.

The user could also use make_input_fn_iterator if they want to customize which input is fed to which replica/worker etc.

dataset that will be distributed evenly across all replicas.

An tf.distribute.InputIterator which returns inputs for each step of the computation. User should call initialize on the returned iterator.


View source

Returns an iterator split across replicas created from an input function.

The input_fn should take an tf.distribute.InputContext object where information about batching and input sharding can be accessed:

def input_fn(input_context):
  batch_size = input_context.get_per_replica_batch_size(global_batch_size)
  d =[[1.]]).repeat().batch(batch_size)
  return d.shard(input_context.num_input_pipelines,
with strategy.scope():
  iterator = strategy.make_input_fn_iterator(input_fn)
  replica_results = strategy.experimental_run(replica_fn, iterator)

The returned by input_fn should have a per-replica batch size, which may be computed using input_context.get_per_replica_batch_size.

input_fn A function taking a tf.distribute.InputContext object and returning a
replication_mode an enum value of tf.distribute.InputReplicationMode. Only PER_WORKER is supported currently, which means there will be a single call to input_fn per worker. Replicas will dequeue from the local on their worker.

An iterator object that should first be .initialize()-ed. It may then either be passed to strategy.experimental_run() or you can iterator.get_next() to get the next value to pass to strategy.extended.call_for_each_replica().


View source

Reduce value across replicas and return result on current device.

strategy = tf.distribute.MirroredStrategy(["GPU:0", "GPU:1"])
def step_fn():
  i = tf.distribute.get_replica_context().replica_id_in_sync_group
  return tf.identity(i)

per_replica_result =
total = strategy.reduce("SUM", per_replica_result, axis=None)
<tf.Tensor: shape=(), dtype=int32, numpy=1>

To see how this would look with multiple replicas, consider the same example with MirroredStrategy with 2 GPUs:

strategy = tf.distribute.MirroredStrategy(devices=["GPU:0", "GPU:1"])
def step_fn():
  i = tf.distribute.get_replica_context().replica_id_in_sync_group
  return tf.identity(i)

per_replica_result =
# Check devices on which per replica result is:
# /job:localhost/replica:0/task:0/device:GPU:0
# /job:localhost/replica:0/task:0/device:GPU:1

total = strategy.reduce("SUM", per_replica_result, axis=None)
# Check device on which reduced result is:
# /job:localhost/replica:0/task:0/device:CPU:0

This API is typically used for aggregating the results returned from different replicas, for reporting etc. For example, loss computed from different replicas can be averaged using this API before printing.

There are a number of different tf.distribute APIs for reducing values across replicas:

What should axis be?

Given a per-replica value returned by run, say a per-example loss, the batch will be divided across all the replicas. This function allows you to aggregate across replicas and optionally also across batch elements by specifying the axis parameter accordingly.

For example, if you have a global batch size of 8 and 2 replicas, values for examples [0, 1, 2, 3] will be on replica 0 and [4, 5, 6, 7] will be on replica 1. With axis=None, reduce will aggregate only across replicas, returning [0+4, 1+5, 2+6, 3+7]. This is useful when each replica is computing a scalar or some other value that doesn't have a "batch" dimension (like a gradient or loss).

strategy.reduce("sum", per_replica_result, axis=None)

Sometimes, you will want to aggregate across both the global batch and all replicas. You can get this behavior by specifying the batch dimension as the axis, typically axis=0. In this case it would return a scalar 0+1+2+3+4+5+6+7.

strategy.reduce("sum", per_replica_result, axis=0)

If there is a last partial batch, you will need to specify an axis so that the resulting shape is consistent across replicas. So if the last batch has size 6 and it is divided into [0, 1, 2, 3] and [4, 5], you would get a shape mismatch unless you specify axis=0. If you specify tf.distribute.ReduceOp.MEAN, using axis=0 will use the correct denominator of 6. Contrast this with computing reduce_mean to get a scalar value on each replica and this function to average those means, which will weigh some values 1/8 and others 1/4.

reduce_op a tf.distribute.ReduceOp value specifying how values should be combined. Allows using string representation of the enum such as "SUM", "MEAN".
value a tf.distribute.DistributedValues instance, e.g. returned by, to be combined into a single tensor. It can also be a regular tensor when used with OneDeviceStrategy or default strategy.
axis specifies the dimension to reduce along within each replica's tensor. Should typically be set to the batch dimension, or None to only reduce across replicas (e.g. if the tensor has no batch dimension).

A Tensor.


View source

Invokes fn on each replica, with the given arguments.

This method is the primary way to distribute your computation with a tf.distribute object. It invokes fn on each replica. If args or kwargs have tf.distribute.DistributedValues, such as those produced by a tf.distribute.DistributedDataset from tf.distribute.Strategy.experimental_distribute_dataset or tf.distribute.Strategy.distribute_datasets_from_function, when fn is executed on a particular replica, it will be executed with the component of tf.distribute.DistributedValues that correspond to that replica.

fn is invoked under a replica context. fn may call tf.distribute.get_replica_context() to access members such as all_reduce. Please see the module-level docstring of tf.distribute for the concept of replica context.

All arguments in args or kwargs can be a nested structure of tensors, e.g. a list of tensors, in which case args and kwargs will be passed to the fn invoked on each replica. Or args or kwargs can be tf.distribute.DistributedValues containing tensors or composite tensors, i.e. tf.compat.v1.TensorInfo.CompositeTensor, in which case each fn call will get the component of a tf.distribute.DistributedValues corresponding to its replica. Note that arbitrary Python values that are not of the types above are not supported.

Example usage:

  1. Constant tensor input.
strategy = tf.distribute.MirroredStrategy(["GPU:0", "GPU:1"])
tensor_input = tf.constant(3.0)
def replica_fn(input):
  return input*2.0
result =, args=(tensor_input,))
  0: <tf.Tensor: shape=(), dtype=float32, numpy=6.0>,
  1: <tf.Tensor: shape=(), dtype=float32, numpy=6.0>
  1. DistributedValues input.
strategy = tf.distribute.MirroredStrategy(["GPU:0", "GPU:1"])
def run():
  def value_fn(value_context):
    return value_context.num_replicas_in_sync
  distributed_values = (
  def replica_fn2(input):
    return input*2
  return, args=(distributed_values,))
result = run()
<tf.Tensor: shape=(), dtype=int32, numpy=4>
  1. Use tf.distribute.ReplicaContext to allreduce values.
strategy = tf.distribute.MirroredStrategy(["gpu:0", "gpu:1"])
def run():
   def value_fn(value_context):
     return tf.constant(value_context.replica_id_in_sync_group)
   distributed_values = (
   def replica_fn(input):
     return tf.distribute.get_replica_context().all_reduce("sum", input)
   return, args=(distributed_values,))
result = run()
  0: <tf.Tensor: shape=(), dtype=int32, numpy=1>,
  1: <tf.Tensor: shape=(), dtype=int32, numpy=1>

fn The function to run on each replica.
args Optional positional arguments to fn. Its element can be a tensor, a nested structure of tensors or a tf.distribute.DistributedValues.
kwargs Optional keyword arguments to fn. Its element can be a tensor, a nested structure of tensors or a tf.distribute.DistributedValues.
options An optional instance of tf.distribute.RunOptions specifying the options to run fn.

Merged return value of fn across replicas. The structure of the return value is the same as the return value from fn. Each element in the structure can either be tf.distribute.DistributedValues, Tensor objects, or Tensors (for example, if running on a single replica).


View source

Context manager to make the strategy current and distribute variables.

This method returns a context manager, and is used as follows:

strategy = tf.distribute.MirroredStrategy(["GPU:0", "GPU:1"])
# Variable created inside scope:
with strategy.scope():
  mirrored_variable = tf.Variable(1.)
  0: <tf.Variable 'Variable:0' shape=() dtype=float32, numpy=1.0>,
  1: <tf.Variable 'Variable/replica_1:0' shape=() dtype=float32, numpy=1.0>
# Variable created outside scope:
regular_variable = tf.Variable(1.)
<tf.Variable 'Variable:0' shape=() dtype=float32, numpy=1.0>

What happens when Strategy.scope is entered?

  • strategy is installed in the global context as the "current" strategy. Inside this scope, tf.distribute.get_strategy() will now return this strategy. Outside this scope, it returns the default no-op strategy.
  • Entering the scope also enters the "cross-replica context". See tf.distribute.StrategyExtended for an explanation on cross-replica and replica contexts.
  • Variable creation inside scope is intercepted by the strategy. Each strategy defines how it wants to affect the variable creation. Sync strategies like MirroredStrategy, TPUStrategy and MultiWorkerMiroredStrategy create variables replicated on each replica, whereas ParameterServerStrategy creates variables on the parameter servers. This is done using a custom tf.variable_creator_scope.
  • In some strategies, a default device scope may also be entered: in MultiWorkerMiroredStrategy, a default device scope of "/CPU:0" is entered on each worker.

What should be in scope and what should be outside?

There are a number of requirements on what needs to happen inside the scope. However, in places where we have information about which strategy is in use, we often enter the scope for the user, so they don't have to do it explicitly (i.e. calling those either inside or outside the scope is OK).

  • Anything that creates variables that should be distributed variables must be called in a strategy.scope. This can be accomplished either by directly calling the variable creating function within the scope context, or by relying on another API like or to automatically enter it for you. Any variable that is created outside scope will not be distributed and may have performance implications. Some common objects that create variables in TF are Models, Optimizers, Metrics. Such objects should always be initialized in the scope, and any functions that may lazily create variables (e.g., Model.__call__(), tracing a tf.function, etc.) should similarly be called within scope. Another source of variable creation can be a checkpoint restore - when variables are created lazily. Note that any variable created inside a strategy captures the strategy information. So reading and writing to these variables outside the strategy.scope can also work seamlessly, without the user having to enter the scope.
  • Some strategy APIs (such as and strategy.reduce) which require to be in a strategy's scope, enter the scope automatically, which means when using those APIs you don't need to explicitly enter the scope yourself.
  • When a tf.keras.Model is created inside a strategy.scope, the Model object captures the scope information. When high level training framework methods such as model.compile,, etc. are then called, the captured scope will be automatically entered, and the associated strategy will be used to distribute the training etc. See a detailed example in distributed keras tutorial. WARNING: Simply calling model(..) does not automatically enter the captured scope -- only high level training framework APIs support this behavior: model.compile,, model.evaluate, model.predict and can all be called inside or outside the scope.
  • The following can be either inside or outside the scope:
    • Creating the input datasets
    • Defining tf.functions that represent your training step
    • Saving APIs such as Loading creates variables, so that should go inside the scope if you want to train the model in a distributed way.
    • Checkpoint saving. As mentioned above - checkpoint.restore may sometimes need to be inside scope if it creates variables.

A context manager.


View source

Returns a copy of config_proto modified for use with this strategy.

The updated config has something needed to run a strategy, e.g. configuration to run collective ops, or device filters to improve distributed training performance.

config_proto a tf.ConfigProto object.

The updated copy of the config_proto.