A class with a collection of APIs that can be called in a replica context.

You can use tf.distribute.get_replica_context to get an instance of ReplicaContext, which can only be called inside the function passed to

strategy = tf.distribute.MirroredStrategy(['GPU:0', 'GPU:1'])
def func():
  replica_context = tf.distribute.get_replica_context()
  return replica_context.replica_id_in_sync_group
  0: <tf.Tensor: shape=(), dtype=int32, numpy=0>,
  1: <tf.Tensor: shape=(), dtype=int32, numpy=1>

strategy A tf.distribute.Strategy.
replica_id_in_sync_group An integer, a Tensor or None. Prefer an integer whenever possible to avoid issues with nested tf.function. It accepts a Tensor only to be compatible with tpu.replicate.

devices Returns the devices this replica is to be executed on, as a tuple of strings. (deprecated)

num_replicas_in_sync Returns number of replicas that are kept in sync.
replica_id_in_sync_group Returns the id of the replica.

This identifies the replica among all replicas that are kept in sync. The value of the replica id can range from 0 to tf.distribute.ReplicaContext.num_replicas_in_sync - 1.

strategy The current tf.distribute.Strategy object.



View source

All-reduces value across all replicas.

strategy = tf.distribute.MirroredStrategy(["GPU:0", "GPU:1"])
def step_fn():
  ctx = tf.distribute.get_replica_context()
  value = tf.identity(1.)
  return ctx.all_reduce(tf.distribute.ReduceOp.SUM, value)
(<tf.Tensor: shape=(), dtype=float32, numpy=2.0>,
 <tf.Tensor: shape=(), dtype=float32, numpy=2.0>)

It supports batched operations. You can pass a list of values and it attempts to batch them when possible. You can also specify options to indicate the desired batching behavior, e.g. batch the values into multiple packs so that they can better overlap with computations.

strategy = tf.distribute.MirroredStrategy(["GPU:0", "GPU:1"])
def step_fn():
  ctx = tf.distribute.get_replica_context()
  value1 = tf.identity(1.)
  value2 = tf.identity(2.)
  return ctx.all_reduce(tf.distribute.ReduceOp.SUM, [value1, value2])
  0: <tf.Tensor: shape=(), dtype=float32, numpy=2.0>,
  1: <tf.Tensor: shape=(), dtype=float32, numpy=2.0>
}, PerReplica:{
  0: <tf.Tensor: shape=(), dtype=float32, numpy=4.0>,
  1: <tf.Tensor: shape=(), dtype=float32, numpy=4.0>

Note that all replicas need to participate in the all-reduce, otherwise this operation hangs. Note that if there're multiple all-reduces, they need to execute in the same order on all replicas. Dispatching all-reduce based on conditions is usually error-prone.

This API currently can only be called in the replica context. Other variants to reduce values across replicas are:

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 nested structure of tf.Tensor which tf.nest.flatten accepts. The structure and the shapes of the tf.Tensor need to be same on all replicas.
options a tf.distribute.experimental.CommunicationOptions. Options to perform collective operations. This overrides the default options if the tf.distribute.Strategy takes one in the constructor. See tf.distribute.experimental.CommunicationOptions for details of the options.

A nested structure of tf.Tensor with the reduced values. The structure is the same as value.


View source

Merge args across replicas and run merge_fn in a cross-replica context.

This allows communication and coordination when there are multiple calls to the step_fn triggered by a call to, ...).

See for an explanation.

If not inside a distributed scope, this is equivalent to:

strategy = tf.distribute.get_strategy()
with cross-replica-context(strategy):
  return merge_fn(strategy, *args, **kwargs)

merge_fn Function that joins arguments from threads that are given as PerReplica. It accepts tf.distribute.Strategy object as the first argument.
args List or tuple with positional per-thread arguments for merge_fn.
kwargs Dict with keyword per-thread arguments for merge_fn.

The return value of merge_fn, except for PerReplica values which are unpacked.