Class StrategyExtended


  • Class tf.contrib.distribute.DistributionStrategyExtended
  • Class tf.distribute.StrategyExtended

Defined in tensorflow/python/distribute/

Additional APIs for algorithms that need to be distribution-aware.

The intent is that you can write an algorithm in a stylized way and it will be usable with a variety of different tf.distribute.Strategy implementations. Each descendant will implement a different strategy for distributing the algorithm across multiple devices/machines. Furthermore, these changes can be hidden inside the specific layers and other library classes that need special treatment to run in a distributed setting, so that most users' model definition code can run unchanged. The tf.distribute.Strategy API works the same way with eager and graph execution.

First let's introduce a few high-level concepts:

  • Data parallelism is where we run multiple copies of the model on different slices of the input data. This is in contrast to model parallelism where we divide up a single copy of a model across multiple devices. Note: we only support data parallelism for now, but hope to add support for model parallelism in the future.
  • A replica is one copy of the model, running on one slice of the input data.
  • Synchronous, or more commonly sync, training is where the updates from each replica are aggregated together before updating the model variables. This is in contrast to asynchronous, or async training, where each replica updates the model variables independently.
  • Furthermore you might run your computation on multiple devices on one machine (or "host"), or on multiple machines/hosts. If you are running on multiple machines, you might have a single master host that drives computation across all of them, or you might have multiple clients driving the computation asynchronously.

To distribute an algorithm, we might use some of these ingredients:

  • Parameter servers: These are hosts that hold a single copy of parameters/variables. All replicas that want to operate on a variable retrieve it at the beginning of a step and send an update to be applied at the end of the step. Can support either sync or async training.
  • Mirrored variables: These are variables that are copied to multiple devices, where we keep the copies in sync by applying the same updates to every copy. Normally would only be used with sync training.
  • Reductions and Allreduce: A reduction is some method of aggregating multiple values into one value, like "sum" or "mean". If doing sync training, we will perform a reduction on the gradients to a parameter from all replicas before applying the update. Allreduce is an algorithm for performing a reduction on values from multiple devices and making the result available on all of those devices.
  • In the future we will have support for TensorFlow's partitioned variables, where a single variable is split across multiple devices.

We have then a few approaches we want to support:

  • Code written (as if) with no knowledge of class tf.distribute.Strategy. This code should work as before, even if some of the layers, etc. used by that code are written to be distribution-aware. This is done by having a default tf.distribute.Strategy that gives ordinary behavior, and by default being in a single replica context.
  • Ordinary model code that you want to run using a specific tf.distribute.Strategy. This can be as simple as:
with my_strategy.scope():
  iterator = my_strategy.make_dataset_iterator(dataset)
  replica_train_ops = my_strategy.extended.call_for_each_replica(
      replica_fn, args=(iterator.get_next(),))
  train_op =

This takes an ordinary dataset and replica_fn and runs it distributed using a particular tf.distribute.Strategy in my_strategy. Any variables created in replica_fn are created using my_strategy's policy, and library functions called by replica_fn can use the get_replica_context() API to get enhanced behavior in this case.

  • If you want to write a distributed algorithm, you may use any of the tf.distribute.Strategy APIs inside a with my_strategy.scope(): block of code.

Lower-level concepts:

  • Wrapped values: In order to represent values parallel across devices (either replicas or the devices associated with a particular value), we wrap them in a "PerReplica" or "Mirrored" object that contains a map from device to values. "PerReplica" is used when the value may be different across replicas, and "Mirrored" when the value are the same.
  • Unwrapping and merging: Consider calling a function fn on multiple replicas, like extended.call_for_each_replica(fn, args=[w]) with an argument w that is a wrapped value. This means w will have a map taking replica device d0 to w0, replica device d1 to w1, etc. extended.call_for_each_replica() unwraps w before calling fn, so it calls fn(w0) on d0, fn(w1) on d1, etc. It then merges the return values from fn(), which can possibly result in wrapped values. For example, let's say fn() returns a tuple with three components: (x, a, v0) from replica 0, (x, b, v1) on replica 1, etc. If the first component is the same object x from every replica, then the first component of the merged result will also be x. If the second component is different (a, b, ...) from each replica, then the merged value will have a wrapped map from replica device to the different values. If the third component is the members of a mirrored variable (v maps d0 to v0, d1 to v1, etc.), then the merged result will be that mirrored variable (v).
  • Replica context vs. Cross-replica context: replica context is when we are in some function that is being called once for each replica. Otherwise we are in cross-replica context, which is useful for calling tf.distribute.Strategy methods which operate across the replicas (like reduce_to()). By default you start in a replica context (the default "single replica context") and then some methods can switch you back and forth, as described below.
  • Worker devices vs. parameter devices: Most replica computations will happen on worker devices. Since we don't yet support model parallelism, there will be one worker device per replica. When using parameter servers (see above), the set of devices holding variables may be different, otherwise the parameter devices might match the worker devices.
  • Non-slot devices are some subset of the parameter devices where we put all the non-slot variables. We need to ensure that all non-slot variables are allocated on the same device, or mirrored across the same set of devices. If you have some variable you want to colocate all the non-slot variables with, you can use colocate_vars_with() to get the remaining non-slot variables on the same device. Otherwise you can use non_slot_devices() to pick a consistent set of devices to pass to both colocate_vars_with() and update_non_slot().

When using a tf.distribute.Strategy, we have a new type dimension called locality that says what values are compatible with which APIs:

  • T: different value for each replica (e.g. a PerReplica-wrapped value).
  • M: value is "mirrored" across replicas, i.e. there are copies with the same value on each replica (e.g. a Mirrored-wrapped value).
  • V(v): value is "mirrored" across all the devices which have a copy of variable v (also a Mirrored-wrapped value, but over parameter devices instead of worker devices).
  • N: value is "mirrored" across all the "non-slot" devices

Rules for methods with respect to locality and single-replica vs. cross-replica context:

  • with d.scope(): default single-replica context -> cross-replica context for d
  • with d.extended.colocate_vars_with(v): in replica/cross-replica context, variables will be created with locality V(v). That is, if we write with d.extended.colocate_vars_with(v1): v2 = tf.get_variable(...), then v2 will have locality V(v1), i.e. locality V(v2) will equal V(v1).
  • with d.extended.colocate_vars_with(d.extended.non_slot_devices(...)): in replica/cross-replica context, variables will be created with locality N
  • v = tf.get_variable(...): in replica/cross-replica context, creates a variable (which by definition will have locality V(v), though will match another locality if inside a colocate_vars_with scope).
  • d.make_dataset_iterator(dataset) (or the deprecated d.distribute_dataset(dataset).make_one_shot_iterator()): in cross-replica context, produces an iterator with locality T
  • d.extended.broadcast_to(t): in cross-replica context, produces a value with locality M
  • d.extended.broadcast_to(t, v): in cross-replica context, produces a value with locality V(v)
  • d.extended.call_for_each_replica(fn, ...): in cross-replica context, runs fn() in a replica context (and so may call get_replica_context() and use its API, including merge_call() to get back to cross-replica context), once for each replica. May use values with locality T or M, and any variable.
  • d.extended.reduce_to(m, t, t): in cross-replica context, accepts t with locality T and produces a value with locality M.
  • d.extended.reduce_to(m, t, v): in cross-replica context, accepts t with locality T and produces a value with locality V(v).
  • d.extended.batch_reduce_to(m, [(t, v)]): seed.extended.reduce_to()`
  • d.extended.update(v, fn, ...): in cross-replica context, runs fn() once for each device v is copied to, all inputs should have locality V(v), output will have locality V(v) as well.
  • d.extended.update_non_slot(d.extended.non_slot_devices(), fn): in cross-replica context, like d.extended.update() except with locality N.
  • d.extended.read_var(v): Gets the (read-only) value of the variable v (on the device determined by the current device scope), aggregating across replicas for replica-local variables. Frequently, this will be done automatically when using v in an expression or fetching it in a cross-replica context, but this function can be used to force that conversion happens at a particular point in time (for example, to add the result of the conversion to a graph collection).

The standard pattern for updating variables is to:

  1. Create an input iterator with d.make_dataset_iterator().
  2. Define each replica d.extended.call_for_each_replica() up to the point of getting a list of gradient, variable pairs.
  3. Call d.extended.reduce_to(VariableAggregation.SUM, t, v) or d.extended.batch_reduce_to() to sum the gradients (with locality T) into values with locality V(v).
  4. Call d.extended.update(v) for each variable to update its value.

Steps 3 and 4 are done automatically by class Optimizer if you call its apply_gradients method in a replica context. Otherwise you can manually call its _distributed_apply method in a cross-replica context.

Another thing you might want to do in the middle of your replica function is an all-reduce of some intermediate value, using d.extended.reduce_to() or d.extended.batch_reduce_to(). You simply provide the same tensor as the input and destination.

Layers should expect to be called in a replica context, and can use the tf.distribute.get_replica_context function to get a tf.distribute.ReplicaContext object. The ReplicaContext object has a merge_call() method for entering cross-replica context where you can use reduce_to() (or batch_reduce_to()) and then optionally update() to update state.

You may use this API whether or not a tf.distribute.Strategy is being used, since there is a default implementation of ReplicaContext and tf.distribute.Strategy.

NOTE for new tf.distribute.Strategy implementations: Please put all logic in a subclass of tf.distribute.StrategyExtended. The only code needed for the tf.distribute.Strategy subclass is for instantiating your subclass of tf.distribute.StrategyExtended in the __init__ method.



Initialize self. See help(type(self)) for accurate signature.



Whether the strategy uses between-graph replication or not.

This is expected to return a constant value that will not be changed throughout its life cycle.



Whether initialization is needed.


Returns the tuple of all devices used to place variables.


Whether checkpointing is needed.


Whether saving summaries is needed.


Returns the tuple of all devices used to for compute replica execution.




Combine multiple reduce_to calls into one for faster execution.



A list of mirrored values, one per pair in value_destination_pairs.



Mirror a tensor on one device to all worker devices.


  • tensor: A Tensor value to broadcast.
  • destinations: A mirrored variable or device string specifying the destination devices to copy tensor to.


A value mirrored to destinations devices.



Run fn once per replica.

fn may call tf.get_replica_context() to access methods such as replica_id_in_sync_group and merge_call().

merge_call() is used to communicate between the replicas and re-enter the cross-replica context. All replicas pause their execution having encountered a merge_call() call. After that the merge_fn-function is executed. Its results are then unwrapped and given back to each replica call. After that execution resumes until fn is complete or encounters another merge_call(). Example:

# Called once in "cross-replica" context.
def merge_fn(distribution, three_plus_replica_id):
  # sum the values across replicas
  return sum(distribution.unwrap(three_plus_replica_id))

# Called once per replica in `distribution`, in a "replica" context.
def fn(three):
  replica_ctx = tf.get_replica_context()
  v = three + replica_ctx.replica_id_in_sync_group
  # Computes the sum of the `v` values across all replicas.
  s = replica_ctx.merge_call(merge_fn, args=(v,))
  return s + v

with distribution.scope():
  # in "cross-replica" context
  merged_results = distribution.call_for_each_replica(fn, args=[3])
  # merged_results has the values from every replica execution of `fn`.
  print(distribution.unwrap(merged_results))  # Prints a list


  • fn: function to run (will be run once per replica).
  • args: Tuple or list with positional arguments for fn.
  • kwargs: Dict with keyword arguments for fn.


Merged return value of fn across all replicas.



Scope that controls which devices variables will be created on.

No operations should be added to the graph inside this scope, it should only be used when creating variables (some implementations work by changing variable creation, others work by using a tf.colocate_with() scope).

This may only be used inside self.scope().

Example usage:

with strategy.scope():
  var1 = tf.get_variable(...)
  with strategy.extended.colocate_vars_with(v1):
    # var2 and var3 will be created on the same device(s) as var1
    var2 = tf.get_variable(...)
    var3 = tf.get_variable(...)

  def fn(v1, v2, v3):
    # operates on v1 from var1, v2 from var2, and v3 from var3

  # `fn` runs on every device `v1` is on, `v2` and `v3` will be there too.
  strategy.extended.update(v1, fn, args=(v2, v3))


  • colocate_with_variable: A created in self.scope(). Variables created while in the returned context manager will be on the same set of devices as colocate_with_variable.


A context manager.



Run fn with input from iterator for iterations times.

This method can be used to run a step function for training a number of times using input from a dataset.


  • fn: function to run using this distribution strategy. The function must have the following signature: def fn(context, inputs). context is an instance of MultiStepContext that will be passed when fn is run. context can be used to specify the outputs to be returned from fn by calling context.set_last_step_output. It can also be used to capture non tensor outputs by context.set_non_tensor_output. See MultiStepContext documentation for more information. inputs will have same type/structure as iterator.get_next(). Typically, fn will use call_for_each_replica method of the strategy to distribute the computation over multiple replicas.
  • iterator: Iterator of a dataset that represents the input for fn. The caller is responsible for initializing the iterator as needed.
  • iterations: (Optional) Number of iterations that fn should be run. Defaults to 1.
  • initial_loop_values: (Optional) Initial values to be passed into the loop that runs fn. Defaults to None. # TODO(priyag): Remove initial_loop_values argument when we have a mechanism to infer the outputs of fn.


Returns the MultiStepContext object which has the following properties, among other things: - run_op: An op that runs fn iterations times. - last_step_outputs: A dictionary containing tensors set using context.set_last_step_output. Evaluating this returns the value of the tensors after the last iteration. - non_tensor_outputs: A dictionatry containing anything that was set by fn by calling context.set_non_tensor_output.



Device(s) for non-slot variables.

Create variables on these devices in a with colocate_vars_with(non_slot_devices(...)): block. Update those using update_non_slot().




Reads the value of a variable.

Returns the aggregate value of a replica-local variable, or the (read-only) value of any other variable.



A tensor representing the value of v, aggregated across replicas if necessary.



Combine (via e.g. sum or mean) values across replicas.


  • reduce_op: Reduction type, an instance of tf.distribute.ReduceOp enum. DEPRECATED but still accepted values: tf.VariableAggregation.SUM, tf.VariableAggregation.MEAN,
  • value: A per-replica value with one value per replica.
  • destinations: A mirrored variable, a per-replica tensor, or a device string. The return value will be copied to all destination devices (or all the devices where the destinations value resides). To perform an all-reduction, pass value to destinations.


A value mirrored to destinations.



Run fn to update var using inputs mirrored to the same devices.

If var is mirrored across multiple devices, then this implements logic like:

results = {}
for device, v in var:
  with tf.device(device):
    # args and kwargs will be unwrapped if they are mirrored.
    results[device] = fn(v, *args, **kwargs)
return merged(results)

Otherwise this returns fn(var, *args, **kwargs) colocated with var.

Neither args nor kwargs may contain per-replica values. If they contain mirrored values, they will be unwrapped before calling fn.


  • var: Variable, possibly mirrored to multiple devices, to operate on.
  • fn: Function to call. Should take the variable as the first argument.
  • args: Tuple or list. Additional positional arguments to pass to fn().
  • kwargs: Dict with keyword arguments to pass to fn().
  • group: Boolean. Defaults to True. If False, the return value will be unwrapped.


By default, the merged return value of fn across all replicas. The merged result has dependencies to make sure that if it is evaluated at all, the side effects (updates) will happen on every replica. If instead "group=False" is specified, this function will return a nest of lists where each list has an element per replica, and the caller is responsible for ensuring all elements are executed.



Runs fn(*args, **kwargs) on colocate_with devices.


  • colocate_with: The return value of non_slot_devices().
  • fn: Function to execute.
  • args: Tuple or list. Positional arguments to pass to fn().
  • kwargs: Dict with keyword arguments to pass to fn().
  • group: Boolean. Defaults to True. If False, the return value will be unwrapped.


Return value of fn, possibly merged across devices.



Returns the container that this per-replica value belongs to.


  • value: A value returned by call_for_each_replica() or a variable created in scope().


A container that value belongs to. If value does not belong to any container (including the case of container having been destroyed), returns the value itself. value in unwrap(value_container(value)) will always be true.