Help protect the Great Barrier Reef with TensorFlow on Kaggle Join Challenge


Options for cross device communications like All-reduce.

Used in the notebooks

Used in the guide

This can be passed to methods like tf.distribute.get_replica_context().all_reduce() to optimize collective operation performance. Note that these are only hints, which may or may not change the actual behavior. Some options only apply to certain strategy and are ignored by others.

One common optimization is to break gradients all-reduce into multiple packs so that weight updates can overlap with gradient all-reduce.


options = tf.distribute.experimental.CommunicationOptions(
    bytes_per_pack=50 * 1024 * 1024,
grads = tf.distribute.get_replica_context().all_reduce(
    'sum', grads, options=options)
optimizer.apply_gradients(zip(grads, vars),

bytes_per_pack a non-negative integer. Breaks collective operations into packs of certain size. If it's zero, the value is determined automatically. This only applies to all-reduce with MultiWorkerMirroredStrategy currently.
timeout_seconds a float or None, timeout in seconds. If not None, the collective raises tf.errors.DeadlineExceededError if it takes longer than this timeout. Zero disables timeout. This can be useful when debugging hanging issues. This should only be used for debugging since it creates a new thread for each collective, i.e. an overhead of timeout_seconds * num_collectives_per_second more threads. This only works for tf.distribute.experimental.MultiWorkerMirroredStrategy.
implementation a tf.distribute.experimental.CommunicationImplementation. This is a hint on the preferred communication implementation. Possible values include AUTO, RING, and NCCL. NCCL is generally more performant for GPU, but doesn't work for CPU. This only works for tf.distribute.experimental.MultiWorkerMirroredStrategy.

ValueError When arguments have invalid value.