tf.raw_ops.TPUReplicateMetadata

Metadata indicating how the TPU computation should be replicated.

tf.raw_ops.TPUReplicateMetadata(
    num_replicas, num_cores_per_replica=1, topology='', use_tpu=True,
    device_assignment=[], computation_shape=[], host_compute_core=[],
    padding_map=[], step_marker_location='STEP_MARK_AT_ENTRY',
    allow_soft_placement=False, name=None
)

This operation holds the metadata common to operations of a tpu.replicate() computation subgraph.

Args:

  • num_replicas: An int that is >= 0. Number of replicas of the computation
  • num_cores_per_replica: An optional int. Defaults to 1. Number of cores per replica. Used for model parallelism.
  • topology: An optional string. Defaults to "". TopologyProto indicating the topology of the TPU pod slice.
  • use_tpu: An optional bool. Defaults to True. Whether to place the computation on the TPU.
  • device_assignment: An optional list of ints. Defaults to []. The assignment of devices for the computation.
  • computation_shape: An optional list of ints. Defaults to []. DEPRECATED. Use num_cores_per_replica instead.
  • host_compute_core: An optional list of strings. Defaults to [].
  • padding_map: An optional list of strings. Defaults to [].
  • step_marker_location: An optional string. Defaults to "STEP_MARK_AT_ENTRY".
  • allow_soft_placement: An optional bool. Defaults to False.
  • name: A name for the operation (optional).

Returns:

The created Operation.