TPUReplicateMetadata

public final class TPUReplicateMetadata

Metadata indicating how the TPU computation should be replicated.

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

Nested Classes

class

TPUReplicateMetadata.Options

Optional attributes for TPUReplicateMetadata

Public Methods

static TPUReplicateMetadata.Options	allowSoftPlacement(Boolean allowSoftPlacement)
static TPUReplicateMetadata.Options	computationShape(List<Long> computationShape)
static TPUReplicateMetadata	create(Scope scope, Long numReplicas, Options... options) Factory method to create a class wrapping a new TPUReplicateMetadata operation.
static TPUReplicateMetadata.Options	deviceAssignment(List<Long> deviceAssignment)
static TPUReplicateMetadata.Options	hostComputeCore(List<String> hostComputeCore)
static TPUReplicateMetadata.Options	numCoresPerReplica(Long numCoresPerReplica)
static TPUReplicateMetadata.Options	paddingMap(List<String> paddingMap)
static TPUReplicateMetadata.Options	stepMarkerLocation(String stepMarkerLocation)
static TPUReplicateMetadata.Options	topology(String topology)
static TPUReplicateMetadata.Options	tpuCompileOptionsProto(String tpuCompileOptionsProto)
static TPUReplicateMetadata.Options	useSpmdForXlaPartitioning(Boolean useSpmdForXlaPartitioning)
static TPUReplicateMetadata.Options	useTpu(Boolean useTpu)

Inherited Methods

From class org.tensorflow.op.PrimitiveOp

final boolean	equals(Object obj)
final int	hashCode()
Operation	op() Returns the underlying Operation
final String	toString()

From class java.lang.Object

boolean	equals(Object arg0)
final Class<?>	getClass()
int	hashCode()
final void	notify()
final void	notifyAll()
String	toString()
final void	wait(long arg0, int arg1)
final void	wait(long arg0)
final void	wait()

Public Methods

public static TPUReplicateMetadata.Options allowSoftPlacement (Boolean allowSoftPlacement)

public static TPUReplicateMetadata.Options computationShape (List<Long> computationShape)

Parameters

computationShape	DEPRECATED. Use num_cores_per_replica instead.

public static TPUReplicateMetadata create (Scope scope, Long numReplicas, Options... options)

Factory method to create a class wrapping a new TPUReplicateMetadata operation.

Parameters

scope	current scope
numReplicas	Number of replicas of the computation
options	carries optional attributes values

Returns

• a new instance of TPUReplicateMetadata

public static TPUReplicateMetadata.Options deviceAssignment (List<Long> deviceAssignment)

Parameters

deviceAssignment	The assignment of devices for the computation.

public static TPUReplicateMetadata.Options hostComputeCore (List<String> hostComputeCore)

public static TPUReplicateMetadata.Options numCoresPerReplica (Long numCoresPerReplica)

Parameters

numCoresPerReplica	Number of cores per replica. Used for model parallelism.

public static TPUReplicateMetadata.Options paddingMap (List<String> paddingMap)

public static TPUReplicateMetadata.Options stepMarkerLocation (String stepMarkerLocation)

public static TPUReplicateMetadata.Options topology (String topology)

Parameters

topology	TopologyProto indicating the topology of the TPU pod slice.

public static TPUReplicateMetadata.Options tpuCompileOptionsProto (String tpuCompileOptionsProto)

public static TPUReplicateMetadata.Options useSpmdForXlaPartitioning (Boolean useSpmdForXlaPartitioning)

public static TPUReplicateMetadata.Options useTpu (Boolean useTpu)

Parameters

useTpu	Whether to place the computation on the TPU.