uint32_t num_replicas = replicate.getN();
  llvm::SmallVector<int32_t, 4> group_assignment_val;
  for (int i = 0; i < num_replicas; ++i) group_assignment_val.push_back(i);
  Value group_assignment = builder.create<ConstOp>(
      block_arg.getLoc(),
      DenseIntElementsAttr::get(
          RankedTensorType::get({1, num_replicas}, builder.getIntegerType(32)),
          group_assignment_val));

	release <- 0
}

func put(dat rat, out *dch) {
	<-out.req
	out.dat <- dat
}

func get(in *dch) rat {
	seqno++
	in.req <- seqno
	return <-in.dat
}

// Get one rat from each of n demand channels

func getn(in []*dch) []rat {
	n := len(in)
	if n != 2 {
		panic("bad n in getn")
	}
	req := new([2]chan int)
	dat := new([2]chan rat)
	out := make([]rat, 2)
	var i int

  StringRef config = recv_op ? recv_op.getConfig() : send_op.getConfig();

  // Create XlaRecvTPUEmbeddingDeduplicationData op.
  OpBuilder builder(region);
  auto output_ty =
      RankedTensorType::get({}, VariantType::get(region->getContext()));
  auto dedup_op = builder.create<XlaRecvTPUEmbeddingDeduplicationDataOp>(
      loc, output_ty, config);

  // Rewrite RecvTPUEmbeddingActivations op to the corresponding internal op.

}

func put(dat item, out *dch) {
	<-out.req
	out.dat <- dat
}

func get(in *dch) *rat {
	seqno++
	in.req <- seqno
	return (<-in.dat).(*rat)
}

// Get one item from each of n demand channels

func getn(in []*dch) []item {
	n := len(in)
	if n != 2 {
		panic("bad n in getn")
	}
	req := make([]chan int, 2)
	dat := make([]chan item, 2)
	out := make([]item, 2)

    tf_device::ReplicateOp replicate_op,
    llvm::ArrayRef<TF::AssignVariableOp> tail_assign_variable_ops) {
  const auto num_replicas = replicate_op.getN();
  auto return_op = llvm::dyn_cast<tf_device::ReturnOp>(
      replicate_op.getRegion().front().getTerminator());

  // Get the new result types.
  // TODO(prakalps): Do not add a value to returned values if it is already
  // returned.

  for (llvm::StringRef device : device_list) {
    state_vars.push_back(builder->create<TF::VarHandleOp>(
        loc,
        llvm::ArrayRef<Type>{RankedTensorType::get(
            {}, TF::ResourceType::get(llvm::ArrayRef<TensorType>{key_type},
                                      builder->getContext()))},
        llvm::ArrayRef<Value>{},
        llvm::ArrayRef<NamedAttribute>{

// invariant. Shape ops are rewritten to be invariant when possible, prior to
// hoisting ops.
void HoistReplicateInvariantOps(tf_device::ReplicateOp replicate_op) {
  const int num_replicas = replicate_op.getN();
  Block* replicate_block = &replicate_op.GetBody();

  // A `ShapeOp` that directly depends on a `tf_device.replicate` param and does
  // not have a virtual device is assumed to return the same shape across all

      continue;

    const auto inputs = partitioned_input.getInputs();
    const bool packed_input = partitioned_input.getIsPacked();
    int num_cores_per_replica = partitioned_input.getN();
    if (num_cores_per_replica_attr) {
      num_cores_per_replica = num_cores_per_replica_attr.getInt();
    } else if (packed_input) {
      return partitioned_input->emitOpError()

    auto& arg = packed_args.emplace_back();
    arg.ssaName = std::get<0>(argAndType);
    arg.type = std::get<1>(argAndType);
  }
  if (parser.parseRegion(body, packed_args)) return failure();

  // Add derived `operandSegmentSizes` attribute based on parsed operands.
  if (!result.attributes.get(kOperandSegmentSizesAttr)) {
    int32_t num_replicated_inputs = replicated_inputs.size() * n;

                                  op)))
        return failure();

      OpBuilder builder(op);
      auto const_op = builder.create<TF::ConstOp>(
          op->getLoc(), DenseIntElementsAttr::get(
                            RankedTensorType::get({}, builder.getI64Type()),
                            {device_ordinal}));
      op->replaceAllUsesWith(const_op);
      op->erase();
      return WalkResult::advance();
    }