// final_op = "Div"
    llvm::SmallVector<int64_t> indices(group_size.getInt());
    std::iota(indices.begin(), indices.end(), 0);

    auto replica_groups = mlir::DenseIntElementsAttr::get(
        mlir::RankedTensorType::get({1, group_size.getInt()},
                                    rewriter.getI64Type()),
        indices);

    {

            doIndexCreate = true;
            doCreate = true;
        }

        if (doCreate) {
            if (doIndexCreate) {
                try {
                    client.admin().indices().prepareCreate(settingsIndexName).setSettings(loadIndexSettings(), XContentType.JSON).execute()
                            .actionGet(getIndicesTimeout());
                } catch (final IOException e) {

  let description = [{
    Get a tensor in resource array.

    indices: the indices in the resource array.
    results: the tensor values for the corresponding indices.

    in_ch and out_ch are for control dependencies.
  }];

  let arguments = (ins
    TFRT_ChainType:$in_ch,
    StrAttr:$device,
    I64ArrayAttr:$indices
  );

  let results = (outs
    TFRT_ChainType:$out_ch,

}

// CHECK-LABEL: func @tensor_array_while_test
// CHECK-SAME: ([[in_chain:%.*]]: !tfrt.chain
func.func @tensor_array_while_test(%indices: tensor<?xi32>, %input_0: tensor<?x?x?xf32>, %input_1: tensor<?x?x?xf32>) -> (tensor<?x?x512xf32>, tensor<?x?x512xf32>) {
  %index = "tf.Const"() {device = "/device:CPU:0", value = dense<0> : tensor<i32>} : () -> (tensor<i32>)

  attributes {tf_saved_model.exported_names = ["test_hoist_varhandleop"]} {
  // CHECK-NOT: tf.VarHandleOp
  // CHECK-NOT: tf.ReadVariableOp
  // CHECK: [[v:%.*]] = "tf._TfrtGetResource"() <{container = [""], indices = [0], shared_name = [""]}> {device = "/CPU:0"} : () -> tensor<i32>
  // CHECK: [[r:%.*]] = "tf.AddV2"({{.*}}, [[v]]) {device = "/CPU:0"} : (tensor<i32>, tensor<i32>) -> tensor<i32>
  // CHECK: return [[r]]

  let summary = "get a tensor in resource array";

  let description = [{
    Get a tensor in resource array.

    indices: the indices in the resource array.
    results: the tensor values for the corresponding indices.
  }];

  let arguments = (ins
    I64ArrayAttr:$indices
  );

  let results = (outs
    Variadic<TFTensorType>:$results
  );

				// best position for error reporting
				pos := ix.Pos()
				if i < len(ix.Indices) {
					pos = ix.Indices[i].Pos()
				}
				check.softErrorf(atPos(pos), InvalidTypeArg, "%v", err)
			} else {
				check.mono.recordInstance(check.pkg, ix.Pos(), inst.TypeParams().list(), inst.TypeArgs().list(), ix.Indices)
			}
		}

		// TODO(rfindley): remove this call: we don't need to call validType here,

            List<Integer> indices = new ArrayList<>();
            List<String> tgt = target.getModules();
            Set<String> excludes = new LinkedHashSet<>(tgt);
            List<String> merged = new ArrayList<>(tgt.size() + src.size());
            merged.addAll(tgt);
            for (int i = 0, n = tgt.size(); i < n; i++) {
                indices.add(i);
            }

// the value of the variable at the time of lowering/compilation)}.
//
// The obtained PjRtBuffers are populated to `args` vector.
// `non_donatable_input_indices` will also be set, which contains the indices of
// the input that should not be donated to output.
//
// There can be three types of input: 1. Tensor with PjRtTensorBuffer; 2.
// Tensor with AsyncValueTensor; 3. Tensor with raw device mem pointer.

    llvm::StringMap<PartitionedCallDecompositionInfo>*);

// Adds the corresponding sizes of tensor list buffers in block's terminator
// to the list of return values. Returns the mapping from the buffer
// indices to the added size indices, which is a list of tuples
// (buffer_return_index, size_return_index, fixed_size).
template <class TerminatorOp>
llvm::SmallVector<std::tuple<int64_t, int64_t, bool>, 8>
AddTensorListSizesToTerminator(