bool needs_reduce_sum = false;
  if (!dnums.lhs.empty()) {
    needs_reduce_sum = true;
    llvm::SmallVector<int32_t> reduce_idcs(dnums.lhs.size());
    for (int64_t i = 0; i < dnums.lhs.size(); ++i) {
      reduce_idcs[i] = dnums.lhs[i];
    }

    lhs = createSumOp(lhs, lhs.getLoc(), reduce_idcs, &rewriter);
  }

  llvm::SmallVector<int32_t> lhs_transpose;
  lhs_transpose.reserve(dnums.lhs_out.size());

    DenseIntElementsAttr dimension = reduce_op.getDimensions();
    SmallVector<int64_t, 4> reduce_dims;
    for (const int64_t& dim : dimension.getValues<int64_t>()) {
      reduce_dims.emplace_back(dim);
    }
    auto dim_type = RankedTensorType::get(
        {static_cast<int64_t>(reduce_dims.size())}, rewriter.getI64Type());
    auto reduction_indices = rewriter.create<TF::ConstOp>(

static Value ApplyReduction(Location loc, Value input,
                            DenseIntElementsAttr reduce_dims,
                            OpBuilder *builder) {
  auto reduce_dims_op = builder->create<ConstantOp>(loc, reduce_dims);
  return builder->create<TF::SumOp>(loc, input, reduce_dims_op,
                                    builder->getBoolAttr(false));
}