if (subtypes.size() != 1) {
    return std::nullopt;
  }
  return subtypes[0].getElementType();
}

// Create an `ConstBytesAttr` which encodes the options
// for the `tf.custom` tensor list op to be created. If the given
// op is not a `tf.TensorList*` op, return empty, although this case
// should never be trigged in practice since patterns are only applied
// on `tf.TensorList*` ops.

        xla_tensor->AllocateShapedBuffer(device_tensor->dtype(), shape, client_,
                                         stream_->parent()->device_ordinal()));

    // The cpu_tensor and literal that we created here hold the data of host
    // tensor in descending layout. The layout could be different from layout in
    // device_tensor (but the logical shape has to be the same). The

    TF_DeleteStatus(status);
  }

  // 6. Read `var` using `ctx_1`. This read should succeed since `ctx_1` was
  // created with `isolate_session_state` set to false, and update should
  // preserve it.
  {
    // Create a handle to `var`, using `ctx_1`.
    TFE_TensorHandle* var_handle =
        CreateVarHandle(ctx_1, remote_device, /*variable_name=*/"var");

        builder, block.getParent()->getLoc(), execute, device);

    builder->create<tf_device::ReturnOp>(block.getParent()->getLoc(),
                                               block_launch_op.getResults());
  }

  return success();
}

// Creates a `tf.TPUCompileSucceededAssert` operation that parses compilation
// status of `compile_op` to check whether compilation is successful.

    ModuleOp module) {
  SymbolTableCollection table;
  SymbolUserMap symbol_map(table, module);
  llvm::DenseMap<func::FuncOp, llvm::DenseSet<func::FuncOp>> caller_callee_map;
  // Creates work queue for determining reachability below.
  std::queue<func::FuncOp> function_worklist;

  for (auto func : module.getOps<func::FuncOp>()) {
    for (auto user : symbol_map.getUsers(func)) {

  auto reshape_sizes = DenseIntElementsAttr::get(reshape_type, new_shape);
  auto reshape_value =
      builder->create<TF::ConstOp>(input.getLoc(), reshape_sizes);
  return builder->create<TF::ReshapeOp>(input.getLoc(), reshape_result_type,
                                        input, reshape_value);
}

// Creates transpose op for shape to depth transform.
TF::TransposeOp GetTransposeOpForConv2DFilter(OpBuilder* builder, Value input) {

      } else {
        if (IsCompatibleTypeWithTFLCastOp(value.getType()) &&
            IsCompatibleTypeWithTFLCastOp(type)) {
          auto cast = b.create<CastOp>(yield_op->getLoc(), type, value);
          args.push_back(cast);
        } else {
          auto cast = b.create<TF::CastOp>(yield_op->getLoc(), type, value);
          args.push_back(cast);
        }
      }
    }
    args.append(new_args.begin(), new_args.end());

}

// Creates a wrapper island enclosing the `sub_op` dependent on
// `control_inputs`.
IslandOp CreateIsland(Operation* sub_op, ValueRange control_inputs,
                      OpBuilder builder) {
  assert(sub_op);
  auto control_type = ControlType::get(builder.getContext());
  auto island = builder.create<IslandOp>(

    // Calculate conjugate of X
    std::string name = "Conj_Log1p_Grad_X";
    TF_RETURN_IF_ERROR(SafeConj(ctx, X, &temp_output, name.c_str()));

    AbstractTensorHandlePtr Conj_X(temp_output);

    // Creates Ones
    name = "OnesLike_Log1p_Grad_X";
    TF_RETURN_IF_ERROR(OnesLike(ctx, Conj_X.get(), &temp_output, name.c_str()));

    AbstractTensorHandlePtr Ones_X(temp_output);

    name = "Add_Log1p_Grad_X";

        auto s_qconst =
            builder.create<SparseQConstOp>(op->getLoc(), cst.getQtypeAttr(),
                                           cst.getValue(), s_param, new_value);
        value.replaceAllUsesWith(s_qconst.getResult());
        cst.erase();
      }

      if (result.needs_densify) {
        auto value = op->getOperand(operand);
        auto densify =
            builder.create<DensifyOp>(op->getLoc(), value.getType(), value);