"/job:localhost/replica:0/task:0/device:CPU:1"};
  ParallelDevice parallel_device(std::move(devices));
  std::unique_ptr<TFE_Op, decltype(&TFE_DeleteOp)> handle_op(
      TFE_NewOp(context.get(), "VarHandleOp", status.get()), TFE_DeleteOp);
  ASSERT_TRUE(TF_GetCode(status.get()) == TF_OK) << TF_Message(status.get());
  TFE_OpSetAttrType(handle_op.get(), "dtype", TF_FLOAT);

        << TF_Message(status.get());
  }

  {
    // Try to pass a non-parallel tensor to TPUReplicatedOutput
    std::unique_ptr<TFE_Op, decltype(&TFE_DeleteOp)> op(
        TFE_NewOp(context.get(), "TPUReplicatedOutput", status.get()),
        TFE_DeleteOp);
    if (TF_GetCode(status.get()) != TF_OK) return;
    TFE_OpSetAttrInt(op.get(), "num_replicas", 1);
    TFE_OpAddInput(op.get(), value_one.get(), status.get());

  CheckTFE_TensorHandleHasFloats(retval_task0, {7, 10, 15, 22});

  TFE_DeleteTensorHandle(retval_task0);
  TFE_DeleteTensorHandle(h0_task0);
  TFE_DeleteTensorHandle(retvals[0]);

  TFE_DeleteOp(matmul);

  TFE_Executor* executor = TFE_ContextGetExecutorForThread(ctx);
  TFE_ExecutorWaitForAllPendingNodes(executor, status);
  ASSERT_EQ(TF_OK, TF_GetCode(status)) << TF_Message(status);

  TFE_TensorHandle* retvals[1] = {nullptr};
  int num_retvals = 1;
  TFE_Execute(func, &retvals[0], &num_retvals, status);
  EXPECT_EQ(TF_GetCode(status), TF_OK) << TF_Message(status);
  ASSERT_EQ(1, num_retvals);
  TFE_DeleteOp(func);
  TFE_DeleteTensorHandle(packed_handle);
  TF_Tensor* t = TFE_TensorHandleResolve(retvals[0], status);
  ASSERT_EQ(TF_GetCode(status), TF_OK) << TF_Message(status);
  TFE_DeleteTensorHandle(retvals[0]);

  TFE_TensorHandle* var_handle = nullptr;
  int num_retvals = 1;
  TFE_Execute(op, &var_handle, &num_retvals, status);
  if (TF_GetCode(status) != TF_OK) return nullptr;
  TFE_DeleteOp(op);
  if (TF_GetCode(status) != TF_OK) return nullptr;
  CHECK_EQ(1, num_retvals);

  // Assign 'value' to it.
  op = TFE_NewOp(ctx, "AssignVariableOp", status);

    // on to one TFE_Op per device and just call TFE_ResetOp to avoid parsing
    // device names repeatedly.
    std::unique_ptr<TFE_Op, decltype(&TFE_DeleteOp)> const_op(
        TFE_NewOp(context, "Const", status), TFE_DeleteOp);
    if (TF_GetCode(status) != TF_OK) return nullptr;
    TFE_OpSetDevice(const_op.get(), underlying_devices_[device_index].c_str(),
                    status);

  status->status = new_op->Reset(op_or_function_name, nullptr);
  if (!status->status.ok()) {
    new_op->Release();
    new_op = nullptr;
  }
  return tensorflow::wrap(new_op);
}

void TFE_DeleteOp(TFE_Op* op) {
  if (op == nullptr) {
    return;
  }

  tensorflow::unwrap(op)->Release();
}

const char* TFE_OpGetName(const TFE_Op* op, TF_Status* status) {

#include "tsl/platform/errors.h"
#include "tsl/platform/thread_annotations.h"

namespace tensorflow {
namespace parallel_device {
namespace {

class OpDeleter {
 public:
  void operator()(TFE_Op* to_delete) const { TFE_DeleteOp(to_delete); }
};

using OpPtr = std::unique_ptr<TFE_Op, OpDeleter>;

class StatusDeleter {
 public:
  void operator()(TF_Status* to_delete) const { TF_DeleteStatus(to_delete); }
};

#include "tensorflow/c/tf_status.h"
#include "tensorflow/c/tf_status_helper.h"

namespace tensorflow {
namespace parallel_device {
namespace {

class OpDeleter {
 public:
  void operator()(TFE_Op* to_delete) const { TFE_DeleteOp(to_delete); }
};

using OpPtr = std::unique_ptr<TFE_Op, OpDeleter>;

using MaybeParallelTensorOwned =
    absl::variant<std::unique_ptr<ParallelTensor>, TensorHandlePtr>;

using MaybeParallelTensorUnowned =