int num_retvals = 1;
  TFE_Execute(matmul.get(), &retval, &num_retvals, status.get());
  ASSERT_TRUE(TF_GetCode(status.get()) == TF_OK) << TF_Message(status.get());
  TFE_DeleteTensorHandle(retval);

  // Custom device: inputs in same custom device works.
  matmul.reset(MatMulOp(context.get(), hcustom0.get(), hcustom0.get()));
  num_retvals = 1;
  executed = false;

// is done.
// TODO(agarwal): change num_retvals to int from int*.
TF_CAPI_EXPORT extern void TFE_Execute(TFE_Op* op, TFE_TensorHandle** retvals,
                                       int* num_retvals, TF_Status* status);

// Add a function (serialized FunctionDef protocol buffer) to ctx so
// that it can be invoked using TFE_Execute.
TF_CAPI_EXPORT extern void TFE_ContextAddFunctionDef(

  TFE_OpSetDevice(matmul, remote_device_name, status);
  EXPECT_EQ(TF_OK, TF_GetCode(status)) << TF_Message(status);

  TFE_TensorHandle* retvals[1];
  int num_retvals = 1;
  TFE_Execute(matmul, &retvals[0], &num_retvals, status);
  EXPECT_EQ(TF_OK, TF_GetCode(status)) << TF_Message(status);

  auto* retval_task0 =
      TFE_TensorHandleCopyToDevice(retvals[0], ctx, local_device_name, status);

  if (remote) {
    TFE_OpSetDevice(func, task1_name, status);
    ASSERT_EQ(TF_GetCode(status), TF_OK) << 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);

  //
  // Attributes are forwarded to executed operations unmodified.
  //
  // The returned optional has a value if and only if `status` evaluates to
  // TF_OK. Bad statuses are forwarded from underlying `TFE_Execute` calls, or
  // if sanity checks on dtypes/metadata fail.
  absl::optional<std::vector<std::unique_ptr<ParallelTensor>>> Execute(
      TFE_Context* context, const std::vector<ParallelTensor*>& inputs,

  if (!device_name.empty()) {
    TFE_OpSetDevice(op, device_name.c_str(), status);
  }
  if (TF_GetCode(status) != TF_OK) return nullptr;
  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);

#include "tensorflow/c/tf_status.h"
#include "tensorflow/c/tf_status_internal.h"
#include "xla/tsl/lib/core/status_test_util.h"
#include "tensorflow/core/platform/test.h"

// NOTE(allenl): These tests currently go through TFE_Execute and so are
// integration testing rather than purely testing the parallel device. They
// correspond fairly well to the implementation, but testing the C++ directly is
// another option.

namespace tensorflow {

    TFE_OpAddInput(op, h, status);
    ASSERT_TRUE(TF_GetCode(status) == TF_OK) << TF_Message(status);

    std::vector<TFE_TensorHandle*> result;
    result.push_back(nullptr);
    int num_retvals = 1;
    TFE_Execute(op, result.data(), &num_retvals, status);
    TFE_DeleteOp(op);
    ASSERT_TRUE(TF_GetCode(status) == TF_OK) << TF_Message(status);
    ASSERT_EQ(num_retvals, 1);

                    absl::optional<int64_t> step_id = absl::nullopt);
  // Block until the previous `StartExecute` operation has executed. Forwards
  // the status from `TFE_Execute` and returns outputs if the status is OK.
  std::vector<TensorHandlePtr> Join(TF_Status* status);

  // Block until all Ops finished running on the thread.
  void AsyncWait(TF_Status* status);

 private:

                                                TF_Status* status) {
  int ret = -1;
  status->status = tensorflow::unwrap(op)->OutputLength(output_name, &ret);
  return ret;
}

void TFE_Execute(TFE_Op* op, TFE_TensorHandle** retvals, int* num_retvals,
                 TF_Status* status) {
  tensorflow::ImmediateExecutionOperation* unwrapped_op =
      tensorflow::unwrap(op);

  status->status =