TF_AbstractTensor* inputs[2] = {at, at};
  TF_OutputList* o = TF_NewOutputList();
  TF_OutputListSetNumOutputs(o, 1, status.get());
  ASSERT_EQ(TF_OK, TF_GetCode(status.get())) << TF_Message(status.get());

  // Execute.
  TF_ExecuteOperation(op, 2, inputs, o, status.get());
  ASSERT_EQ(TF_OK, TF_GetCode(status.get())) << TF_Message(status.get());

  // Clean up operation and inputs.
  TF_DeleteAbstractOp(op);

  TFE_TensorHandle* var_handle = nullptr;
  int num_retvals = 1;
  executed = false;
  TFE_Execute(op.get(), &var_handle, &num_retvals, status.get());
  ASSERT_TRUE(TF_GetCode(status.get()) == TF_OK) << TF_Message(status.get());
  ASSERT_TRUE(executed);
  auto handle_cleaner = tensorflow::gtl::MakeCleanup(
      [var_handle]() { TFE_DeleteTensorHandle(var_handle); });

    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);

// 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(

  // expected.
  tf_gcs_filesystem::RamFileBlockCache cache1(
      8, 16, 2 /* max staleness */, fetcher,
      [&env]() { return env->NowSeconds(); });
  // Execute the first read to load the block.
  TF_EXPECT_OK(ReadCache(&cache1, "", 0, 1, &out));
  EXPECT_EQ(calls, 1);
  // Now advance the clock one second at a time and redo the read. The call

      "SetAttrFunctionList has not been "
      "implemented yet.");
}
Status Execute(AbstractOperation* op_, AbstractContext* ctx,
               absl::Span<AbstractTensorHandle*> retvals, int* num_retvals,
               ForwardOperation* forward_op_, Tape* tape,
               const GradientRegistry& registry) {
  TF_RETURN_IF_ERROR(op_->Execute(retvals, num_retvals));
  for (int i = 0; i < *num_retvals; i++) {

  parallel_device.Execute(context.get(), handle_inputs, "ReadVariableOp",
                          TFE_OpGetAttrs(read_op.get()),
                          /*expected_max_outputs=*/1, status.get());
  ASSERT_FALSE(TF_GetCode(status.get()) == TF_OK);
  TF_SetStatus(status.get(), TF_OK, "");

  // Check that ops still run successfully on the device.
  parallel_device.Execute(context.get(), std::vector<ParallelTensor*>(),

  // 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,
      const char* operation_name, const TFE_OpAttrs* attributes,
      int expected_max_outputs, TF_Status* status) const;

  // 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:
  void Run();

  void Execute(TFE_Context* context, const char* operation_name,

  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);