// Create a new tensor_handle.
  TFE_TensorHandle* h0_task0_new = TestMatrixTensorHandle(ctx);

  // Check that copying it to the old remote device (named localhost) fails.
  TFE_TensorHandleCopyToDevice(h0_task0_new, ctx, remote_device_name, status);
  EXPECT_NE(TF_OK, TF_GetCode(status)) << TF_Message(status);

  // Copying and executing on the new remote device works.
  const char new_remote_device_name[] =

  // scheduled on that same threadpool, causing a deadlock in cases where the
  // caller of event_mgr->ThenExecute() blocks on the completion of the callback
  // (as in the case of ConstOp kernel creation on GPU, which involves copying a
  // CPU tensor to GPU).
  // Setting a larger thread pool does not help with the Swift caller, as we use
  // a different TFE context for each thread of execution (for running graph

  ASSERT_TRUE(executed);
  auto handle_cleaner = tensorflow::gtl::MakeCleanup(
      [var_handle]() { TFE_DeleteTensorHandle(var_handle); });

  // Assign to the variable, copying to the custom device.
  std::unique_ptr<TFE_TensorHandle, decltype(&TFE_DeleteTensorHandle)> one(
      TestScalarTensorHandle(context.get(), 111.f), TFE_DeleteTensorHandle);

// Currently this is an error, and un-packing ParallelTensors must be performed
// explicitly by running a TPUReplicatedOutput operation on the parallel device.
//
// TODO(allenl): There are some use-cases that are only supported by copying to
// host at the moment (e.g. debug print on a tensor, .numpy(), etc.). We either
// need to return something here or address these use-cases one by one.
TFE_TensorHandle* CopyTensorFromParallelDevice(TFE_Context* context,

  ASSERT_EQ(TF_GetCode(status.get()), TF_OK) << TF_Message(status.get());

  TensorHandlePtr cpu_value(FloatTensorHandle(3., status.get()));
  ASSERT_EQ(TF_GetCode(status.get()), TF_OK) << TF_Message(status.get());

  // Copying on to a parallel device must be explicit.
  TensorHandlePtr failed_copy_on_result(TFE_TensorHandleCopyToDevice(
      cpu_value.get(), context.get(), device_name, status.get()));

          temp_file_name.c_str());
  TF_SetStatus(status, TF_OK, "");
}

// TODO(vnvo2409): We could download into a local temporary file and use
// memory-mapping.
void NewReadOnlyMemoryRegionFromFile(const TF_Filesystem* filesystem,
                                     const char* path,
                                     TF_ReadOnlyMemoryRegion* region,