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[] =
      "/job:worker/replica:0/task:1/device:CPU:0";
  const char new_local_device_name[] =
      "/job:worker/replica:0/task:0/device:CPU:0";

  AbstractTensorHandle* backprop;
  TF_RETURN_IF_ERROR(ops::SparseSoftmaxCrossEntropyWithLogits(
      ctx, inputs[0], inputs[1], &loss, &backprop,
      "SparseSoftmaxCrossEntropyWithLogits"));
  // `gradient_checker` only works with model that returns only 1 tensor.
  // Although, `ops::SparseSoftmaxCrossEntropyWithLogits` returns 2 tensors, the
  // second tensor isn't needed for computing gradient so we could safely drop
  // it.
  outputs[0] = loss;

  tf_gcs_filesystem::ExpiringLRUCache<int> cache(
      1, 0, [&env]() { return env->NowSeconds(); });
  env->SetNowSeconds(1);
  // Verify that replacement of an existing element works, and updates the
  // timestamp of the entry.
  cache.Insert(key, 41);
  env->SetNowSeconds(2);
  cache.Insert(key, 42);
  // 1 second after the most recent insertion, the entry is still valid.
  env->SetNowSeconds(3);

    TF_SetStatus(status, TF_OK, "");
    return n;
  };
  std::vector<char> out;
  std::unique_ptr<NowSecondsEnv> env(new NowSecondsEnv);
  // Create a cache with max staleness of 2 seconds, and verify that it works as
  // expected.
  tf_gcs_filesystem::RamFileBlockCache cache1(
      8, 16, 2 /* max staleness */, fetcher,
      [&env]() { return env->NowSeconds(); });
  // Execute the first read to load the block.

    TFE_ExecutorClearError(executor);
    TFE_ExecutorWaitForAllPendingNodes(executor, status);
    ASSERT_EQ(TF_OK, TF_GetCode(status)) << TF_Message(status);
    TFE_DeleteExecutor(executor);
  }
  // Following works in async mode since TFE_ContextAsyncClearError was called.
  TF_SetStatus(status, TF_OK, "");
  if (retvals[0] != nullptr) {
    TFE_DeleteTensorHandle(retvals[0]);
  }
  retvals[0] = nullptr;

  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;
  TFE_Execute(matmul.get(), &retval, &num_retvals, status.get());

        // each thread would cause collectives to deadlock. For consistency we
        // create a new sync executor for every thread.
        //
        // TODO(allenl): We should have an async API that works with the
        // parallel device.
        device_(device),
        executor_(
            TFE_NewExecutor(is_async, /*enable_streaming_enqueue=*/true,