ASSERT_EQ(outputs->size(), 1);
  ParallelTensor* parallel_result = (*outputs)[0].get();
  ExpectScalarEq<float>(parallel_result->tensor(0), 3.);
  ExpectScalarEq<float>(parallel_result->tensor(1), 3.);

  run_collective = parallel_device.ScalarsFromSequence<bool>(
      {true, false}, context.get(), status.get());
  parallel_device.Execute(context.get(),

  // tensor are combined into a single entry.
  const TensorSliceReader::VarToShapeMap& GetVariableToShapeMap() const;

  // Returns a map from variable names to their data types.  Slices of a
  // partitioned tensor are combined into a single entry.
  const TensorSliceReader::VarToDataTypeMap& GetVariableToDataTypeMap() const;

  // Attempts to look up the tensor named "name" and stores the found result in

TF_Tensor* BoolTensor(int32_t v);

// Create a tensor with values of type TF_INT8 provided by `values`.
TF_Tensor* Int8Tensor(const int64_t* dims, int num_dims, const char* values);

// Create a tensor with values of type TF_INT32 provided by `values`.
TF_Tensor* Int32Tensor(const int64_t* dims, int num_dims,
                       const int32_t* values);

// Create 1 dimensional tensor with values from `values`

// two component devices, running `x = TPUReplicatedInput(inputs=[a, b])` on the
// parallel device creates a parallel tensor `x` with `a` on the first of
// `underlying_devices` and `b` on the second. Running `a_unpacked, b_unpacked =
// TPUReplicatedOutput(input=x, num_replicas=2)` un-packs the parallel tensor
// into its components.
//
// The filled `device` struct and the allocated `device_info` struct may be

                                                   size_t* len);

// On success, dequeues a tensor from a TF-managed FifoQueue given by
// `tensor_id`, associated with `session`. There must be a graph node named
// "fifo_queue_dequeue_<tensor_id>", to be executed by this API call.

// Caller must call TF_DeleteTensor() over the returned tensor. If the queue is
// empty, this call is blocked.
//

                          const char* name) {
  unique_tensor_ptr tensor(BoolTensor(v), TF_DeleteTensor);
  return Const(tensor.get(), graph, s, name);
}

TF_Operation* ScalarConst(int32_t v, TF_Graph* graph, TF_Status* s,
                          const char* name) {
  unique_tensor_ptr tensor(Int32Tensor(v), TF_DeleteTensor);
  return Const(tensor.get(), graph, s, name);
}

  TensorHandlePtr size_three(
      VectorFloatTensorHandle(size_three_value, status.get()));
  ASSERT_EQ(TF_GetCode(status.get()), TF_OK) << TF_Message(status.get());

  // Try to combine these values into a single parallel tensor.
  std::array<TFE_TensorHandle*, 2> components{size_two.get(), size_three.get()};
  TensorHandlePtr combined_value = CreatePerDeviceValues(
      context.get(), components, device_name, status.get());

    (*data)[i] = i;
  }
}

// Fills out_dims with the dimensions of the given tensor.
void GetDims(const TF_Tensor* t, int64_t* out_dims) {
  int num_dims = TF_NumDims(t);
  for (int i = 0; i < num_dims; i++) {
    out_dims[i] = TF_Dim(t, i);
  }
}

// Runs model as is if output is a scalar,
// else sums the output tensor before returning.
absl::Status RunAndMaybeSum(AbstractContext* ctx, Model forward,

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

  TF_Tensor* tensor =
      TF_AllocateTensor(TF_FLOAT, dims.data(), dims.size(), sizeof(float));
  float tensor_data[] = {1};
  memcpy(TF_TensorData(tensor), tensor_data, TF_TensorByteSize(tensor));
  TFE_TensorHandle* tensor_handle = TFE_NewTensorHandle(tensor, status);
  ASSERT_EQ(TF_OK, TF_GetCode(status)) << TF_Message(status);

    // Create C++ Tensor
    Tensor src(tensorflow::DT_STRING, TensorShape(dims));
    for (int64_t i = 0; i < src.NumElements(); ++i) {
      src.flat<tstring>()(i) = data[i];
    }
    TF_Tensor* dst = TF_TensorFromTensor(src, &status);
    ASSERT_TRUE(status.ok()) << status.message();

    // Convert back to a C++ Tensor and ensure we get expected output.
    Tensor output;