test cpu-only node with multi-device dynet

Makefile

DYNET_PATH ?= /home/vlad/code/dynet
EIGEN_PATH ?= /home/vlad/code/eigen

CC = g++
DEBUG = -g
INCLUDES = -I$(DYNET_PATH) -I$(EIGEN_PATH) 
LIBS = -L$(DYNET_PATH)/build-cuda/dynet/ 
CFLAGS = -O3 -Wall -Wno-sign-compare -Wno-int-in-bool-context -c -fmessage-length=0 $(INCLUDES) -DEIGEN_FAST_MATH -fPIC -fno-finite-math-only -Wno-missing-braces -std=c++11 -funroll-loops
LFLAGS = $(LIBS) -ldynet

ifdef MKL_PATH
INCLUDES = $(INCLUDES) -I$(MKL_PATH)/include 
LIBS = $(LIBS) -L$(MKL_PATH)/lib/intel64
LFLAGS = $(LFLAGS) -lmkl_intel_lp64 -lmkl_sequential -lmkl_core;
CFLAGS = $(CFLAGS) -DEIGEN_USE_MKL_ALL
endif

all: multidev 

multidev: $(OBJS) multidev.o
	$(CC) $(OBJS) multidev.o $(LFLAGS) -o multidev

multidev.o : multidev.cc
	$(CC) $(CFLAGS) multidev.cc
clean:
	rm -f *.o *~ multidev 

multidev.cc

#include <dynet/dynet.h>
#include <dynet/expr.h>

#include <iostream>

using namespace std;
using namespace dynet;

int main(int argc, char** argv)
{
    dynet::initialize(argc, argv);
    Device* cpu_device = get_device_manager()->get_global_device("CPU");
    Device* gpu_device = get_device_manager()->get_global_device("GPU:0");


    const unsigned DIM = 5;

    ParameterCollection m;
    Parameter p_W = m.add_parameters({DIM, DIM}, gpu_device);
    Parameter p_o = m.add_parameters({DIM}, gpu_device);

    ComputationGraph cg;
    auto W = parameter(cg, p_W);
    auto o = parameter(cg, p_o);

    auto x = dynet::random_normal(cg, {DIM});
    auto y = W * x;

    y = to_device(y, cpu_device);
    auto z = sparsemax(y);
    z = to_device(z, gpu_device);

    auto l = dot_product(o, z);

    cg.forward(l);
    cout << "z: " << endl << z.value() << endl;
    cout << "l: " << l.value() << endl;
    cg.backward(l);
    cout << "grad W" << endl << W.gradient() << endl;

    return 0;
}