Cuda same kernel, but different results with __constant__

Question

How can cudaMemcpyToSymbol just make this ??

// head.h
#include <stdio.h>
__constant__ float const_mem[1];

__global__ void k0();   //I will declare it in main.cu
__global__ void k1();   //I will declare it in separate.cu

//separate.cu
#include "head.h"

__global__ void k0() {
     printf("%f\n", const_mem[0]);
}

//main.cu
#include "head.h"

__global__ void k1() {
     printf("%f\n", const_mem[0]);
}

int main() {
     float arr[] = {5};
     cudaMemcpyToSymbol(const_mem, arr, sizeof(float));
     k0<<<1,1>>>();
     k1<<<1,1>>>();
}

Compilation : nvcc main.cu separate.cu

output of sudo nvprof ./a.out (./a.out gives litteraly nothing)

0.000000
5.000000

That mean that kernel writed in an other transition unit is not accessing const_memory ... but how is it possible ??