How to use multiprocess when input is a torch.tensor with gradient?

Question

import torch.multiprocessing as mp
import torch
import time

class test():
    def __init__(self,X,Y):
        self.X=X #.share_memory_()
        self.Y=Y
        self.n=X.shape[0]
        self.result()        
    def result(self):
        self.result=torch.zeros([self.n])
        pool = mp.Pool()
        for i in range(self.n):
            pool.apply_async(self.f, (i,))
        print('start')
        pool.close()
        pool.join() 
        print('result',self.result)
        
    def f(self,i):
        time.sleep(2)
        xi=self.X[i]
        yi=self.Y[i]
        self.result[i]=xi+yi

If X do not have gradient, it works correctly.

if __name__ == '__main__':
    from temp import *
    import torch
    X=torch.tensor([1.0,2.0,3.0])
    Y=X+1
    A=test(X,Y)

It will print [3.0,5.0,7.0].

But if X has gradient

if __name__ == '__main__':
    from temp import *
    import torch
    X=torch.tensor([1.0,2.0,3.0]).requires_grad_()
    Y=X+1
    A=test(X,Y)

Then it print [0,0,0].

When X has gradient, or X is in 'cuda', it will print [0,0,0,0], which is not correct.