How to get gpu data from torch.multiprocessing.Queue?

Question

Here is my code:

import torch
import torch.multiprocessing as mp

mp.set_start_method('spawn', force=True)

def job(device, q, event):
    x = torch.ByteTensor([1,9,5]).to(device)
    x.share_memory_()
    print("in job:", x)
    q.put(x)
    event.wait()

       
def main():
    device = torch.device("cuda" if torch.cuda.is_available else "cpu")
 
    num_processes = 4
    processes = []
    q = mp.SimpleQueue()
    event = mp.Event()
   
    for rank in range(num_processes):
        p = mp.Process(target=job, args=(device, q, event))
        p.start()
        processes.append(p)

    for i in range(num_processes):
        y = q.get()
        print("in main:", y)
        del y 
    event.set()
    
    for p in processes:
        p.join()
    
if __name__ =="__main__":
    main()

Got the outputs:

in job: tensor([1, 9, 5], device='cuda:0', dtype=torch.uint8)
in job: tensor([1, 9, 5], device='cuda:0', dtype=torch.uint8)
in job: tensor([1, 9, 5], device='cuda:0', dtype=torch.uint8)
in job: tensor([1, 9, 5], device='cuda:0', dtype=torch.uint8)
in main: tensor([0, 0, 0], device='cuda:0', dtype=torch.uint8)
in main: tensor([0, 0, 0], device='cuda:0', dtype=torch.uint8)
in main: tensor([0, 0, 0], device='cuda:0', dtype=torch.uint8)
in main: tensor([0, 0, 0], device='cuda:0', dtype=torch.uint8)

When tensor is on cpu, good.But when I send it to CUDA, I can't read the tensor content from main process.What triggers gpu data lost through queue? I think it could be some mechanism of multiprocessing memory space but Im not quite familiar with it.