Windows & WSL2: zeroed CUDA tensors in spawned processes

windows
1

Dear PyTorch Team,

Please help to find the root cause of the issue described below.

Issue description

Executing the following script using the PyTorch 2.7, CUDA 12.6:

import torch
import torch.multiprocessing as mp

def producer(shared_queue, producer_done, consumer_done):
    for i in range(1, 10):
        t = torch.tensor([i, i*2], device="cuda:0")
        print(f"#{i}: Tensor to send: {t}")
        shared_queue.put(t)
    producer_done.set()
    consumer_done.wait()

def consumer(shared_queue, producer_done, consumer_done):
    producer_done.wait()
    for i in range(1, 10):
        t = shared_queue.get()
        print(f"#{i}: Tensor received: {t}")

    consumer_done.set()

if __name__ == "__main__":

    mp.set_start_method("spawn", True)

    shared_queue = torch.multiprocessing.Queue()
    consumer_done = mp.Event()
    producer_done = mp.Event()

    producer_process = mp.Process(target=producer, args=(shared_queue, producer_done, consumer_done))
    consumer_process = mp.Process(target=consumer, args=(shared_queue, producer_done, consumer_done))
    
    producer_process.start()
    consumer_process.start()   
    producer_process.join()
    consumer_process.join()

constantly results in the same output on both Windows and WSL2 (Ubuntu 24.04):

#1: Tensor to send: tensor([1, 2], device='cuda:0')
#2: Tensor to send: tensor([0, 0], device='cuda:0')
#3: Tensor to send: tensor([3, 6], device='cuda:0')
#4: Tensor to send: tensor([4, 8], device='cuda:0')
#5: Tensor to send: tensor([ 5, 10], device='cuda:0')
#6: Tensor to send: tensor([ 6, 12], device='cuda:0')
#7: Tensor to send: tensor([ 7, 14], device='cuda:0')
#8: Tensor to send: tensor([ 8, 16], device='cuda:0')
#9: Tensor to send: tensor([ 9, 18], device='cuda:0')
#1: Tensor received: tensor([0, 0], device='cuda:0')
#2: Tensor received: tensor([0, 0], device='cuda:0')
#3: Tensor received: tensor([3, 6], device='cuda:0')
#4: Tensor received: tensor([4, 8], device='cuda:0')
#5: Tensor received: tensor([ 5, 10], device='cuda:0')
#6: Tensor received: tensor([ 6, 12], device='cuda:0')
#7: Tensor received: tensor([ 7, 14], device='cuda:0')
#8: Tensor received: tensor([ 8, 16], device='cuda:0')
#9: Tensor received: tensor([ 9, 18], device='cuda:0')

The first two received tensors are zeroed, but not only received - also the second one that’s sent is zeroed as well. No issues on native Ubuntu setup.

Some findings

Thank you in advance for your assistance