CUDA multiprocessing: training multiple model in different processes in single GPU

[visoft (Cristi Vicas)]

Hi! I had similar issue. After some googling I found this:

A call to torch.cuda.is_available makes an unrelated multi-processing computation crash?

Looks like set_start_method did not work for me but mp = mp.get_context('spawn') did. Hope that provides some help.

2nd link showed the solution:

def some_method():
    mp = torch.multiprocessing.get_context('forkserver')  # <-- This does the magic
    pool = mp.Pool(processes=1)  
    ........

Hope it helps!