RuntimeError: Cannot re-initialize CUDA in forked subprocess.

[pustar]

RuntimeError: Cannot re-initialize CUDA in forked subprocess. To use CUDA with multiprocessing, you must use the 'spawn' start method

how to solve this problem?

an answer says: torch.multiprocessing.set_start_method('spawn')
but where to add this line?

[aryamazaheri]

In this line, you can change fork to either spawn or forkserver to make the code work.

pycls/pycls/core/distributed.py

Line 144 in 0ddcc2b

mp_runner(single_proc_run, args=args, nprocs=num_proc, start_method="fork")