Yes both dill or cloudpickle don’t have these limitations, using cloudpickle instead of dill is ok, but the problem is that we can’t easily make torch elastic use them over pickle, right? https://github.com/pytorch/pytorch/blob/a9c7e882acd94ca816db143da985febc2f20d37a/torch/multiprocessing/spawn.py#L177

import multiprocessing

def start_processes(fn, args=(), nprocs=1, join=True, daemon=False, start_method='spawn'):
    mp = multiprocessing.get_context(start_method)
    ...
    process.start()

Or do you mean using my trick but with cloudpickle?