Python multiprocessing 'RuntimeError: An attempt has been made to start a new process before the current process has finished its bootstrapping phase' (Windows/macOS spawn, missing if __name__ == '__main__')

Cause: With the spawn and forkserver start methods, each child process starts a fresh interpreter and imports your main module. If the top level of that module creates processes, a Pool or a ProcessPoolExecutor without protection, the child re-runs that code and tries to start more children, which Python stops with this RuntimeError. Spawn is the only method on Windows and has been the default on macOS since Python 3.8; from Python 3.14 the default on Linux is forkserver, so code that was fine on Linux with fork can suddenly hit this. Fix (best first): 1. Put the entry point behind a guard: define functions and classes at module level, then put the code that starts processes inside if __name__ == '__main__': (usually a main() call). 2. Do the same in scripts that use multiprocessing.Pool, ProcessPoolExecutor, or libraries that use them internally (PyTorch DataLoader with num_workers > 0, joblib, and similar). 3. Call set_start_method() at most once and only inside the guard, or avoid changing the global method by using ctx = multiprocessing.get_context("spawn") and ctx.Process / ctx.Pool. 4. Put worker functions at module level in an importable file; functions defined in a REPL or lambdas do not work with these start methods, and ProcessPoolExecutor does not work in the interactive interpreter. 5. For frozen executables (PyInstaller and similar) on Windows, call multiprocessing.freeze_support() right after the guard line. 6. Keep heavy work (loading models, opening connections) out of module top level, since it will repeat in every child. 7. Only if you truly need fork on POSIX, request it explicitly with get_context("fork"), noting it can be unsafe with threads and is deprecated-warned from 3.12 when multithreaded. Verify: the script runs from the command line without the RuntimeError, and the number of child processes equals what you requested. Sources: https://docs.python.org/3/library/multiprocessing.html https://docs.python.org/3/library/concurrent.futures.html