Lightning-AI/pytorch-lightning · error · ValueError

The start method '{self._start_method}' is not available on

Error message

The start method '{self._start_method}' is not available on this platform. Available methods are: {', '.join(mp.get_all_start_methods())}

What it means

The _MultiProcessingLauncher validates its start_method argument against multiprocessing.get_all_start_methods() at construction time. Not all platforms support all start methods: 'forkserver' and 'fork' are Unix-only, and 'fork' is unavailable on macOS spawn-default setups that restrict it. Passing an unsupported method raises this ValueError immediately.

Source

Thrown at src/lightning/fabric/strategies/launchers/multiprocessing.py:71

    Args:
        strategy: A reference to the strategy that is used together with this launcher.
        start_method: The method how to start the processes.
            - 'spawn': The default start method. Requires all objects to be pickleable.
            - 'fork': Preferable for IPython/Jupyter environments where 'spawn' is not available. Not available on
              the Windows platform for example.
            - 'forkserver': Alternative implementation to 'fork'.

    """

    def __init__(
        self,
        strategy: "ParallelStrategy",
        start_method: Literal["spawn", "fork", "forkserver"] = "spawn",
    ) -> None:
        self._strategy = strategy
        self._start_method = start_method
        if start_method not in mp.get_all_start_methods():
            raise ValueError(
                f"The start method '{self._start_method}' is not available on this platform. Available methods are:"
                f" {', '.join(mp.get_all_start_methods())}"
            )

    @property
    @override
    def is_interactive_compatible(self) -> bool:
        # The start method 'spawn' is not supported in interactive environments
        # The start method 'fork' is the only one supported in Jupyter environments, with constraints around CUDA
        # initialization. For more context, see https://github.com/Lightning-AI/pytorch-lightning/issues/7550
        return self._start_method == "fork"

    @override
    def launch(self, function: Callable, *args: Any, **kwargs: Any) -> Any:
        """Launches processes that run the given function in parallel.

        The function is allowed to have a return value. However, when all processes join, only the return value
        of worker process 0 gets returned from this `launch` method in the main process.

View on GitHub (pinned to 9fed5c27d2)

Solutions

1. Switch to start_method='spawn', which is available on all platforms
2. Remove the explicit start_method argument to use the default ('spawn')
3. On Windows, accept that only 'spawn' is generally available and adapt the code to be spawn-safe (guard main, picklable objects)

Example fix

# before
launcher = _MultiProcessingLauncher(strategy, start_method="forkserver")

# after
import multiprocessing as mp
method = "forkserver" if "forkserver" in mp.get_all_start_methods() else "spawn"
launcher = _MultiProcessingLauncher(strategy, start_method=method)

import multiprocessing as mp
allowed = mp.get_all_start_methods()
if start_method not in allowed:
    start_method = "spawn"  # universally available fallback

def safe_start_method(m: str) -> Literal["spawn", "fork", "forkserver"]:
    import multiprocessing as mp
    return m if m in mp.get_all_start_methods() else "spawn"  # type: ignore[return-value]

When it happens

Trigger: Creating a strategy/launcher with launcher('forkserver') or _MultiProcessingLauncher(strategy, start_method='forkserver') on Windows, or a start method filtered out by the platform's mp.get_all_start_methods().

Common situations: Running a script developed on Linux under Windows; explicitly setting start_method in Fabric/strategy kwargs that is not supported by the OS; CI matrix runs on Windows hitting Unix-only start methods.

Related errors

• The start method '{self._start_method}' is not available on
• Cannot re-initialize CUDA in forked subprocess. To use CUDA
• Lightning can't create new processes if CUDA is already init
• Launching multiple processes with the 'spawn' start method r
• You selected `Trainer(strategy='{strategy_flag}')` but proce

AI-assisted analysis of Lightning-AI/pytorch-lightning@9fed5c27d2 (2026-08-28). Data as JSON: /api/errors/022ca51dd8dff2ec. Report an issue: GitHub.