Lightning-AI/pytorch-lightning · error · MisconfigurationException

Schedule should return a `torch.profiler.ProfilerAction`. Fo

Error message

Schedule should return a `torch.profiler.ProfilerAction`. Found: {action}

What it means

_init_kineto probes the user schedule by calling schedule(0) and requires the return value to be a torch.profiler.ProfilerAction enum member. Lightning's ScheduleWrapper relies on ProfilerAction transitions, so a custom callable returning something else (None, string, int) fails validation immediately.

Source

Thrown at src/lightning/pytorch/profilers/pytorch.py:345

            valid_table_keys = set(inspect.signature(EventList.table).parameters.keys()) - {
                "self",
                "sort_by",
                "row_limit",
            }
            if key not in valid_table_keys:
                raise KeyError(f"Found invalid table_kwargs key: {key}. Should be within {valid_table_keys}.")

    def _init_kineto(self, profiler_kwargs: Any) -> None:
        has_schedule = "schedule" in profiler_kwargs
        self._has_on_trace_ready = "on_trace_ready" in profiler_kwargs

        schedule = profiler_kwargs.get("schedule", None)
        if schedule is not None:
            if not callable(schedule):
                raise MisconfigurationException(f"Schedule should be a callable. Found: {schedule}")
            action = schedule(0)
            if not isinstance(action, ProfilerAction):
                raise MisconfigurationException(
                    f"Schedule should return a `torch.profiler.ProfilerAction`. Found: {action}"
                )
        self._default_schedule()
        schedule = schedule if has_schedule else self._default_schedule()
        self._schedule = ScheduleWrapper(schedule) if schedule is not None else schedule
        self._profiler_kwargs["schedule"] = self._schedule

        activities = profiler_kwargs.get("activities", None)
        self._profiler_kwargs["activities"] = activities or self._default_activities()
        self._export_to_flame_graph = profiler_kwargs.get("export_to_flame_graph", False)
        self._metric = profiler_kwargs.get("metric", "self_cpu_time_total")
        with_stack = profiler_kwargs.get("with_stack", False) or self._export_to_flame_graph
        self._profiler_kwargs["with_stack"] = with_stack

    @property
    def _total_steps(self) -> Union[int, float]:
        assert self._schedule is not None
        assert self._lightning_module is not None

View on GitHub (pinned to 9fed5c27d2)

Solutions

1. Import from torch.profiler import ProfilerAction and return ProfilerAction members (NONE, WARMUP, RECORD, RECORD_AND_SAVE) for every step input
2. Handle all branches of your callable so it never returns None
3. Prefer composing torch.profiler.schedule() rather than writing a custom callable

Example fix

# before
profiler = PyTorchProfiler(schedule=lambda step: "WARMUP" if step < 3 else "RECORD")  # invalid

# after
from torch.profiler import ProfilerAction
profiler = PyTorchProfiler(
    schedule=lambda step: ProfilerAction.WARMUP if step < 3 else ProfilerAction.RECORD_AND_SAVE
)

from torch.profiler import ProfilerAction
assert schedule(0) is not None and isinstance(schedule(0), ProfilerAction)

def returns_profiler_action(schedule) -> bool:
    from torch.profiler import ProfilerAction
    try:
        return isinstance(schedule(0), ProfilerAction)
    except Exception:
        return False

When it happens

Trigger: Passing a hand-written schedule callable such as lambda step: None or a function returning a string/int instead of torch.profiler.ProfilerAction (e.g. ProfilerAction.WARMUP vs 'WARMUP').

Common situations: Writing a custom schedule without importing torch.profiler.ProfilerAction; returning the enum's name string; early-returning None in one branch of a conditional custom schedule.

Related errors

• Schedule should be a callable. Found: {schedule}
• You are trying to use `ScheduleWrapper` which require kineto
• The PyTorch Profiler default schedule will be overridden as
• You requested to find {num_devices} devices but this machine
• The FSDP strategy can only work with the `FSDPPrecision` plu

AI-assisted analysis of Lightning-AI/pytorch-lightning@9fed5c27d2 (2026-08-28). Data as JSON: /api/errors/b8b90d11f95186fb. Report an issue: GitHub.