microsoft/qlib · error · RuntimeError

You can trying to get a state from a dead environment wrappe

Error message

You can trying to get a state from a dead environment wrapper.

What it means

RuntimeError in `EnvWrapper.reset` (qlib/rl/utils/env_wrapper.py:154). The wrapper pulls the next initial state from `seed_iterator`; once the iterator is exhausted it is set to None, marking the wrapper dead (it should be recycled by the vector env). A reset on such a dead wrapper raises immediately instead of silently reusing stale state.

Source

Thrown at qlib/rl/utils/env_wrapper.py:154

        self.status: EnvWrapperStatus = cast(EnvWrapperStatus, None)

    @property
    def action_space(self) -> Space:
        return self.action_interpreter.action_space

    @property
    def observation_space(self) -> Space:
        return self.state_interpreter.observation_space

    def reset(self, **kwargs: Any) -> ObsType:
        """
        Try to get a state from state queue, and init the simulator with this state.
        If the queue is exhausted, generate an invalid (nan) observation.
        """

        try:
            if self.seed_iterator is None:
                raise RuntimeError("You can trying to get a state from a dead environment wrapper.")

            # TODO: simulator/observation might need seed to prefetch something
            # as only seed has the ability to do the work beforehands

            # NOTE: though logger is reset here, logs in this function won't work,
            # because we can't send them outside.
            # See https://github.com/thu-ml/tianshou/issues/605
            self.logger.reset()

            if self.seed_iterator is SEED_INTERATOR_MISSING:
                # no initial state
                initial_state = None
                self.simulator = cast(Callable[[], Simulator], self.simulator_fn)()
            else:
                initial_state = next(cast(Iterator[InitialStateType], self.seed_iterator))
                self.simulator = self.simulator_fn(initial_state)

            self.status = EnvWrapperStatus(

View on GitHub (pinned to 79633dd950)

Solutions

Let the vector environment (e.g. qlib's BaseVectorEnv / tianshou runner) own wrapper recycling; never reset wrappers yourself after exhaustion.
Ensure the seed iterator yields at least as many initial states as resets requested in the phase (train loop steps; FiniteVectorEnv sizes val/test to the iterator).
In custom loops, track exhausted wrappers (reset returns a NaN observation when the iterator merely ends) and only treat the RuntimeError as a lifecycle bug to fix upstream.

Example fix

// before
for _ in range(n):
    obs = wrapper.reset()  # n > number of seeds -> dead wrapper
// after
while True:
    obs = wrapper.reset()
    if is_invalid(obs):  # queue exhausted: stop instead of resetting again
        break

Defensive patterns

Strategy: try-catch

Validate before calling

def safe_reset(wrapper):
    """Returns None when the wrapper is dead and should be recycled."""
    try:
        return wrapper.reset()
    except RuntimeError as e:
        if "dead environment wrapper" in str(e):
            return None
        raise

Type guard

def wrapper_is_alive(wrapper) -> bool:
    return wrapper.seed_iterator is not None

Try / catch

try:
    obs = wrapper.reset()
except RuntimeError as e:
    if "dead environment wrapper" in str(e):
        log.debug("wrapper exhausted; recycling")
    else:
        raise

Prevention

Let the vector env own wrapper recycling; don't hold or reset wrappers manually.
Size seed iterators to the number of resets your loop performs.
Treat exhausted wrappers (NaN obs) as terminal, not as an error to retry.

When it happens

Trigger: Calling `env_wrapper.reset()` after its seed iterator raised StopIteration (i.e., after the finite seed set was consumed); manually driving an EnvWrapper outside a proper vector-env lifecycle; holding references to wrappers that the vector env already retired.

Common situations: Custom training loops that manage envs themselves and lose track of which wrappers are exhausted; seed iterators yielding fewer states than the number of resets requested; async/stale references after a NestedExecutor restarts environments.

Related errors

AI-assisted analysis of microsoft/qlib@79633dd950 (2026-08-15). Data as JSON: /api/errors/ccd3534c911057e8. Report an issue: GitHub.