sgl-project/sglang · error · AttributeError

Subclasses of BaseScheduler must define '{attr}' property

Error message

Subclasses of BaseScheduler must define '{attr}' property

What it means

BaseScheduler.__init__ enforces that every subclass defines the properties timesteps, order, and num_train_timesteps (typically as @property). If a subclass sets none of them in __init__ and does not define the properties, the abstract-contract check raises AttributeError immediately at construction.

Source

Thrown at python/sglang/multimodal_gen/runtime/models/schedulers/base.py:21

# SPDX-License-Identifier: Apache-2.0

from abc import ABC, abstractmethod

import torch


class BaseScheduler(ABC):
    timesteps: torch.Tensor
    order: int
    num_train_timesteps: int

    def __init__(self, *args, **kwargs) -> None:
        # Check if subclass has defined all required properties
        required_attributes = ["timesteps", "order", "num_train_timesteps"]

        for attr in required_attributes:
            if not hasattr(self, attr):
                raise AttributeError(
                    f"Subclasses of BaseScheduler must define '{attr}' property"
                )

    @abstractmethod
    def set_shift(self, shift: float) -> None:
        pass

    @abstractmethod
    def set_timesteps(self, *args, **kwargs) -> None:
        pass

    @abstractmethod
    def scale_model_input(
        self, sample: torch.Tensor, timestep: int | None = None
    ) -> torch.Tensor:
        pass

View on GitHub (pinned to 0132848349)

Solutions

1. Define timesteps, order, and num_train_timesteps as attributes or @property on the subclass
2. Call super().__init__() AFTER assigning the required attributes in the subclass __init__ (the check uses hasattr on self)
3. Initialize them to None only if the property is defined — a plain missing attribute will fail; prefer explicit property definitions
4. Add a quick construction smoke test for each new scheduler

Example fix

// before
class MySched(BaseScheduler):
    def __init__(self):
        super().__init__()  # fails: attrs not set yet
        self.num_train_timesteps = 1000
// after
class MySched(BaseScheduler):
    num_train_timesteps: int = 1000
    order = 1
    def __init__(self):
        self.timesteps = None
        super().__init__()

required = {'timesteps', 'order', 'num_train_timesteps'}
assert all(hasattr(sched, a) for a in required), f'scheduler missing: {required - set(dir(sched))}'

def satisfies_base_scheduler(s: type) -> bool:
    return all(any(a in vars(c) for c in s.__mro__) for a in ('timesteps', 'order', 'num_train_timesteps'))

try:
    sched = MyScheduler(...)
except AttributeError as e:
    raise TypeError(f'{type(sched).__name__} incomplete BaseScheduler impl') from e

When it happens

Trigger: Subclassing BaseScheduler (e.g. a custom flow-match scheduler) and overriding __init__ without calling super().__init__() AFTER setting self.timesteps/etc., or simply not defining the three properties/attributes at all.

Common situations: Writing a custom scheduler for a new diffusion model; refactoring an existing scheduler so timesteps becomes a lazily-computed property and the attribute is removed; forgetting super().__init__() ordering (check runs before attributes exist if super is called first).

Related errors

• {output_batch.error}
• action policy returned no output
• Expected {request_count} outputs, got {output_count} from sc
• denoising_strength must be positive
• Must pass a value for `mu` when `use_dynamic_shifting` is Tr

AI-assisted analysis of sgl-project/sglang@0132848349 (2026-08-28). Data as JSON: /api/errors/46c8ace7024771db. Report an issue: GitHub.