Class MagCacheConfig

src/diffusers/hooks/mag_cache.py:84–137

r""" Configuration for [MagCache](https://github.com/Zehong-Ma/MagCache). Args: threshold (`float`, defaults to `0.06`): The threshold for the accumulated error. If the accumulated error is below this threshold, the block computation is skipped. A higher thre

@dataclass
class MagCacheConfig:
    r"""
    Configuration for [MagCache](https://github.com/Zehong-Ma/MagCache).

    Args:
        threshold (`float`, defaults to `0.06`):
            The threshold for the accumulated error. If the accumulated error is below this threshold, the block
            computation is skipped. A higher threshold allows for more aggressive skipping (faster) but may degrade
            quality.
        max_skip_steps (`int`, defaults to `3`):
            The maximum number of consecutive steps that can be skipped (K in the paper).
        retention_ratio (`float`, defaults to `0.2`):
            The fraction of initial steps during which skipping is disabled to ensure stability. For example, if
            `num_inference_steps` is 28 and `retention_ratio` is 0.2, the first 6 steps will never be skipped.
        num_inference_steps (`int`, defaults to `28`):
            The number of inference steps used in the pipeline. This is required to interpolate `mag_ratios` correctly.
        mag_ratios (`torch.Tensor`, *optional*):
            The pre-computed magnitude ratios for the model. These are checkpoint-dependent. If not provided, you must
            set `calibrate=True` to calculate them for your specific model. For Flux models, you can use
            `diffusers.hooks.mag_cache.FLUX_MAG_RATIOS`.
        calibrate (`bool`, defaults to `False`):
            If True, enables calibration mode. In this mode, no blocks are skipped. Instead, the hook calculates the
            magnitude ratios for the current run and logs them at the end. Use this to obtain `mag_ratios` for new
            models or schedulers.
    """

    threshold: float = 0.06
    max_skip_steps: int = 3
    retention_ratio: float = 0.2
    num_inference_steps: int = 28
    mag_ratios: Optional[Union[torch.Tensor, List[float]]] = None
    calibrate: bool = False

    def __post_init__(self):
        # User MUST provide ratios OR enable calibration.
        if self.mag_ratios is None and not self.calibrate:
            raise ValueError(
                " `mag_ratios` must be provided for MagCache inference because these ratios are model-dependent.\n"
                "To get them for your model:\n"
                "1. Initialize `MagCacheConfig(calibrate=True, ...)`\n"
                "2. Run inference on your model once.\n"
                "3. Copy the printed ratios array and pass it to `mag_ratios` in the config.\n"
                "For Flux models, you can import `FLUX_MAG_RATIOS` from `diffusers.hooks.mag_cache`."
            )

        if not self.calibrate and self.mag_ratios is not None:
            if not torch.is_tensor(self.mag_ratios):
                self.mag_ratios = torch.tensor(self.mag_ratios)

            if len(self.mag_ratios) != self.num_inference_steps:
                logger.debug(
                    f"Interpolating mag_ratios from length {len(self.mag_ratios)} to {self.num_inference_steps}"
                )
                self.mag_ratios = nearest_interp(self.mag_ratios, self.num_inference_steps)


class MagCacheState(BaseState):