lllyasviel/ControlNet · error · ValueError

Invalid ema_decay value: {}

Error message

Invalid ema_decay value: {}

What it means

Raised by the EMA-tracking AdamW optimizer when ema_decay is outside [0.0, 1.0]. This custom parameter tracks an exponential moving average of the weights during training (used at inference for smoothed weights), so its decay factor must be a valid probability in the closed interval [0, 1].

Source

Thrown at ldm/util.py:107

class AdamWwithEMAandWings(optim.Optimizer):
    # credit to https://gist.github.com/crowsonkb/65f7265353f403714fce3b2595e0b298
    def __init__(self, params, lr=1.e-3, betas=(0.9, 0.999), eps=1.e-8,  # TODO: check hyperparameters before using
                 weight_decay=1.e-2, amsgrad=False, ema_decay=0.9999,   # ema decay to match previous code
                 ema_power=1., param_names=()):
        """AdamW that saves EMA versions of the parameters."""
        if not 0.0 <= lr:
            raise ValueError("Invalid learning rate: {}".format(lr))
        if not 0.0 <= eps:
            raise ValueError("Invalid epsilon value: {}".format(eps))
        if not 0.0 <= betas[0] < 1.0:
            raise ValueError("Invalid beta parameter at index 0: {}".format(betas[0]))
        if not 0.0 <= betas[1] < 1.0:
            raise ValueError("Invalid beta parameter at index 1: {}".format(betas[1]))
        if not 0.0 <= weight_decay:
            raise ValueError("Invalid weight_decay value: {}".format(weight_decay))
        if not 0.0 <= ema_decay <= 1.0:
            raise ValueError("Invalid ema_decay value: {}".format(ema_decay))
        defaults = dict(lr=lr, betas=betas, eps=eps,
                        weight_decay=weight_decay, amsgrad=amsgrad, ema_decay=ema_decay,
                        ema_power=ema_power, param_names=param_names)
        super().__init__(params, defaults)

    def __setstate__(self, state):
        super().__setstate__(state)
        for group in self.param_groups:
            group.setdefault('amsgrad', False)

    @torch.no_grad()
    def step(self, closure=None):
        """Performs a single optimization step.
        Args:
            closure (callable, optional): A closure that reevaluates the model
                and returns the loss.
        """
        loss = None

View on GitHub (pinned to ed85cd1e25)

Solutions

1. Set ema_decay to a value in [0, 1]; the constructor default is 0.9999
2. If ramping decay per step, clamp: ema_decay = min(max(decay, 0.0), 1.0)
3. Double-check you didn't pass ema_power (typically 1.0 or 3/4) into the ema_decay slot

Example fix

# before
opt = AdamW(params, ema_decay=min(1.0 + k * 0.01, ...))  # can exceed 1.0
# after
opt = AdamW(params, ema_decay=min(max(0.9999 + k * 1e-5, 0.0), 1.0))

assert 0.0 <= ema_decay <= 1.0, f'bad ema_decay: {ema_decay}'

When it happens

Trigger: Constructing AdamW(params, ema_decay=1.2) or ema_decay=-0.5, or computing ema_decay dynamically (e.g. d * scale in EMA-adjustment logic) so that it leaves the valid range.

Common situations: Custom training loops that ramp EMA decay over time and overshoot 1.0, configs mixing up ema_decay with ema_power, or ported hyperparameters from torch_ema where valid ranges differ.

Related errors

• Invalid learning rate: {}
• Invalid epsilon value: {}
• Invalid beta parameter at index 0: {}
• Invalid beta parameter at index 1: {}
• Invalid weight_decay value: {}

AI-assisted analysis of lllyasviel/ControlNet@ed85cd1e25 (2026-08-27). Data as JSON: /api/errors/b90fc61fe661bc45. Report an issue: GitHub.