Lightning-AI/pytorch-lightning · error · MisconfigurationException

`precision='bf16-mixed'` does not use a scaler, found {scale

Error message

`precision='bf16-mixed'` does not use a scaler, found {scaler}.

What it means

MixedPrecisionPlugin (the AMP precision plugin) was constructed with a GradScaler while precision is 'bf16-mixed'. Bfloat16 training has a much wider dynamic range than fp16, so loss scaling is unnecessary, and Lightning rejects the scaler to prevent silently mis-scaled gradients. Pass scaler=None (or omit it) when using bf16.

Source

Thrown at src/lightning/pytorch/plugins/precision/amp.py:75

    """

    def __init__(
        self,
        precision: Literal["16-mixed", "bf16-mixed"],
        device: str,
        scaler: Optional["torch.amp.GradScaler"] = None,
    ) -> None:
        if precision not in ("16-mixed", "bf16-mixed"):
            raise ValueError(
                f"`Passed `{type(self).__name__}(precision={precision!r})`."
                f" Precision must be '16-mixed' or 'bf16-mixed'."
            )

        self.precision = precision
        if scaler is None and self.precision == "16-mixed":
            scaler = torch.amp.GradScaler(device=device)
        if scaler is not None and self.precision == "bf16-mixed":
            raise MisconfigurationException(f"`precision='bf16-mixed'` does not use a scaler, found {scaler}.")
        self.device = device
        self.scaler = scaler

    @override
    def pre_backward(self, tensor: Tensor, module: "pl.LightningModule") -> Tensor:  # type: ignore[override]
        if self.scaler is not None:
            tensor = self.scaler.scale(tensor)
        return super().pre_backward(tensor, module)

    @override
    def optimizer_step(  # type: ignore[override]
        self,
        optimizer: Optimizable,
        model: "pl.LightningModule",
        closure: Callable[[], Any],
        **kwargs: Any,
    ) -> Any:
        if self.scaler is None:

View on GitHub (pinned to 9fed5c27d2)

Solutions

1. Remove the scaler argument when using precision='bf16-mixed'
2. If you need a scaler, switch precision to '16-mixed'
3. Build the scaler conditionally: only create one when precision == '16-mixed'

Example fix

# before
plugin = MixedPrecisionPlugin(precision='bf16-mixed', scaler=torch.amp.GradScaler('cuda'))

# after
plugin = MixedPrecisionPlugin(precision='bf16-mixed', scaler=None)

from lightning.pytorch.plugins import MixedPrecisionPlugin

def make_plugin(precision, scaler=None):
    if precision == '16-mixed' and scaler is None:
        scaler = torch.amp.GradScaler('cuda')
    if precision != '16-mixed':
        scaler = None  # bf16/fp32 never take a scaler
    return MixedPrecisionPlugin(precision=precision, scaler=scaler)

def is_scaler_compatible_precision(precision: str) -> bool:
    return precision == '16-mixed'

When it happens

Trigger: Instantiating MixedPrecisionPlugin(precision='bf16-mixed', scaler=torch.amp.GradScaler(...)) or a custom plugin subclass; also when reusing an fp16 ('16-mixed') plugin config after switching Trainer(precision='bf16-mixed').

Common situations: Migrating a working 16-mixed AMP setup to bf16-mixed without removing the scaler; copy-pasted plugin configs from older Lightning versions (<2.0) where scaler+bf16 was tolerated; programmatically swapping precision strings while keeping a scaler object.

Related errors

• `precision='bf16-mixed'` does not use a scaler, found {scale
• `Passed `{type(self).__name__}(precision={precision!r})`. Pr
• AMP and the LBFGS optimizer are not compatible.
• The current optimizer, {type(optimizer).__qualname__}, does
• `precision={precision!r}` does not use a scaler, found {scal

AI-assisted analysis of Lightning-AI/pytorch-lightning@9fed5c27d2 (2026-08-28). Data as JSON: /api/errors/36f50db3ba8bbb8e. Report an issue: GitHub.