Lightning-AI/pytorch-lightning · error · ValueError

`precision={precision!r}` does not use a scaler, found {scal

Error message

`precision={precision!r}` does not use a scaler, found {scaler}.

What it means

FSDPMixedPrecisionPlugin received an explicit ShardedGradScaler while precision is anything other than '16-mixed'. Only fp16 mixed precision needs loss scaling; bf16/fp32 modes must be constructed without a scaler, so the plugin raises to prevent a mis-scaled FSDP run.

Source

Thrown at src/lightning/pytorch/plugins/precision/fsdp.py:64

    Raises:
        ValueError:
            If unsupported ``precision`` is provided.

    """

    def __init__(self, precision: _PRECISION_INPUT, scaler: Optional["ShardedGradScaler"] = None) -> None:
        supported_precision = get_args(_PRECISION_INPUT)
        if precision not in supported_precision:
            raise ValueError(
                f"`precision={precision!r})` is not supported in FSDP."
                f" `precision` must be one of: {supported_precision}."
            )

        from torch.distributed.fsdp.sharded_grad_scaler import ShardedGradScaler

        if scaler is not None and self.precision != "16-mixed":
            raise ValueError(f"`precision={precision!r}` does not use a scaler, found {scaler}.")

        self.scaler = ShardedGradScaler() if scaler is None and precision == "16-mixed" else None
        self.precision = precision

        precision_to_type = {
            "bf16-mixed": torch.float32,
            "16-mixed": torch.float32,
            "bf16-true": torch.bfloat16,
            "16-true": torch.float16,
            "32-true": torch.float32,
        }
        self._desired_input_dtype = precision_to_type[self.precision]

    @override
    def convert_module(self, module: Module) -> Module:
        if "true" in self.precision:
            return module.to(dtype=self._desired_input_dtype)
        return module

View on GitHub (pinned to 9fed5c27d2)

Solutions

1. Pass scaler=None (or omit it) when precision is not '16-mixed'
2. Keep the scaler only for precision='16-mixed'
3. Construct the scaler conditionally based on the precision string

Example fix

# before
plugin = FSDPMixedPrecisionPlugin(precision='bf16-mixed', scaler=ShardedGradScaler())

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

def make_fsdp_plugin(precision, scaler=None):
    if precision != '16-mixed':
        scaler = None  # only fp16 mixed precision uses ShardedGradScaler
    return FSDPMixedPrecisionPlugin(precision=precision, scaler=scaler)

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

When it happens

Trigger: FSDPMixedPrecisionPlugin(precision='bf16-mixed', scaler=ShardedGradScaler()) or precision='32-true' with a scaler; any non-'16-mixed' precision combined with a non-None scaler argument.

Common situations: Switching an FSDP fp16 recipe to bf16-mixed but keeping the ShardedGradScaler in the config; templated plugin construction that always passes a scaler; older examples that shipped with explicit scalers.

Related errors

• `precision='bf16-mixed'` does not use a scaler, found {scale
• Was unable to infer precision type, received {self.precision
• The optimizer has references to the model's meta-device para
• `precision={precision!r})` is not supported in FSDP. `precis
• `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/183f59b057486f63. Report an issue: GitHub.