diff --git a/torchao/float8/fsdp_utils.py b/torchao/float8/fsdp_utils.py index 9fde9922af..acc4333f9c 100644 --- a/torchao/float8/fsdp_utils.py +++ b/torchao/float8/fsdp_utils.py @@ -45,7 +45,7 @@ def precompute_float8_dynamic_scale_for_fsdp(module: nn.Module) -> None: isinstance(m, Float8Linear) and m.scaling_type_weight is ScalingType.DELAYED for m in module.modules() ): - raise NotImplementedError("Only supports delayed scaling") + raise NotImplementedError("Only supports dynamic scaling") float8_linears: List[Float8Linear] = [ m for m in module.modules()