fix(normalizers): cast baseMVA to float32 to support MPS backend

[emmanuelbadmus]

The MPS (Apple Silicon) backend for PyTorch does not support float64. This PR ensures baseMVA is cast to float32 in the normalizer to prevent crashes on Mac GPUs.

[Copilot]

Pull request overview

This PR aims to prevent PyTorch MPS (Apple Silicon GPU) crashes by avoiding float64 usage for the baseMVA scaling factor during normalization.

Changes:

• Updates HeteroDataMVANormalizer.transform() to store data.baseMVA as a torch.float32 tensor instead of a Python scalar.

Comments suppressed due to low confidence (1)

gridfm_graphkit/datasets/normalizers.py:244

• Casting data.baseMVA to float32 here can make inverse_transform() fail the exact equality check (data.baseMVA != self.baseMVA).any() due to float32 rounding vs self.baseMVA (often a higher-precision Python/NumPy float), causing a false mismatch error. Either keep data.baseMVA in the same precision as self.baseMVA, or cast/compare using a tolerance. Also note that the MPS float64 crash is more directly caused by self.baseMVA being a np.float64 from np.percentile(...) (see fit()); casting self.baseMVA to a Python float/float32 before the division ops would address dtype promotion at the source.

        data.baseMVA = torch.tensor(self.baseMVA, dtype=torch.float32)
        data.is_normalized = True

    def inverse_transform(self, data: HeteroData):
        if self.baseMVA is None or self.baseMVA == 0:
            raise ValueError("BaseMVA not properly set")

        if not data.is_normalized.all():
            raise ValueError("Attempting to denormalize data which is not normalized")

        if (data.baseMVA != self.baseMVA).any():
            raise ValueError(
                f"Normalizer baseMVA was {self.baseMVA} but Data object baseMVA is {data.baseMVA}",
            )

---

💡 Add Copilot custom instructions for smarter, more guided reviews. Learn how to get started.

[Copilot]

data.baseMVA is changed from a Python scalar to a torch.Tensor. This breaks LoadGridParamsFromPath.forward() (gridfm_graphkit/datasets/transforms.py:114), which assigns data.baseMVA into HeteroDataMVANormalizer.baseMVA; HeteroDataMVANormalizer.transform() then does if self.baseMVA is None or self.baseMVA == 0: and will error when self.baseMVA is a tensor. Consider keeping data.baseMVA as a Python float (or ensuring callers use .item()), and instead cast the scalar used in normalization ops to a float32 tensor on the right device/dtype.