PyTorch 2.6.0: Weights only load failed

[hongbo-miao]

When upgrade from PyTorch 2.5.1 to 2.6.0, for code like

from ogb.graphproppred import PygGraphPropPredDataset

dataset = PygGraphPropPredDataset(name = 'ogbg-molhiv')

I got error:

Traceback (most recent call last):
  File "/home/runner/work/hongbomiao.com/hongbomiao.com/machine-learning/graph-neural-network/src/main.py", line 103, in main
    dataset, split_idx = fetch_dataset(config)
                         ^^^^^^^^^^^^^^^^^^^^^
  File "/home/runner/work/hongbomiao.com/hongbomiao.com/machine-learning/graph-neural-network/src/model/data_loader.py", line 13, in fetch_dataset
    dataset = PygGraphPropPredDataset(name=config.dataset)
              ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
  File "/home/runner/work/hongbomiao.com/hongbomiao.com/machine-learning/graph-neural-network/.venv/lib/python3.12/site-packages/ogb/graphproppred/dataset_pyg.py", line 68, in __init__
    self.data, self.slices = torch.load(self.processed_paths[0])
                             ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
  File "/home/runner/work/hongbomiao.com/hongbomiao.com/machine-learning/graph-neural-network/.venv/lib/python3.12/site-packages/torch/serialization.py", line 1470, in load
    raise pickle.UnpicklingError(_get_wo_message(str(e))) from None
_pickle.UnpicklingError: Weights only load failed. This file can still be loaded, to do so you have two options, do those steps only if you trust the source of the checkpoint. 
	(1) In PyTorch 2.6, we changed the default value of the `weights_only` argument in `torch.load` from `False` to `True`. Re-running `torch.load` with `weights_only` set to `False` will likely succeed, but it can result in arbitrary code execution. Do it only if you got the file from a trusted source.
	(2) Alternatively, to load with `weights_only=True` please check the recommended steps in the following error message.
	WeightsUnpickler error: Unsupported global: GLOBAL torch_geometric.data.data.DataEdgeAttr was not an allowed global by default. Please use `torch.serialization.add_safe_globals([DataEdgeAttr])` or the `torch.serialization.safe_globals([DataEdgeAttr])` context manager to allowlist this global if you trust this class/function.

I am wondering if the library should handle this issue in a future version, or if we should add something like (not tested codes):

from ogb.graphproppred import PygGraphPropPredDataset
from torch_geometric.data.data import DataEdgeAttr

torch.serialization.add_safe_globals([DataEdgeAttr])
dataset = PygGraphPropPredDataset(name = 'ogbg-molhiv')

What would be the recommended approach? Thank you! ☺️

[alexbarghi-nv]

I'm having this problem too.

[EMZEDI]

Having the same issue. This needs to be fixed, though a workaround for now could be the following:

import torch
from torch_geometric.data.storage import GlobalStorage
from torch_geometric.data.data import DataEdgeAttr, DataTensorAttr

torch.serialization.add_safe_globals([GlobalStorage, DataEdgeAttr, DataTensorAttr])

# the rest of the code

• It made it work for me.

[benjaminbuzek]

Having the same issue. This needs to be fixed, though a workaround for now could be the following:

import torch
from torch_geometric.data.storage import GlobalStorage
from torch_geometric.data.data import DataEdgeAttr, DataTensorAttr

torch.serialization.add_safe_globals([GlobalStorage, DataEdgeAttr, DataTensorAttr])

the rest of the code

• It made it work for me.

I ran into the same issue with the version >=1.3.6 and can confirm that this resolved the issue for me.