Debian Patches

Status for pytorch-geometric/2.7.0-2

Patch Description Author Forwarded Bugs Origin Last update
improve-python3.13-compatibility.patch Python 3.13 deprecates calling `_eval_type` without `type_params`. Dhyey Mavani <ddmavani2003@gmail.com> yes
0001-Bug-DeprecationWarning-Crash-Fix-10583.patch [Bug] DeprecationWarning Crash Fix (#10583)
While working on fixing a segmentation fault in the test
```
test/datasets/test_molecule_gpt_dataset.py::test_molecule_gpt_dataset
```
Following @puririshi98’s suggestion, I created a minimal Python script,
a.py, that reproduces the same segmentation fault:
```
import sentencepiece as spm

def test_spm():
tokenizer = spm.SentencePieceProcessor()
print("SentencePieceProcessor created:", tokenizer)

if __name__ == "__main__":
test_spm()
```
When executed directly with Python, the script runs successfully
regardless of the location of `a.py`.
However, when executed with pytest, it fails if `a.py` is located in
`/opt/pyg/pytorch_geometric` or any of its subdirectories.
```
python3 a.py <<======= OK
python3 /opt/pyg/pytorch_geometric/test/datasets/a.py <<======= OK

pytest a.py <<======= OK
pytest /opt/pyg/pytorch_geometric/test/datasets/a.py <<======= SegFault
```

The rest was straightforward.

- I found that the presence of `pyproject.toml` alters this behavior;
- I traced it to a specific line in that file (thanks to @puririshi98,
again);
- the solution was obtained with [the help of
ChatGPT](https://chatgpt.com/share/69694e3f-293c-8010-9fec-86382af39186).

After applying this change, the following test runs successfully:
```
pytest /opt/pyg/pytorch_geometric/test/datasets/test_molecule_gpt_dataset.py::test_molecule_gpt_dataset
```		 Andrei Ivanov <32910461+drivanov@users.noreply.github.com> no 2026-01-15
0002-Fix-test_batch_norm-to-check-module-attributes-inste.patch Fix `test_batch_norm` to check module attributes instead of relying on string representation (#10657)

The `test_batch_norm` was recently changed in
[PR#10647](https://github.com/pyg-team/pytorch_geometric/pull/10647)
(**Fix CI failing with PyTorch 2.12+**), but it still fails with some
PyTorh 2.12 versions, and
`torch-2.12.0a0+0291f960b6.nvinternal.main.46693258` is one of them.

Current PR:
- Replaces brittle str(norm) checks with direct assertions on
num_features, eps, momentum, affine, track_running_stats, weight, and
bias.
- Removes dependence on `WITH_PT212` flag.
- Works correctly with both PyG wrapper `BatchNorm` and native
`torch.nn.BatchNorm1d`.
- Version-agnostic and future-proof across PyTorch builds.

The fix was tested with :
- torch-2.12.0a0+0291f960b6.nvinternal.main.46693258
- torch 2.10.0a0+a36e1d39eb.nv26.1.42222806		 Andrei Ivanov <32910461+drivanov@users.noreply.github.com> no 2026-04-08

All known versions for source package 'pytorch-geometric'

Links