wheels segfault when on CUDA

Expected behavior and actual behavior.

Something is still wrong with the 1.8.5.post1 wheels: When importing from Shapely in a certain way, I later-on get segfaults when running code on the GPU.

(My guess is that somehow the way the post-release was done in https://github.com/shapely/shapely/issues/1567 is unusual and broken.)

Steps to reproduce the problem.

python -c "from shapely.geometry import Polygon; import torch; torch.randn(10).cuda()"
Segmentation fault (core dumped)

(This is the minimal example I could come up with. There are lots of real segfaults when doing other CUDA stuff with this version of Shapely.)

It goes away with importing all of Shapely.

It also goes away when installing 1.8.5.post1 from source (either pip install --no-binary shapely --force-reinstall shapely or pip install --force-reinstall git+https://github.com/shapely/shapely@maint-1.8).

Operating system

This is Linux with Python 3.8. (I tried different versions of Pytorch and CUDA toolkit – always the same.)

Shapely version and provenance

1.8.5.post1 wheel as published on PyPI

[jorisvandenbossche]

(BTW, import shapely of course does not do anything, because it’s just the namespace pkg on 1.8.x. But from shapely import * works without crashing. So that’s the workaround I actually meant.)

What I think is happening here is that from shapely import * is also not doing anything (for the same reason). So the reason that this “workaround” seems to work, is also because it just not yet imports anything, just like import shapely (and in contrast with import shapely.geometry).

So the issue is still that in general importing actual shapely functionality before importing torch gives segfaults.

[bertsky]

I certainly understand you cannot require that right now (certainly given we are still in betal), but I would certainly recommend in already trying it out, to ensure we find those other issues and can fix them.

Yes, the problem is that (in my experience) most errors only happen rarely, as they depend on so much from the data side. As soon as we have our data regression test ready, we’ll include prereleases in CI to get early warnings.