Segmentation fault when `juliacall` is imported before `torch`

[fiktor]

Minimal example:

$ python -c 'import juliacall; import torch; print("Done.")'
Segmentation fault (core dumped)

The segfault happens at 0x00007fffa673c342 in __kmp_env_blk_init () in libgomp.so.1

Importing only one of the two modules does not result in an error.

Note that this issue is related but different from #215: here juliacall is imported before torch, and there torch is imported before juliacall.

Environment details:

• Fedora release 37 (Thirty Seven)
• Python 3.10.11 (installed using a mamba environment)
• Torch version 1.11.0.post202 (pytorch-gpu, build cuda112py310h68407e5_202)
• JuliaCall version 0.9.15 (build pypi_0)
• Julia version 1.9.4

Similar issues:

• JuliaPy.#215: Heap corruption on Python when torch is imported before juliacall, but not the reverse FAQ
• JuliaPy.#518: Segmentation Fault When Using PyJulia Inside of PyTorch Custom Autograd Function

[fiktor]

Attached are other information which might be relevant to reproduce the bug: gdb trace and the list of installed packages (in plaintext and as environment.yml renamed environment.yml.txt to pass github filetype filter).

environment.yml.txt
mamba_list.txt
trace.txt

[fiktor]

It looks like this problem was mentioned in pytorch.#78829: while the issue 78829 is about the order "import torch; import juliacall", the comment cited above says

In the Linux scenario the issue occurs when Julia is loaded before pytorch within the same process.

which is consistent with this issue. Three more observations:

• The crash happens when the following line in torch/__init__.py is executed:

198:    from torch._C import *  # noqa: F403

• Both juliacall and torch are loading the same libgomp.so.1
• There are some function names which are loaded by both juliacall and torch but exported by different libraries. Example: juliacall loads (indirectly) libopenblas64_.so and torch loads libmkl_gnu_thread.so.2. Both of these two libraries export symbol omp_in_parallel. I am not sure whether it is related to the segfault.

[mkitti]

You appear to be using juliacall which is unrelated to the pyjulia package. I'm going to transfer this issue to https://github.com/JuliaPy/PythonCall.jl

There are a number of possibilities here. You could for example configure Julia to use MKL.

[MilesCranmer]

I think this is fixed on the most recent Julia/PyTorch, please re-open otherwise.

(@mkitti I appear to not have any admin rights in this repo FYI; could you close this?)

[mkitti]

Could you link the pull request and/or commit which fixed this?

Also, you should ask @cjdoris to close this before asking me.

[MilesCranmer]

I have no idea what fixed it, it’s been hard to figure out both why it happened, and why it stopped. Maybe PyTorch isn’t exporting LLVM symbolis, or maybe Julia is more resistant to them, I’m really not sure. But I can’t reproduce it anymore so it’s as good as solved for now.

(I tagged you to close because you were in the thread already and this is an issue we’ve both looked at; I’m not sure @cjdoris is aware of it (?), and certainly wouldn’t be able to reproduce it now anyways. But yeah normally I would ask him!)

[cjdoris]

Closing by request.