Segmentation error for torch==2.2.1 on MacOs

[CloseChoice]

🐛 Describe the bug

At shap, we have run into problems with our CI jobs on macOs, e.g. see here. I tracked this down to an issue with torch==2.2.1.

Here is code to reproduce the issue (this works on torch==2.2.0):

import time

import torch
from sklearn.datasets import fetch_california_housing


def test_something():
    X, y = fetch_california_housing(return_X_y=True)
    torch.tensor(X)
    time.sleep(3)

(execute with python -m pytest <filename>)

Stacktrace:

bash-3.2$ python -m pytest tests/explainers/test_segfault_minimal_example2.py                                                                                                                               
=========================================================================================== test session starts ============================================================================================
platform darwin -- Python 3.11.8, pytest-8.1.0, pluggy-1.4.0
Matplotlib: 3.8.3
Freetype: 2.6.1
rootdir: /Users/runner/work/shap/shap
configfile: pyproject.toml
plugins: cov-4.1.0, mpl-0.17.0
collected 1 item                                                                                                                                                                                           

tests/explainers/test_segfault_minimal_example2.py Fatal Python error: Segmentation fault

Thread 0x00000001140ad600 (most recent call first):
  File "/Users/runner/work/shap/shap/tests/explainers/test_segfault_minimal_example2.py", line 8 in test_something
  File "/Library/Frameworks/Python.framework/Versions/3.11/lib/python3.11/site-packages/_pytest/python.py", line 194 in pytest_pyfunc_call
  File "/Library/Frameworks/Python.framework/Versions/3.11/lib/python3.11/site-packages/pluggy/_callers.py", line 102 in _multicall
  File "/Library/Frameworks/Python.framework/Versions/3.11/lib/python3.11/site-packages/pluggy/_manager.py", line 119 in _hookexec
  File "/Library/Frameworks/Python.framework/Versions/3.11/lib/python3.11/site-packages/pluggy/_hooks.py", line 501 in __call__
  File "/Library/Frameworks/Python.framework/Versions/3.11/lib/python3.11/site-packages/_pytest/python.py", line 1769 in runtestSegmentation fault: 11

Versions

PyTorch version: 2.2.1
Is debug build: False
CUDA used to build PyTorch: None
ROCM used to build PyTorch: N/A

OS: macOS 12.7.3 (x86_64)
GCC version: Could not collect
Clang version: 14.0.0 (clang-1400.0.29.202)
CMake version: version 3.28.3
Libc version: N/A

Python version: 3.11.8 (v3.11.8:db85d51d3e, Feb  6 2024, 18:02:37) [Clang 13.0.0 (clang-1300.0.29.30)] (64-bit runtime)
Python platform: macOS-12.7.3-x86_64-i386-64bit
Is CUDA available: False
CUDA runtime version: No CUDA
CUDA_MODULE_LOADING set to: N/A
GPU models and configuration: No CUDA
Nvidia driver version: No CUDA
cuDNN version: No CUDA
HIP runtime version: N/A
MIOpen runtime version: N/A
Is XNNPACK available: True

CPU:
Intel(R) Xeon(R) CPU E5-1650 v2 @ 3.50GHz

Versions of relevant libraries:
[pip3] numpy==1.26.4
[pip3] torch==2.2.1
[pip3] torchvision==0.17.0
[conda] No relevant packages

cc @malfet @albanD @frank-wei @jgong5 @mingfeima @XiaobingSuper @sanchitintel @ashokei @jingxu10

[malfet]

Is this reproducible if one uses Apple Silicon M1 runners? (Though Torch-2.2 is the last release to support Intel Macs per #114602 )

At least I can not reproduce it on M1, trying it in x86 Rosetta mode.
Can not reproduce it in Rosetta environment either:

arch -arch x86_64 "/Applications/Python 3.11//IDLE.app/Contents/MacOS/Python" -mpytest ~/test/bug-121101.py

Nor can I repro in GitHub CI: https://github.com/malfet/deleteme/actions/runs/8150940508/job/22278030319?pr=79

[connortann]

I can reproduce in GitHub CI (over in the shap repo) with a slightly different setup:

• Update macos tests to latest pytorch shap/shap#3568

I'll see if I can identify what the relevant difference is between that job and your run above- perhaps it's related to having different dependencies installed.

[connortann]

I've done a bit of debugging with GitHub CI, and I think I have a found an environment for the minimal reproducible example above.

The test snippet above passes in an environment created with pip install pytest torch scikit-learn, but fails if the env also includes lightgbm.

The examples below ran on GitHub Actions with macos-latest, python=3.11.8, torch 2.2.1.

Passing run

Example passing run: https://github.com/shap/shap/actions/runs/8248044359/job/22557508223
Output of pip list:

Package           Version
----------------- -----------
certifi           2024.2.2
filelock          3.13.1
fsspec            2024.2.0
iniconfig         2.0.0
Jinja2            3.1.3
joblib            1.3.2
MarkupSafe        2.1.5
mpmath            1.3.0
networkx          3.2.1
numpy             1.26.4
packaging         24.0
pip               24.0
pluggy            1.4.0
pytest            8.1.1
scikit-learn      1.4.1.post1
scipy             1.12.0
setuptools        65.5.0
sympy             1.12
threadpoolctl     3.3.0
torch             2.2.1
typing_extensions 4.10.0

Failing run

Example failing run: https://github.com/shap/shap/actions/runs/8248015803/job/22557423230
Output of pip list (identical apart from lightgbm):

Package           Version
----------------- -----------
certifi           2024.2.2
filelock          3.13.1
fsspec            2024.2.0
iniconfig         2.0.0
Jinja2            3.1.3
joblib            1.3.2
lightgbm          4.3.0
MarkupSafe        2.1.5
mpmath            1.3.0
networkx          3.2.1
numpy             1.26.4
packaging         24.0
pip               24.0
pluggy            1.4.0
pytest            8.1.1
scikit-learn      1.4.1.post1
scipy             1.12.0
setuptools        65.5.0
sympy             1.12
threadpoolctl     3.3.0
torch             2.2.1
typing_extensions 4.10.0