Importing xgboost before torch + openmp causes seg fault

[ikrommyd]

I'm on apple sillicon in a fresh environment that I just created.

uv venv --python 3.12
source .venv/bin/activate
uv pip install xgboost torch

Running the following seg faults:

import xgboost
import torch

M = torch.tensor([[2.0, 0.5],
                  [0.5, 1.0]],
                 dtype=torch.double, device="cpu")

L = torch.linalg.cholesky(M)

print("Cholesky succeeded:", L)

Switching the import order (torch before xgboost) fixes the problem. Also doing export OMP_NUM_THREADS=1 fixes the problem.

From my quick debugging session I think it's that xgboost loads libomp from homebrew first while then torch loads it's own version. I don't know if torch or xgboost should do something differently here, I'm just opening the issue to raise awareness.

[ikrommyd]

-> otool -L .venv/lib/python3.12/site-packages/xgboost/lib/libxgboost.dylib | grep libomp
        @rpath/libomp.dylib (compatibility version 5.0.0, current version 5.0.0)

-> otool -L .venv/lib/python3.12/site-packages/torch/lib/libomp.dylib | grep libomp

.venv/lib/python3.12/site-packages/torch/lib/libomp.dylib:
        /opt/homebrew/opt/libomp/lib/libomp.dylib (compatibility version 5.0.0, current version 5.0.0)

and indeed when loading

-> DYLD_PRINT_LIBRARIES=1 python3 - <<EOF 2>&1 | grep libomp.dylib
import xgboost
import torch
EOF


dyld[58427]: <AE900456-D54F-3657-A54C-21B02B390E20> /opt/homebrew/Cellar/libomp/20.1.5/lib/libomp.dylib
dyld[58427]: <2448B10B-E7F4-3091-9302-765D47BFB0B8> /Users/iason/.venv/lib/python3.12/site-packages/torch/lib/libomp.dylib

while with the opposite import order

-> DYLD_PRINT_LIBRARIES=1 python3 - <<EOF 2>&1 | grep libomp.dylib
import torch
import xgboost
EOF

dyld[59185]: <2448B10B-E7F4-3091-9302-765D47BFB0B8> /Users/iason/.venv/lib/python3.12/site-packages/torch/lib/libomp.dylib

[matthewfeickert]

I think you're right about different symbols getting used, given that if everything is installed from conda-forge (on linux-64) then there's no issue

$ docker run --rm -ti ghcr.io/prefix-dev/pixi:latest 
root@31162fd0e8eb:/# pixi init example
Created /example/pixi.toml
root@31162fd0e8eb:/# cd example/
root@31162fd0e8eb:/example# pixi add xgboost pytorch
Added xgboost >=3.0.2,<4
Added pytorch >=2.7.0,<3
root@31162fd0e8eb:/example# pixi global install vim
Global environments as specified in '/root/.pixi/manifests/pixi-global.toml'
└── vim: 9.1.1397 (installed)
    └─ exposes: ex, rview, rvim, view, vim, vimdiff, vimtutor, xxd
root@31162fd0e8eb:/example# vim example.py
root@31162fd0e8eb:/example# cat example.py 
import xgboost
import torch

M = torch.tensor([[2.0, 0.5],
                  [0.5, 1.0]],
                 dtype=torch.double, device="cpu")

L = torch.linalg.cholesky(M)

print("Cholesky succeeded:", L)
root@31162fd0e8eb:/example# pixi run python example.py 
Cholesky succeeded: tensor([[1.4142, 0.0000],
        [0.3536, 0.9354]], dtype=torch.float64)
root@31162fd0e8eb:/example# find .pixi/ -type f -iname "libomp.*"
.pixi/envs/default/lib/libomp.so

Same as for osx-arm64:

% pixi init /tmp/example
✔ Created /private/tmp/example/pixi.toml
% cd /tmp/example
% pixi add xgboost pytorch
✔ Added xgboost >=3.0.2,<4
✔ Added pytorch >=2.7.0,<3
% vim example.py
% pixi run python example.py
Cholesky succeeded: tensor([[1.4142, 0.0000],
        [0.3536, 0.9354]], dtype=torch.float64)

[ikrommyd]

Yeah installing the packages from conda-forge is totally fine cause there is an openmp version coming with them from conda-forge. I'm talking about plain pypi installation here.

[trivialfis]

There is ongoing work on the Python wheel side for more flexible dependencies. cc @hcho3 https://github.com/wheelnext

For now, Python projects must package all dependencies into a wheel, which can create conflicts between packages. Until the wheel format starts supporting native dependencies, you should use conda/mamba to resolve these kinds of issues.