On:

• Ubuntu 24.04
• Intel Data Center GPU Max

Usage scenario:

• User installs pytorch of some version
• User installs oneAPI of some version
• User intends to use pytorch and oneAPI for separate sub-tasks (i.e. there is no user intent to use installed oneAPI in pytorch workflow and vice versa). However, pytorch and oneAPI sub-tasks might be ingredients of bigger common workflow.
• User activates oneAPI environment within existing pytorch environment

Issue:

• If activated oneAPI version does not match the version of oneAPI used in the selected version of pytorch user will see the following failure:

$ pip install torch==2.8 --index-url https://download.pytorch.org/whl/xpu
$ pip install numpy
$ python -c "import torch; print(torch.xpu.is_available())"
True

$ . /opt/intel/oneapi/2025.2/oneapi-vars.sh
$ python -c "import torch; print(torch.xpu.is_available())"
Traceback (most recent call last):
  File "<string>", line 1, in <module>
  File "/home/dvrogozh/pt2.8/lib/python3.12/site-packages/torch/__init__.py", line 416, in <module>
    from torch._C import *  # noqa: F403
    ^^^^^^^^^^^^^^^^^^^^^^
ImportError: /opt/intel/oneapi/2025.2/lib/libur_loader.so.0: version `LIBUR_LOADER_0.11' not found (required by /home/dvrogozh/pt2.8/lib/python3.12/site-packages/torch/lib/../../../../libsycl.so.8)

Expectation:

• Pytorch is using RPATH pointing to pinned versions of oneAPI runtime distributed with pytorch itself. Expectation is that pytorch will continue to run regardless whether user installed and sourced environment of some other version of oneAPI.

CC: @riverliuintel @chuanqi129