intel
2.8.10+xpu wheel invalid
Describe the bug
Installing via pip results in an error citing the wheel is invalid.
"ERROR: Wheel 'intel-extension-for-pytorch' located at /tmp/pip-unpack-n6ux5tob/intel_extension_for_pytorch-2.8.10+xpu-cp311-cp311-linux_x86_64.whl is invalid"
I have purged the pip cache and tried multiple times to eliminate the likelihood that the download was corrupted and I did attempt the pip install from https://pytorch-extension.intel.com/release-whl/stable/xpu/us
It appears to complete the download but directly afterward, I receive the above error.
This is being installed on Arch Linux into a python 3.11 venv.
SIDENOTE: While trying to start ComfyUI afterward, it produces an error stating that LIBUR_LOADER_0.11 is not found when trying to import torch (torch._C).
Versions
ERROR! Intel® Extension for PyTorch* needs to work with PyTorch 2.7.*, but PyTorch 2.8.0+xpu is found. Please switch to the matching version and run again.
UPDATE
After further investigation, it seems that the intel-oneapi-basekit package in the Arch Linux repositories provide version 10 of libur_loader.so with the compiler. We have the intel-oneapi-base-toolkit package available in the "Extra-Testing" repository, but that package provides 2025.2 which provides libur_loader.so.0.12 (version 12).
There is no package (that I can find) currently available in Arch Linux that will provide libur_loader.so.0.11 which is required by torch==2.8.0+xpu.
Additionally, I cannot download the 2025.1.3 Intel-OneAPI-Base package from Intel.com. When I choose the 2025.1.3 version from the dropdown menu, the site seems to disregard the choice and downloads the offline installer for 2025.2 anyway.
Solution1: Change the requirement for libur_loader.so to either versions 10 or 12...or even better, just make it >= 10, if possible.
Solution2: Make the 2025.1.3 package available for download.
I haven't seen the same error on pip install, however I was also seeing this error when running python scripts that load the torch library:
ImportError: /opt/intel/oneapi/compiler/2025.2/lib/libur_loader.so.0: version LIBUR_LOADER_0.11' not found`
It turns out that this error was happening because I was running:
source /opt/intel/oneapi/setvars.sh
When I run the pytorch wheel in an environment where I haven't run the setvars script, everything runs fine.
Please refer to the installation guide for details: https://pytorch-extension.intel.com/installation?platform=gpu&version=v2.8.10%2Bxpu&os=linux%2Fwsl2&package=pip
When you are using prebuilt wheels, please don't source oneAPI env variables, the dependent libraries will be installed automatically when you install IPEX.
If you build IPEX from source, please follow https://pytorch-extension.intel.com/installation?platform=gpu&version=v2.8.10%2Bxpu&os=linux%2Fwsl2&package=source. The supported oneAPI version is 2025.1.
