Intel GPU acceleration is missing (XPU not enabled)

pytorch has recently added support for Intel GPU acceleration 1.

To test if my Intel Arc GPU can be used by pytorch I can run python -c 'import torch; torch.tensor([1.0, 2.0]).to("xpu")'

That test fails today with python-pytorch 2.6.0-8 like this:

$ python -c 'import torch; torch.tensor([1.0, 2.0]).to("xpu")'
Traceback (most recent call last):
  File "<string>", line 1, in <module>
    import torch; torch.tensor([1.0, 2.0]).to("xpu")
                  ~~~~~~~~~~~~~~~~~~~~~~~~~~~^^^^^^^
  File "/usr/lib/python3.13/site-packages/torch/xpu/__init__.py", line 118, in _lazy_init
    raise AssertionError("Torch not compiled with XPU enabled")
AssertionError: Torch not compiled with XPU enabled

added priority3-normal severity4-low statusunconfirmed labels

assigned to @freswa, @toolybird, @gromit, and @tpkessler

I've taken a crack at this myself: greyltc/python-pytorch@main...add-xpu

This build "works" (passes my quick xpu test above). I used the full intel-oneapi-buildkit environment at runtime though in my testing (. /opt/intel/oneapi/setvars.sh). So probably there are runtime dependency improvements to be made here (I highly doubt the whole intel-oneapi-buildkit is needed at runtime).

I don't have AMD or CUDA hardware to test with (and honestly I commented out the builds for those packages in my testing, so I can't even say if those packages build with my changes).

added scopefeature statusconfirmed labels and removed statusunconfirmed label

assigned to @svenstaro and @kgizdov and unassigned @freswa, @toolybird, and @gromit

my XPU build changes seem to have issues. If I build with USE_GLOG=ON, the build fails, but if I build with USE_GLOG=OFF the build completes but I get a runtime error:

ERROR: flag 'logtostderr' was defined more than once (in files '/build/python-pytorch/src/pytorch-opt/c10/util/Logging.cpp' and '/usr/src/debug/google-glog/glog-0.7.1/src/flags.cc').

I've made a few more hacks to work around the glog issues I was having (in https://gitlab.archlinux.org/greyltc/python-pytorch/-/tree/wip-xpu ), but now I can't get past this runtime error:

RuntimeError: Native API failed. Native API returns: 20 (UR_RESULT_ERROR_DEVICE_LOST)

Although I get the same error with a precompiled xpu torch installed via python -m pip install --pre torch torchvision torchaudio --index-url https://download.pytorch.org/whl/nightly/xpu so maybe I've got other issues.