Specify folder for git clone in compile-pytorch-ipex.sh; update IPEX pin

[anmyachev]

Closes #2651

[anmyachev]

By default was pytorch

[whitneywhtsang]

Closes #2651

The change LGTM, but it doesn't resolve the problem reported by #2651.

[anmyachev]

@Stonepia just to clarify, did you made force push to dev/triton-test-3.0 branch? intel/intel-extension-for-pytorch@15ef7db looks similar to the previous that we use.

[anmyachev]

@whitneywhtsang could you try with the last change?

[whitneywhtsang]

@whitneywhtsang could you try with the last change?

It passes the reported problem, it is at the step [ 74%] Built target dnnl_gpu_ocl for a while, will reply again when it works e2e.

[whitneywhtsang]

[ 74%] Built target dnnl_gpu_jit
make: *** [Makefile:136: all] Error 2
Traceback (most recent call last):
  File "/home/jovyan/intel-xpu-backend-for-triton/.scripts_cache/intel-extension-for-pytorch/setup.py", line 1168, in <module>
    setup(
  File "/home/jovyan/.conda/envs/python-3.12/lib/python3.12/site-packages/setuptools/__init__.py", line 117, in setup
    return distutils.core.setup(**attrs)
           ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^

[Stonepia]

Oh sorry for that. Yes we made a force push to the dev/triton-test-3.0 branch, because there is a token leak, that should be our fault. We will fix that.

[Stonepia]

Hi @anmyachev , I really apology for not informing you about the force push. Because I thought Triton had already switched to the stock PT.

About this build problem specifically, I have checked about the code change, seems there is no change that would cause the break. I don't see anything that we missed.

Most likely, is that during that version of IPEX, the CPU build would break the XPU version of IPEX. In addition, there might be because of the one API version.

Could you try these things? I could build on my machine when the CPU part of IPEX is closed.

1. What oneAPI are you using? The dev bundle 0.5.3 could work on my machine.
2. Could you try the following flags before building?

# I think this flag will be ok. We don't need CPU.
export BUILD_WITH_CPU=OFF
# more flags to close other things.  You could close these things
# I don't think these things would affect Triton dev.
export USE_XETLA=OFF
export USE_PTI=OFF
export USE_KINETO=OFF

Again, so sorry for the issue. If there is anything I could help, please tell me.

[Stonepia]

I didn't try to build with the former version of oneAPI (like 2023.2). I think the oneAPI version might explain why the former build could pass, and now it can't. From the code, I checked that and there should not have missing parts.

[anmyachev]

Hi @Stonepia,

thanks for confirming! Don't worry, it's not that bad. I was able to build code with PTDB, even with python 3.12.

[ 74%] Built target dnnl_gpu_jit
make: *** [Makefile:136: all] Error 2
Traceback (most recent call last):
  File "/home/jovyan/intel-xpu-backend-for-triton/.scripts_cache/intel-extension-for-pytorch/setup.py", line 1168, in <module>
    setup(
  File "/home/jovyan/.conda/envs/python-3.12/lib/python3.12/site-packages/setuptools/__init__.py", line 117, in setup
    return distutils.core.setup(**attrs)
           ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^

@whitneywhtsang I can't reproduce that problem. If it's still there, I guess we can continue in a separate thread.

[whitneywhtsang]

@whitneywhtsang I can't reproduce that problem. If it's still there, I guess we can continue in a separate thread.

Sure, my environment is Agama 1032.19, DLE 2025.0.0.

[anmyachev]

@whitneywhtsang I can't reproduce that problem. If it's still there, I guess we can continue in a separate thread.

Sure, my environment is Agama 1032.19, DLE 2025.0.0.

Well, this is kind of expected, our IPEX branch doesn’t involve working with 2025 compiler, or I missed something?

[whitneywhtsang]

Well, this is kind of expected, our IPEX branch doesn’t involve working with 2025 compiler, or I missed something?

Make sense, verified it works for me with PTDB.