Cannot compile IPEX for Intel Arc A370M (DG2) - Compile errors #790
Comments
|
Hi,
|
|
@staropram Please follow the instructions from @jingxu10. The collect_env.py script tells us what versions of all libraries you have installed. The full compilation log tells us what went wrong during compilation. In general, these are the steps to install IPEX, where the link you provided is correct. For Arc A370M, you need to compile from source since there isn't a pre-built binary for it. https://pytorch-extension.intel.com/installation?platform=gpu&version=v2.5.10%2Bxpu&os=linux%2Fwsl2&package=source Just to be 100% certain, the compile script for IPEX v2.5.10+xpu is here: https://github.com/intel/intel-extension-for-pytorch/blob/v2.5.10%2Bxpu/scripts/compile_bundle.sh |
|
Thank you both for the help, here's the collect_env.py output: @jingxu10 I was using "2.5.10" above. I will build again from scratch using the exact compile bundle linked above from a conda env instead (just in case) and upload the build log. |
|
OK, I have tried again using the exact Here is and Any help you could give me would be appreciated. Thanks. |
|
This might be a red herring, I asked chatgpt to analyse the log file and it thinks this is the error:
Although ninja doesn't actually say anything until the linking step:
Is chatgpt sending us down a rabbit hole here or is it onto something? Otherwise, does the script have an easy way to produce more verbose output? I'll think about that while I await your observations of the above log and chatgpt's idea. Ash |
|
@staropram from your collect_env.log, your GCC version is 14.2.0. Can you try with GCC 11? |
|
I just get other errors, before we go down that route, can I please ask are you using icpx for host and device compilation? Or are you using gcc, which version? Clang? |
|
Just GCC 11 is sufficient. If you're still getting errors with that, please share with us the full compilation log. |
Describe the issue
Dear Intel IPEX Gods/Godesses,
Have pity on a humble man of the digital path, striving to incise truth and meaning from the chaos of build hell. Help me, oh wise ones, to find the correct incantations.
I have been struggling for the last 2 days trying to get intel-extension-for-pytorch working/compiled without success.
I went down a few wrong avenues, tried the existing packages (no detection of A370M), tried to just clone the repo and build but without success. Then I discovered intel's official build instructions tada, and this
compile_bundle.shapproach...So where am I, let's begin there:
compile_bundle.sh)compile_bundle.shuses both pip and conda, I was using apython -m venvenvironment not conda, so then I had to PATH the bin dir in from miniconda3 in order for your script to install libpng via this on line 264:conda install -y conda-forge::libpng conda-forge::libjpeg-turbo. It looked like I got away with this conda venv inception trick, but do I need to go full jungle anaconda on this and start again?So anyway, I'm running the script like this:
./compile_bundle.sh /home/ash/intel/oneapi dg2Is that right, oneapi installed from the official installer and that's where he lives, and then dg2 is supposedly the right mapping for AOT for the A370M, is that all correct?
So here's the first important question: is this actually supposed to work with the A370M? If it works am I going to get a GPU accelerated pytorch on my A370M? I'm not training grok5 here, I'm just playing with some teeny tiny models, but any speedup would be very welcome. If it won't ever work then I may as well stop now and buy an old gaming laptop with an NVIDIA GPU .
So I run the compilation and it starts compiling 8509 objects, the laptop has an i7-12700H but it takes some time as you'd imagine, and I lost track but this was the failure point:
I have no idea what's gone wrong here. Any help would be much appreciated.
Ash
The text was updated successfully, but these errors were encountered: