Intel GPU pytorch installation on Windows

[TheRain_MA]

Dear developers and users,

Hello!

I followed the GPU-version pytorch installation procedure posted on Github, however, there are so many conflicts like "expected torch version can not be downloaded", "no torchvision GPU version for windows", or some other errors. 

I am not a professional computer science learner, so I wonder if someone could update the installation instructions, or if there is a day that everything will go into an easy way.

Thank you very much!

Best Regards,

MA

[RahulC_Intel]

Hi Xibo, could you please share the below details for further investigation: 

• OS build on which you are installing the software  
• Is the installer an online or an offline package?  
• Version of Microsoft Visual Studio installed along with it's build version
• Exact version of pytorch which you are installting in your system
• Other versions of Intel® Parallel Studio XE or Intel® oneAPI already installed on the system   
• Installation Log files.  
  • The default path for a log file for a oneAPI installation is: C:\Program Files (x86)\Intel\oneAPI\logs  
  • The default path for a log file for a Parallel Studio XE installation is: %temp%\pset_tmp_PSXE<version>_<username> 

[TheRain_MA]

Hello RahulC_Intel

Thank you for your quick respond!

Finally I successfully ran the .bat file under the guidance from [Installation Guide (Windows)], and I got three files named <intel-extension-for-pytorch> <pytorch> <vision>, but the packs might not be installed correctly. It showed that "ModuleNotFoundError: No module named 'torch".

I rechecked the files, found that there was no <dist> folder in any of three files.

Could you please help me find the reason?

Here are information you need.

        ·  Windows 10.

        ·  I used [Install via compiling from source].

        ·  2022 Community.

        ·  I don't know since I used [Install via compiling from source].

        ·  oneAPI version : latest

·  The whole installation procedure and commond was conducted under a conda environment with Miniconda Prompt, python version I built in the environment was 3.9.

·  I uploaded the log file a oneAPI installation.

·  I didn't install Python since I got Miniconda, was that a possible reason? 

Thank you for your help.

Yours sincerely,

MA

[RahulC_Intel]

Hi Xibo, yes you need to install python as well. Kindly refer to the link: https://intel.github.io/intel-extension-for-pytorch/xpu/latest/tutorials/installations/windows.html, and follow the installation steps for Intel® Extension for PyTorch.

[TheRain_MA]

Hello RahulC_Intel

Thank you for your reply and suggestions.

I think I have followed the steps strictly, but they can't work well.

· Via pip command

        I can only import <torch>.

        <inter_extension_for_pytorch> could not work well and showed as "module can't be found". However, the <inter_extension_for_pytorch> pack actually existed in the environment examined by Pycharm.

· Via compiling from source

    All three packs could not be installed. Lots of files were generated but there was no dist file and .whl. You may check the log file I uploaded if it helps.

Also, I tried to follow the tuition about using DirectML to make Pytorch possible. But I found some packs could not work as well.

However, the trial of using DirectML to make TensorFlow possible was much better, and so far I haven't found serious issues.

Anyway, I think I could just wait, and looking forward to a new version of <inter_extension_for_pytorch on Windows> which concludes <torch>, <torchvision>, and < inter_extension_for_pytorch> without a complex installation procedure.

And many thanks for the team's hard work so far and future.

Best Regard

[RahulC_Intel]

Hi Xibo, we have sent you a private message. Kindly check your inbox.

[RahulC_Intel]

Hi Xibo, thank you for the update. Could you please share the below details for further investigation:

• Are you using Build Tool for Visual Studio 2022.
• Are you using Intel® Arc™ A-Series Graphics.
• Exact version of Python.
• Are you using Intel® oneAPI Base Toolkit 2023.2.0.

We would like to know.

[RahulC_Intel]

Hi XIBO, unfortunately we were unable to hear back from you. 

If you have any further query, please post a new question as this thread will no longer be monitored by Intel®.