IPEX xpu wheel files for windows using the wrong version of Pytorch? Also no torchvision

[Mindset-Official]

Describe the issue

python -m pip install torch==2.0.0a0 intel_extension_for_pytorch==2.0.110+gitba7f6c1 -f https://developer.intel.com/ipex-whl-stable-xpu

wheel files seem to include torch 2.0.0a0 but notes say that xpu is only available for torch 2.0.1a0

Python 3.10.6 (tags/v3.10.6:9c7b4bd, Aug 1 2022, 21:53:49) [MSC v.1932 64 bit (AMD64)] on win32
Type "help", "copyright", "credits" or "license" for more information.

import torch
import intel_extension_for_pytorch
Traceback (most recent call last):
File "", line 1, in
File "C:\Users\mymin\AppData\Local\Programs\Python\Python310\lib\site-packages\intel_extension_for_pytorch_init_.py", line 89, in
raise err
OSError: [WinError 126] The specified module could not be found. Error loading "C:\Users\mymin\AppData\Local\Programs\Python\Python310\lib\site-packages\intel_extension_for_pytorch\bin\intel-ext-pt-gpu.dll" or one of its dependencies.
torch.xpu.is_available()
Traceback (most recent call last):
File "", line 1, in
AttributeError: module 'torch' has no attribute 'xpu'

Also there seems to be no version of torchvision for windows? Not sure if this is a mistake or just part of the expiremental nature? as is xpu is not detected and doesn't work on windows so it defaults to cpu still.

[jingxu10]

Because of the experimental nature, the pytorch wheel shipped is 2.0.1 code base, but didn't get version updated in its version.txt file. torchvision wheels are not available at this moment as well.

@min-jean-cho could you check the error above?

OSError: [WinError 126] The specified module could not be found. Error loading "C:\Users\mymin\AppData\Local\Programs\Python\Python310\lib\site-packages\intel_extension_for_pytorch\bin\intel-ext-pt-gpu.dll" or one of its dependencies.

[min-jean-cho]

Thanks @Mindset-Official for trying out. Could you please double check if you have activated oneAPI environment, call "{YOUR_PATH_TO_ONEAPI}\setvars.bat" (e.g., call "C:\Program Files (x86)\Intel\oneAPI\setvars.bat")?

[Mindset-Official]

Okay, thanks. I was able to get it to run by deleting python and installing miniconda as default, I then ran everything in oneapi environment and

Python 3.10.12 | packaged by Anaconda, Inc. | (main, Jul 5 2023, 19:01:18) [MSC v.1916 64 bit (AMD64)] on win32
Type "help", "copyright", "credits" or "license" for more information.

import torch
import intel_extension_for_pytorch
torch.xpu.is_available()
True

appreciate the help and response. I will give it a shot and see how everything works in windows.

[min-jean-cho]

Glad to hear that! Let us know if any other questions.

[Nuullll]

Since Intel hasn't released a corresponding torchvision wheel for windows, the SD Web UI tries to install the latest torchvision == 0.15.2, which requires torch == 2.0.1. So pip install torchvision will uninstall the intel torch because of incompatible version and install torch == 2.0.1 instead. Then intel_extension_for_pytorch bails out because of wrong dll version from torch 2.0.1.

I have to hack torchvision-0.15.1.dist-info/METADATA in the python site-packages to force it compatible with torch == 2.0.0a0.

Metadata-Version: 2.1
Name: torchvision
Version: 0.15.1
Summary: image and video datasets and models for torch deep learning
Home-page: https://github.com/pytorch/vision
Author: PyTorch Core Team
Author-email: soumith@pytorch.org
License: BSD
Requires-Python: >=3.8
License-File: LICENSE
Requires-Dist: numpy
Requires-Dist: requests
- Requires-Dist: torch (==2.0.0)
+ Requires-Dist: torch (==2.0.0a0)
Requires-Dist: pillow (!=8.3.*,>=5.3.0)
Provides-Extra: scipy
Requires-Dist: scipy ; extra == 'scipy'

[Nuullll]

So we really need an intel torchvision wheel for windows. Is that planned in the near future?

[Mindset-Official]

Since Intel hasn't released a corresponding torchvision wheel for windows, the SD Web UI tries to install the latest torchvision == 0.15.2, which requires torch == 2.0.1. So pip install torchvision will uninstall the intel torch because of incompatible version and install torch == 2.0.1 instead. Then intel_extension_for_pytorch bails out because of wrong dll version from torch 2.0.1.

I have to hack torchvision-0.15.1.dist-info/METADATA in the python site-packages to force it compatible with torch == 2.0.0a0.

Metadata-Version: 2.1
Name: torchvision
Version: 0.15.1
Summary: image and video datasets and models for torch deep learning
Home-page: https://github.com/pytorch/vision
Author: PyTorch Core Team
Author-email: soumith@pytorch.org
License: BSD
Requires-Python: >=3.8
License-File: LICENSE
Requires-Dist: numpy
Requires-Dist: requests
- Requires-Dist: torch (==2.0.0)
+ Requires-Dist: torch (==2.0.0a0)
Requires-Dist: pillow (!=8.3.*,>=5.3.0)
Provides-Extra: scipy
Requires-Dist: scipy ; extra == 'scipy'

Try this <@323514682155466753> 'torch==2.0.0a0 torchvision==0.15.2a0 intel_extension_for_pytorch==2.0.110+gitba7f6c1 -f https://developer.intel.com/ipex-whl-stable-xpu' 0.15.2a0 seems to install correctly for me. It should atleast install without error

[Nuullll]

Thanks @Mindset-Official. torchvision-0.15.2a0+cxx11.abi-cp310-cp310-linux_x86_64.whl seems to be a linux wheel, I doubt it would work as expected on native windows. Anyway I'd give it a try.

Update:

ERROR: torchvision-0.15.2a0+cxx11.abi-cp310-cp310-linux_x86_64.whl is not a supported wheel on this platform.

[charygu]

Okay, thanks. I was able to get it to run by deleting python and installing miniconda as default, I then ran everything in oneapi environment and

Python 3.10.12 | packaged by Anaconda, Inc. | (main, Jul 5 2023, 19:01:18) [MSC v.1916 64 bit (AMD64)] on win32 Type "help", "copyright", "credits" or "license" for more information.

import torch
import intel_extension_for_pytorch
torch.xpu.is_available()
True

appreciate the help and response. I will give it a shot and see how everything works in windows.

Hi, Mindset-Official, Could you tell me how you fix that issue? Did you get this fixed on native Windows?

[Mindset-Official]

Okay, thanks. I was able to get it to run by deleting python and installing miniconda as default, I then ran everything in oneapi environment and
Python 3.10.12 | packaged by Anaconda, Inc. | (main, Jul 5 2023, 19:01:18) [MSC v.1916 64 bit (AMD64)] on win32 Type "help", "copyright", "credits" or "license" for more information.

import torch
import intel_extension_for_pytorch
torch.xpu.is_available()
True

appreciate the help and response. I will give it a shot and see how everything works in windows.

Hi, Mindset-Official, Could you tell me how you fix that issue? Did you get this fixed on native Windows?

Yeah, what I did in that post fixed it. You can probably run it without miniconda but I left it as is. Definitely make sure to run the environment and call those variables every time you run.