What's exactly the dependencies if install Intel optimization of PyTorch?

[ChenHaifeng]

I am trying to integration oneAPI deep learning optimizations to CloudTik (https://github.com/oap-project/cloudtik)

While I found the different installation options are very confusing and it's not clear what's the must have components that I need install and configure for Intel optimizations of PyTorch.

To be clear the question:

1. When I install through apt: sudo apt-get install intel-aikit-pytorch

I got a list of apt dependencies which are installed in the /opt/intel/oneapi and there are conda environment for pytorch under it.

2. When I install through conda: conda install -c intel intel-aikit-pytorch=2023.1.0

I got a list of conda dependencies which are all installed under the current conda environment.

The dependencies are seems to be quite different (user cannot track whether it is identical with #1 since the layouts are already quite different)

3. When I install through conda: conda install -c intel-extension-for-pytorch=1.13.10

The list of dependencies are different with #2 but we know intel-aikit-pytorch includes nerual-compressor.

4. When I install through pip: pip install intel-extension-for-pytorch==1.13.100

The list of pip packages are huge different than #3 in libraries such as intelxxx dependencies in #3 but not in pip dependencies. for example, intelpython thing.

Since there are many options, and every option are quite different from key dependences point of view. 

As oneAPI user (developer who uses oneAPI component in a project), I can hardly figure out what's the best choice if I want use one component (like Intel optimizations for PyTorch) and it's minimum requirement.

For developers to use one component, we usually have an existing environment to integrate and will hardly to take everything (intel-aikit for example) as a whole because it's too large and more complex for conflictions)

Instead we want to integrate what we need, for example Intel optimizations for PyTorch or Intel optimizations for Tensorflow. We need to know what's the minimum requirement for each component. And ideally, we need consistent dependencies and layout for different installation options so that whatever option I choose, I get the logical identical thing instead of quite different thing and get very confused.

Hopefully, someone develops these can give some information on the fundamental differences on what we get between the above options.

I hope we can visit these aspects in the design of the packages and improve the use experiences to reflect the fundamental dependencies of each component. 

[RemyaP_Intel]

Hi,

Thanks for sharing your observations. We have also observed the same behavior. We are checking on this internally. We will get back to you.

Regards,

Remya Premdas

[Munara_T_Intel]

Thank you for your post. We are still looking into this question, and will update once we have more information.

[Munara_T_Intel · Accepted solution]

Hi,

Intel Pytorch team suggests to use package from pypi, which is maintained by their team directly. You can always get latest and clean IPEX (Intel Extension for PyTorch) with pip install.

Thanks