Packaging difference conda-forge's and pytorch's distribution of pytorch

[jjerphan]

Hi,

Thank you for packaging pytorch on conda-forge!

I have few questions about this distribution with respect to the official one from pytorch:

• What is the plus-value of installing conda-forge's builds of pytorch instead of pytorch's?
• What was the original reason behind redistributing pytorch on conda-forge?
• Do you know a summary of the differences between conda-forge's and the original distribution of pytorch?

[hmaarrfk]

What is the plus-value of installing conda-forge's builds of pytorch instead of pytorch's?

If your computation stack depends on some optimized C binaries and they aren't available on pypi, then conda allows you to have your own "channel" and build in accordance to your requirements.

pytorch's channel depends on anaconda's "default" channel optimizations.

on conda-forge, we depend on conda-forge's optimizations.

On conda-forge, it is a community lead development so you and others can have an impact on the development. For example, I believe we were the first to have Apple silicon (arm) packages.

What was the original reason behind redistributing pytorch on conda-forge?

The above.

Do you know a summary of the differences between conda-forge's and the original distribution of pytorch?

• arm compatibility on linux
• lack of windows compatibility (help appreciated)
• usage of shared libprotobuf
• Probably a few more.

[jjerphan]

Thank you for those elements.

pytorch and torchvision (and a few other packages from NVIDIA it seems) are distributed on conda-forge, but the rest of the packages are distributed on the pytorch and nvidia channels.

In the case of pytorch, to which extends is it possible to mix packages coming from different channels? Installing as many packages as possible from conda-forge and then from pytorch is possible, but would such an installation be prone to problems?

[hmaarrfk]

yes prone to problems if you mix conda-forge + defaults