Hello All!
The “Install PyTorch” grid on the main pytorch.org page only offers pip installation (and I
believe that conda support has been discontinued). However, the language just above the
grid, “Anaconda is our recommended package manager”, hasn’t been updated to reflect
this change.
Best.
K. Frank
Thanks for providing the feedback @KFrank!
While the conda packages were removed, the Anaconda package manager can still be used to create and manage virtual environments. Nevertheless, we might want to update the section with newer recommendations.
Would you have a favorite package manager for your PyTorch environments you could share here?
Hi @ptrblck!
The short answer is that I use the pip3 command from the install grid to install into multiple
conda environments. I don’t set up anything unusual or fancy – I just take what the standard
pytorch installation command gives me.
The install grid used to produce a conda installation command, so I started with that. Recently
that option was dropped so I’ve been using pip3 for the actual installation. But I’ve found conda
environments useful, so I run pip3 install in a conda environment.
My workflow for a new pytorch version is to create a new conda environment and install into
that. This allows me to switch back and forth between pytorch versions. For example, in
addition to the current 2.9.0, I’ve kept 2.8.0 and a couple of other older versions around in
case I want to look for a possible regression. This also lets me try a nightly or something
like the xpu version without it colliding with the stable release.
Best.
K. Frank
Thank you for sharing your workflows. Your current workflow matches exactly what I’m doing. I understand the docs statement of mentioning Anaconda as a package manager matches exactly your use case as well and does not recommend using the (dropped) conda package itself. Let me know if I misunderstand the docs.
i feel like it wouldn’t be too controversial at this point to settle on pip and uv
Maybe for commercial usecases, one nuance to follow is that, Its good to install conda from miniforge rather than anaconda binary just to have conda-forge as default channel and stay away from non-commercial default channels in anaconda.