it's available via the default channel? how?

Yeah actually this is pretty odd, and maybe my conda has gone weird somehow (running Ubuntu 14.04, conda with Python 3.6 and CUDA 8.0). Having run conda install pytorch previously, I now get:

conda install pytorch
Fetching package metadata .........
Solving package specifications: .

# All requested packages already installed.
# packages in environment at /home/ka709/miniconda3:
#
pytorch                   0.1.12                   py36_0

Versus:

conda install pytorch -c soumith
Fetching package metadata ...........
Solving package specifications: .
Warning: 2 possible package resolutions (only showing differing packages):
  - soumith::pytorch-0.1.12-py361_2cu80
  - soumith::pytorch-0.1.12-py36_2cu80

Package plan for installation in environment /home/ka709/miniconda3:

The following packages will be UPDATED:

    pytorch: 0.1.12-py36_0 --> 0.1.12-py361_2cu80 soumith [cuda80]

Installing with conda install pytorch -c soumith, and then running conda install pytorch gives:

conda install pytorch
Fetching package metadata .........
Solving package specifications: .

Package plan for installation in environment /home/ka709/miniconda3:

The following packages will be SUPERSEDED by a higher-priority channel:

    pytorch: 0.1.12-py361_2cu80 soumith [cuda80] --> 0.1.12-py36_0

paging @kalefranz could I request you to please disable or remove the anaconda pytorch and pytorch-gpu packages? they are not fully done yet:

• they dont have an OSX builds
• they separately package CPU and GPU bits (which we absolutely don't recommend, it's confusing to the user in many ways. If you insist pytorch has to have GPU support and pytorch-cpu has to be the CPU only build)
• the packages only are build against CUDA 7.5, we need to provide CUDA 7.5 and CUDA 8.0 versions because 7.5 doesn't support newer cards.
• We don’t want to ship nomkl packages. PyTorch without a good BLAS library crawls on the CPU side, and doesn’t have LAPACK routines enabled.

Jonathan sent me an email about these packages a couple of months ago (when they were private), but I didn't get time to look into his package scripts yet (i am a bit overworked, but it is on my TODO list). Someone made these packages public on the anaconda side a couple of days ago .
Reference URL: https://anaconda.org/anaconda/pytorch/files

Yep if I just want to naively do conda update --all now then I get:

The following packages will be SUPERSEDED by a higher-priority channel:

    pytorch:     0.2.0-py36h6d07548_2cu80 soumith [cuda80] --> 0.1.12-py36cuda7.5cudnn6.0_1
    torchvision: 0.1.9-py36h7584368_1     soumith --> 0.1.8-py36_0

Looking at the download count, people are actually pulling these new packages (uploaded 4 days ago).

reported this to the conda folks :(
There's not much I can do other than reporting it to them, and they're slow in issuing new packages (their staging / testing process takes some time), and they dont have Cuda8 support yet.

So what is the recommended process? I see on the pytorch page that -c soumith is still recommended. However, I am working in a context where proxy blocks access to https://conda.anaconda.org/soumith/linux-64/repodata.json but not to the default channels. I care just about a CPU version on RHL 6.7 (for now) -- though of course I would prefer pytorch with optimized libraries. Should I go ahead and use the default channel?

the default channel has 0.1.12 which is a very old channel. if you only require CPU and you have access only to default channel, build from source: https://github.com/pytorch/pytorch#from-source