`pytorch-lightning=1.7` doesn't build against `pytorch=1.12`

[lcolbois]

Comment:

Hello,
I am unsure whether pytorch-lightning=1.7 is supposed to be compatible with pytorch=1.12, but the build is not possible as of now.
E.g:

CONDA_OVERRIDE_CUDA=11.2 mamba create -c conda-forge -n test pytorch-lightning=1.7 pytorch=1.12
...
Encountered problems while solving:
  - package pytorch-1.12.0-cpu_py310h75c9ab6_0 requires libprotobuf >=3.20.1,<3.21.0a0, but none of the providers can be installed

I think the conflict is between pytorch=1.12 that seem to require libprotobuf>=3.20 and this package's tensorboard>=2.9.1 constraint that adds an underlying pin to protobuf < 3.20.

I am uncertain whether this can be solved here or should be done in tensorboard or pytorch feedstocks.

Also, is there a way to setup the CI so that it attempts a build with the latest pytorch conda-forge release ?

[Borda]

Interesting, we shall find some better solution for the protobuf dependency
cc: @carmocca

[anjos]

The command above works for me as of today. What doesn't work, and it seems to be related, is having an environment with both those and tensorflow. Downgrading pytorch-lightning to 1.5 makes it work again:

# OK
$ CONDA_OVERRIDE_CUDA=11.2 mamba create --dry-run -c conda-forge -n test pytorch-lightning=1.5 pytorch=1.12 tensorflow=2.9

# Not OK
CONDA_OVERRIDE_CUDA=11.2 mamba create --dry-run -c conda-forge -n test pytorch-lightning=1.7 pytorch=1.12 tensorflow=2.9
Encountered problems while solving:
  - package tensorflow-base-2.9.1-cpu_py310h8df3ab6_0 requires libprotobuf >=3.20.1,<3.21.0a0, but none of the providers can be installed

Any clues?

[anjos]

I'm sensing this is a tensorflow issue and non-related to this package? If so, you can probably close this one.

[Borda]

Looking at the report again, this seems unrelated to Lightning as Conda is saying that torch and libprotobuf are not compatible, and it was the user request to use pytorch=1.12