Conda installation wants to downgrade to PyTorch 1.12.1

[niphiwi]

On a fresh conda environment with only PyTorch 2.1.2 cpuonly (and its dependencies installed), I want to install Pytorch Cluster.

conda install pytorch-cluster -c pyg throws me the following error:

conda install pytorch-cluster -c pyg
Channels:
 - pyg
 - defaults
 - pytorch
Platform: win-64
Collecting package metadata (repodata.json): done
Solving environment: \ warning  libmamba Added empty dependency for problem type SOLVER_RULE_UPDATE
failed

LibMambaUnsatisfiableError: Encountered problems while solving:
  - nothing provides python_abi 3.6.* *_cp36m needed by pytorch-cluster-1.5.9-py36_torch_1.8.0_cpu

Could not solve for environment specs
The following packages are incompatible
├─ pin-1 is installable and it requires
│  └─ python 3.11.* , which can be installed;
└─ pytorch-cluster is not installable because there are no viable options
   ├─ pytorch-cluster 1.5.9 would require
   │  └─ python_abi 3.6.* *_cp36m, which does not exist (perhaps a missing channel);
   ├─ pytorch-cluster 1.5.9 would require
   │  └─ python_abi 3.7.* *_cp37m, which does not exist (perhaps a missing channel);
   ├─ pytorch-cluster 1.5.9 would require
   │  └─ python_abi 3.8.* *_cp38, which does not exist (perhaps a missing channel);
   ├─ pytorch-cluster 1.5.9 would require
   │  └─ python_abi 3.9.* *_cp39, which does not exist (perhaps a missing channel);
   ├─ pytorch-cluster 1.5.9 would require
   │  └─ cudatoolkit 11.1.* , which does not exist (perhaps a missing channel);
   ├─ pytorch-cluster 1.6.0 would require
   │  └─ python >=3.7,<3.8.0a0 , which conflicts with any installable versions previously reported;
   ├─ pytorch-cluster 1.6.0 would require
   │  └─ python >=3.8,<3.9.0a0 , which conflicts with any installable versions previously reported;
   ├─ pytorch-cluster 1.6.0 would require
   │  └─ python >=3.9,<3.10.0a0 , which conflicts with any installable versions previously reported;
   ├─ pytorch-cluster 1.6.0 would require
   │  └─ cudatoolkit 11.5.* , which does not exist (perhaps a missing channel);
   ├─ pytorch-cluster 1.6.0 would require
   │  └─ python >=3.10,<3.11.0a0 , which conflicts with any installable versions previously reported;
   └─ pytorch-cluster 1.6.0 would require
      └─ cudatoolkit 11.6.* , which does not exist (perhaps a missing channel).

Pins seem to be involved in the conflict. Currently pinned specs:
 - python 3.11.* (labeled as 'pin-1')

My understanding is that my problem is the incompatibility with Python 3.11. Therefore I downgrade to Python 3.10:

conda install python=3.10

Now, however, it tries to downgrade to PyTorch 1.12.1. Why is that? I was able to install PyTorch Cluster with Pip in the past, but what's the problem with conda that I am missing?

The following NEW packages will be INSTALLED:

  cudatoolkit        pkgs/main/win-64::cudatoolkit-11.3.1-h59b6b97_2
  future             pkgs/main/win-64::future-0.18.3-py310haa95532_0
  pytorch-cluster    pyg/win-64::pytorch-cluster-1.6.0-py310_torch_1.12.0_cu113
  six                pkgs/main/noarch::six-1.16.0-pyhd3eb1b0_1

The following packages will be DOWNGRADED:

  intel-openmp                      2023.1.0-h59b6b97_46320 --> 2021.4.0-haa95532_3556
  mkl                               2023.1.0-h6b88ed4_46358 --> 2021.4.0-haa95532_640
  mkl-service                         2.4.0-py310h2bbff1b_1 --> 2.4.0-py310h2bbff1b_0
  mkl_fft                             1.3.8-py310h2bbff1b_0 --> 1.3.1-py310ha0764ea_0
  mkl_random                          1.2.4-py310h59b6b97_0 --> 1.2.2-py310h4ed8f06_0
  numpy                              1.26.3-py310h055cbcc_0 --> 1.24.3-py310hdc03b94_0
  numpy-base                         1.26.3-py310h65a83cf_0 --> 1.24.3-py310h3caf3d7_0
  pytorch                         2.1.0-cpu_py310hb0bdfb8_0 --> 1.12.1-cpu_py310h5e1f01c_1
  torchaudio                                2.1.0-py310_cpu --> 0.12.1-py310_cpu

[rusty1s]

We currently have trouble building conda wheels on Windows, that's why it is probably downgrading you to an earlier PyTorch version. Would installing from pip work for you for now?

[niphiwi]

Installing from pip is okay for me (and works). I was just curious about the behavior and did not find a related issue here. But then it's probably related to the warning in the PyG documentation ("Conda packages are currently not available for Windows and M1/M2/M3 macs")?

[rusty1s]

Yes :(

[e-eight]

I am having this same issue on Linux.