Different results for conda install and mamba install

[MortenHannemose]

Using conda to install pytorch without a GPU finds the following packages:
conda install pytorch torchvision cpuonly -c pytorch

    package                    |            build
    ---------------------------|-----------------
    cpuonly-1.0                |                0           2 KB  pytorch
    pytorch-1.6.0              |      py3.8_cpu_0       142.0 MB  pytorch
    torchvision-0.7.0          |         py38_cpu         5.8 MB  pytorch
    ------------------------------------------------------------
                                           Total:       148.0 MB

When executing the same command with mamba:
mamba install pytorch torchvision cpuonly -c pytorch
it finds the CUDA versions of the packages (intended for GPU)

  Package      Version  Build                   Channel                Size
-----------------------------------------------------------------------------
  Install:
-----------------------------------------------------------------------------

  cpuonly          1.0  0                       pytorch/noarch         2 KB
  cudatoolkit  10.2.89  h74a9793_1              pkgs/main/win-64     317 MB
  pytorch        1.6.0  py3.8_cuda102_cudnn7_0  pytorch/win-64       705 MB
  torchvision    0.7.0  py38_cu102              pytorch/win-64         6 MB

  Summary:

  Install: 5 packages

  Total download: 1 GB

I had expected that conda and mamba would find the same packages.

[wolfv]

Yes, pytorch (unfortunately) still uses track_features. We already have an issue for that here #336

[wolfv]

if you know the pytorch guys, please ping them on the linked issue!