PyTorch 2.8 with CUDA can't be installed on linux-aarch64 due to libcudss conflict

[traversaro]

Solution to issue cannot be found in the documentation.

• I checked the documentation.

Issue

While working on torchcodec migration to PyTorch 2.8, I experienced the following error:

traversaro@IITBMP014LW012:~$ CONDA_SUBDIR=linux-aarch64 conda create -n testpytorch pytorch=2.8.*=cuda*
Retrieving notices: done
Channels:
 - conda-forge
Platform: linux-aarch64
Collecting package metadata (repodata.json): done
Solving environment: failed

LibMambaUnsatisfiableError: Encountered problems while solving:
  - package libtorch-2.8.0-cuda129_generic_h4d1ebf7_200 requires libcudss >=0.6.0.5,<0.6.1.0a0, but none of the providers can be installed

Could not solve for environment specs
The following package could not be installed
└─ pytorch =2.8 cuda* is not installable because it requires
   ├─ libcudss0 >=0.4.0.2,<0.4.1.0a0 *, which requires
   │  └─ libcudss <0.0.0a0 *, which can be installed;
   └─ libtorch ==2.8.0 cuda129_generic_h4d1ebf7_200, which requires
      └─ libcudss >=0.6.0.5,<0.6.1.0a0 *, which conflicts with any installable versions previously reported.

It seems that pytorch requires libcudds0, while libtorch requires libcudss.

Installed packages

I can't install the environment.

Environment info

traversaro@IITBMP014LW012:~$ conda info

     active environment : base
    active env location : /home/traversaro/.pixi/envs/conda
            shell level : 1
       user config file : /home/traversaro/.condarc
 populated config files : /home/traversaro/.condarc
          conda version : 25.7.0
    conda-build version : not installed
         python version : 3.13.7.final.0
                 solver : libmamba (default)
       virtual packages : __archspec=1=aarch64
                          __conda=25.7.0=0
                          __cuda=12.3=0
                          __glibc=2.39=0
                          __linux=6.6.87.2=0
                          __unix=0=0
       base environment : /home/traversaro/.pixi/envs/conda  (writable)
      conda av data dir : /home/traversaro/.pixi/envs/conda/etc/conda
  conda av metadata url : None
           channel URLs : https://conda.anaconda.org/conda-forge/linux-aarch64
                          https://conda.anaconda.org/conda-forge/noarch
          package cache : /home/traversaro/.pixi/envs/conda/pkgs
                          /home/traversaro/.conda/pkgs
       envs directories : /home/traversaro/.pixi/envs/conda/envs
                          /home/traversaro/.conda/envs
               platform : linux-aarch64
             user-agent : conda/25.7.0 requests/2.32.5 CPython/3.13.7 Linux/6.6.87.2-microsoft-standard-WSL2 ubuntu/24.04.2 glibc/2.39 solver/libmamba conda-libmamba-solver/25.4.0 libmambapy/2.3.2
                UID:GID : 1002:1002
             netrc file : None
           offline mode : False

[h-vetinari]

This reminds me of conda-forge/conda-forge-repodata-patches-feedstock#1081 / conda-forge/torchvision-feedstock#123

[weiji14]

Looking at https://anaconda.org/conda-forge/pytorch/files?version=2.8.0, it seems like the linux-aarch64 builds of pytorch-2.8.0 (build_number 200) is depending on libcudss0 instead of libcudss like the linux-64 variants. Possibly related to conda-forge/torchvision-feedstock#123 (comment) which said:

libcudss0 was pulled in because libcudss-dev 0.4 was used, for some reason, instead of 0.6. I'm not sure why the solver decided to go with an older version of libcudss-dev

Not sure if this can be fixed by a repodata-patch. Might need to wait for the next pytorch rebuild on conda-forge?

[traversaro]

If the ABI between libcudss 0.4 and 0.6 is indeed different, this can't be solved by a repodata patch but we need a rebuild of pytorch 2.8.0 linux-aarch64 cuda builds.

[mgorny]

Well, "good news" is that we'll be building new packages due to #424 anyway.

[traversaro]

I briefly checked #425 and it seems that all builds there are using libcudss 0.7.0, so this issue should be solved by that PR.

[traversaro]

By the way, a bit of but it seems that libcudss-dev have quite strict run_exports pin: https://prefix.dev/channels/conda-forge/packages/libcudss-dev . I wonder if it make sense to add it to the conda-forge-pinnings.

[h-vetinari]

I wonder if it make sense to add [libcudss-dev] to the conda-forge-pinnings.

That would be great!

[traversaro]

Problem solved, see conda-forge/torchcodec-feedstock#35 .