cudnn 9 migration???

[hmaarrfk]

Comment:

i'm getting a similar error conda-forge/pytorch-cpu-feedstock#259

How can i get around this? I haven't seen this i other CUDNN feedstocks.

    Encountered problems while solving:
      - nothing provides __glibc >=2.28,<3.0.a0 needed by cudnn-9.2.1.18-h0e55af2_0

    Could not solve for environment specs
    The following packages are incompatible
    ├─ cuda-version 12.0.*  is requested and can be installed;
    └─ cudnn 9.*  is not installable because there are no viable options
       ├─ cudnn 9.2.1.18 would require
       │  └─ cuda-version >=11.0,<12.0a0 , which conflicts with any installable versions previously reported;
       └─ cudnn 9.2.1.18 would require
          └─ __glibc >=2.28,<3.0.a0 , which is missing on the system.

[billysuh7]

@hmaarrfk it appears you tried to install on linux-aarch64 (based on h0e55af2_0 ). Would you let me know:

1. exact command you issued
2. output of "conda config --show channels"
3. output of conda "config --show channel_priority"
4. output of "conda list" in your environment

[hmaarrfk]

With the help of h-vetinari, I think i got things working.

The thing is that the default images do not have stdlib 2.28 and require you to specify a newer linux image for compilation.

Eventually we had to add
https://github.com/conda-forge/pytorch-cpu-feedstock/blob/main/recipe/conda_build_config.yaml#L7

I think ultimately, things are moving forward with some opt-in mechanisms to allow alma8 which should alleviate this
conda-forge/docker-images#287

But today, if people want to use CUDNN9 and aarch, they must specify that giant matrix.

[jakirkham]

Thanks Billy and Mark! 🙏

Yeah we need to make it easier for folks to grab the new image and stdlib. Started some work in this direction with PR: conda-forge/conda-forge-pinning-feedstock#6548

Please feel free to try it out and share feedback there

[jakirkham]

In conda-forge, we have moved to the latest Docker image for source builds (currently Alma Linux 9): conda-forge/conda-forge-pinning-feedstock#6626

Also cuDNN has updated its GLIBC requirement in PR: #96

Plus GLIBC compatibility is tested for with the logic added in PR: #97

Think these collectively have resolved the issue above. So will close this out

Please let us know if any other issues come up