Dev pytorch from forge

[Old-Shatterhand]

Description

The current dev branch cannot be published on conda-forge due to PyTorch not being installed from conda-forge but from the pytorch channel. Therefore, I changed the PyTorch installation channel from PyTorch to conda-forge in the CI/CD pipeline to test if the test succeeds with the new dependencies.

Todos

• [ ] If the tests succeed, one has to adjust the conda-forge upload to use PyTorch from conda-forge.

Questions

No questions so far.

Status

• [ ] Ready to go

[Old-Shatterhand]

In the first round, the conda setup failed (Link to CI/CD)

I have no clear idea, why this breaks. Maybe, for every run, a cached environment is updated instead of a new one is created? Or mamba uses cached builds from previous downloads? Maybe, @dominiquesydow or @mbackenkoehler can help? My only guess is to rename the environments in the CI/CD pipeline and in devtools/test_env.yml. Let's see if that helps.

[mbackenkoehler]

@Old-Shatterhand There is no caching of conda environments or packages.

Pasting the error from the raw logs here:

2023-06-06T09:04:25.6344822Z Could not solve for environment specs
2023-06-06T09:04:25.6345202Z The following packages are incompatible
2023-06-06T09:04:25.6345856Z ├─ mkl <2022  is requested and can be installed;
2023-06-06T09:04:25.6346344Z └─ pytorch 1.13.0  is uninstallable because there are no viable options
2023-06-06T09:04:25.6346750Z    ├─ pytorch 1.13.0 would require
2023-06-06T09:04:25.6347273Z    │  └─ mkl >=2022.2.1,<2023.0a0 , which conflicts with any installable versions previously reported;
2023-06-06T09:04:25.6347694Z    └─ pytorch 1.13.0 would require
2023-06-06T09:04:25.6348044Z       └─ __cuda  , which is missing on the system.

[Old-Shatterhand]

Hm, then this run will not succeed as well...