Matthias Fey

For Mac, the `CUDA` code path should never get called, so it is very surprising to me it searches for `nvcc`. What happens if you try to install via ```...

You should run ``` conda install pytorch=2.2.1 pytorch-cuda=11.8 -c pytorch -c nvidia pip install pyg_lib torch_scatter torch_sparse -f https://data.pyg.org/whl/torch-2.2.0+cu118.html ```

Got it, sorry :) As a workaround, you can do manual installation of `pyg-lib` via ``` pip install ninja wheel cmake pip install git+https://github.com/pyg-team/pyg-lib.git ``` It looks like pre-built `pyg-lib`...

Do you want to build with CUDA or without? it looks for a CUDA implementation but cannot find it. If you want to install without CUDA, try ``` FORCE_CUDA=0 pip...

Please see my comment in https://github.com/pyg-team/pyg-lib/issues/211#issuecomment-1478934645.

I hit the same issue today :( I tried many different things, but in all scenarios the final binary was searching for a Python folder in `/Library/Frameworks/Python.framework/Versions`. Installing it from...

@DamianSzwichtenberg

Yeah, that's because we don't yet provide pre-built M1 wheels, and `pyg-lib` is not yet on PyPi. You can try to run via ``` pip install git+https://github.com/pyg-team/pyg-lib.git

You need to install PyTorch beforehand before being able to install `pyg-lib` from source.

I am a bit confused. We currently don't use `pyproject.toml` for `pyg-lib`. Not totally sure why it complains about it TBH.