Is there a way to specify the cuda toolkit version like 11.6 in the requirements?
I believe you have to create your own strategy to install the CUDA Toolkit version you desire for the environment.
Cuda toolkit can be installed via conda but not pip iirc, so you could add it to
All the requirements are combined in the setup.py so how would the GitHub actions differentiate between pip and conda installations?
conda_requires is only used for conda installs.
A minor technicality here is that the new releases of cuda (cuda >= 11.3.0) are found as “cuda-toolkit” (at “nvidia” channel) instead of the previous (cuda <=11.3.1) “cudatoolkit” (at default channel).
conda install -c “nvidia” cuda-toolkit==11.6
I don’t know if there is a way for the settings.ini to exchange channels during setup.
That technical issue can be overcome by adding the “nvidia” channel to conda prior to setup.
conda config --add channels nvidia
In that way, the conda install process can search all available channels registered in the machine.