The right way to manage CUDA environment with conda.

There are several principal rules:

1. Conda manages Python distribution better than it manages python packages.
2. Always install all python packages via pip.

If you only need to use the runtime and don’t need tools such as nvcc, then you need to install cuda-toolkit. conda provides the relevant packages at https://anaconda.org/nvidia/cuda-toolkit, and you need to install the corresponding version of cuda-toolkit using a command like conda install -c nvidia/label/cuda-11.8.0 cuda-toolkit to install the corresponding version.

If you need to use nvcc, e.g. to compile a PyTorch plugin, then you need to install cuda, which is available from conda at https://anaconda.org/nvidia/cuda, and you need to install the corresponding version using a command like conda install -c nvidia/label/cuda-11.8.0 cuda cuda

According to this issue, you need to use the -c parameter to specify the channel to install the package, otherwise, if you follow the example on the anaconda website (using :: to specify the channel), it will definitely install version 12.4.