Which CUDA driver APIs or CUDA runtime APIs Pytorch calls upon during training or predicting

Hello developers,
May I ask which CUDA driver APIs or CUDA runtime APIs Pytorch calls upon during training or predicting?

You could search for all cuda* and cu* calls in the code base, profile your workload and check the kernel calls in e.g. Nsight Systems, or be more specific with your question in case you are interested in a few specific calls.