What does cuda.is_initialized() actually check?

The documentation for the function torch.cuda.is_initialized() tells me

"Returns whether PyTorch’s CUDA state has been initialized."

but what does that actually mean and when/for what do I need that information?

If I run this on a couple of different machines which support CUDA and where pytorch runs perfectly fine, I always get the return value False.

So: what exactly is this function telling me?

is_initialized() checks if the internal _initialized flag was already set to true from _lazy_init() and also if the current process is in a “bad fork”, i.e. if you are trying to re-initialize CUDA in forked subprocess.