How can we dynamically determine the CUDA driver version?

Since online services use inconsistent CUDA versions, yet a specific CUDA version must be specified at compile time, is it possible to support a compilation method that does not require explicitly specifying the CUDA version?