I seem to recall ~6 years ago I had to recompile an ML library because my CPU didn’t support AVX2. So even though the release binaries required AVX2, the source

I seem to recall ~6 years ago I had to recompile an ML library because my CPU didn’t support AVX2. So even though the release binaries required AVX2, the source code was written in a way that it could work without it. I wonder whether this required specific effort by the developers of that library, or whether it was taken care of by some underlying tooling? Can you write CUDA code today that uses the latest features but can be compiled for older cards as well?