Dynamic linking of SIMD wrapper library for different architectures?

Question

This stackoverflow hints that the Eigen vector library will need to know the targeted ISA at compile time to compile the SIMD intrinsics. As we are running our application on different servers using different intel ISA's, some supporting only AVX2 while others support avx512, what is the advised way to go about if we want to continually deploy or app on the whole server cluster? I have identified two ways but maybe there are more:

• We could wrap Eigen in two precompiled shared lib where one covers the avx2 and the other the avx512 case, picking at runtime (since it's a header only library)
• We could statically compile the whole application for two ISA's and hand it out individually.

Are there more, maybe better alternatives?