llama.cpp is a generalized inference engine; therefore, model-specific optimizations should remain outside its core architecture. When developing a bespoke inference application, it is standard practice to fuse operations and heavily optimize the compute kernels for the specific target model.