Currently it is possible to select which CUDA runtime is used by adding `` to the CMake invocation. This is fine for simple test programs, but it has its issues. For example, it is hard to set this variable programmatically: * Using `string(APPEND CMAKE_CUDA_FLAGS " --cudart shared")` has simply no effect. The workaround is to make sure that CMAKE_CUDA_FLAGS already is a Cache variable by creating it first: `set(CMAKE_CUDA_FLAGS "" CACHE STRING "")` * Setting `CMAKE_CUDA_FLAGS` after a `project(name LANGUAGES CUDA)` or `enable_language(CUDA)` has no effect. It *must* be set before the project or enable_language call. * Assume I set this variable in a script (for example as a default value), but want to allow the user to override this. There is no easy way without parsing/regexing/stripping the CMAKE_CUDA_FLAGS to override it, since specifying it twice is forbidden by NVCC. I believe there should be a simple variable called `CMAKE_CUDA_USE_STATIC_CUDA_RUNTIME` or similar, that allows a simple way to switch between a dynamic and static runtimes.