Run cuda kernel on tvm runtime

Hi, I tried to use auto-scheduler to tune some custom cuda ops for a tensorflow network, but encountered into a problem that tensorflow requires some of the the kernel function parameter to be const pointers. Dirty method modifying the .cu file works but tiresome. If anyone knows that will methods below work?

Not output the tuned .cu file and compile with nvcc, just run the tuned cuda kernel on tvm runtime, or Add somewhat configuration to my auto-scheduler python code?