how to use nvrtc to run a sm90 kernel

Hi

I want to use nvrtc to compile a sm90 kernel in runtime. The problem is that I don't have the kernel instance on host thus can't run to_underlying_arguments to get kernel param to launch the kernel.

When running a sm80 code, I can get it's params directly since Params is equal to Arguments, but for a sm90 kernel I needs to construct tma_load struct before I launch the kernel.

I see there is a nvrtc example in cutlass which use python to launch a nvcc progress to compile host api for the kernel such as initialize into a .so. But I have a c++ application and can't launch any child progress in it.

So is there anyway to runtime compile a sm90 kernel in my application. I see make_tma_copy is defined as CUDA_HOST_RTC which can be used in nvrtc, but it will call cuTensorMapEncodeTiled which can only called on host, and nvrtcCompileProgram can only compile a kernel function. Is there any example to use make_tma_copy in nvrtc, thanks.

Start with the CUTLASS NVRTC example and trace how it handles host-side setup. Compare the sm80 argument path with run_to_underlying_arguments, make_tma_copy, cuTensorMapEncodeTiled, and nvrtcCompileProgram. Done means a documented C++ runtime-compilation path for sm90, or a clearly stated limitation if the requested flow is unsupported.