How to insert CUDA kernel code/lib slice into a triton kernel?

[juinshell]

Hello, guys! I'm trying to fuse some cuda device api call to a triton gpu kernel. I found this way. However, libdevice.bc is llvm ir(nvvm ir) code, It is hard for me to generate nvvm ir from .cu file(refer) .

So I sincerely ask if there is a way to call code snippets of the cuda kernel in tirton, like getting threadblock index, some sync op, or thirdparty api like nvshmem.