Is there any way to insert thirdparty api call in Triton?

[juinshell]

Hello! Currently I need to use some NVIDIA Library API(like NVSHMEM) that I think not be support currently in Triton. So I want to know If the user needs to insert some custom or third-party api call (with parameters related to threadIdx or blockIdx) into triton cuda kernel, can this be done as efficiently and concisely as possible?

The followings are some references I found:

• https://github.com/triton-lang/triton/blob/main/python/triton/language/extra/cuda/utils.py
• for nvidia backend, will generate ir in the order ["ttir", "ttgir", "llir", "ptx", "cubin"], user may focus more on maintaining api call message in "ttir" and "ttgir", and do something when generating llir?(not for sure)

Thanks for any help anyone could give.

[juinshell]

https://github.com/triton-lang/triton/blob/main/python/triton/tools/compile.py can export CUBIN as a CUDA kernel which cannot be changed by the user. I thought the seem goes equally to export PTX. It may lead to complex errors in execution and be difficult to debug.