Only static version of nvptxcompiler installed with CUDA 11.6 on Linux?

I have version 11.6.55 installed on my Devuan Linux system.

While I see both the static and dynamic versions of libraries such as nvrtc under /usr/local/cuda-11.6/lib64 - for libnvptxcompiler, I only see the static version(/usr/local/cuda-11.6/lib64/libnvptxcompiler_static.a). Why is that? Can I get the dynamic version?

It seems that library is needed for statically linking nvrtc codes, but not for dynamically linking nvrtc codes. Therefore even if a dynamic version were provided there don’t seem to be any build instructions that would include it.

It seems that the dynamic linking process provides all necessary support via some other provided path, such as the dynamically linked nvrtc library.

But I don’t need it for NVRTC, I need it directly - that’s what underlies the PTX compilation API:

https://docs.nvidia.com/cuda/ptx-compiler-api

Follow the provided build instructions. No dynamic nvptxcompiler library is needed following those build instructions, and those are the only documented/supported build instructions.

If you’d like to see a change to CUDA please file a RFE/bug.

Yes, I realize using the static library works, and the build instructions show us we can use it. My question is - why is this library static-only, while NVRTC, which does something similar, is provided both as a static and as a dynamic library? Anyway, I’ll file an RFE.

Filed: https://developer.nvidia.com/nvidia_bug/3772055.