librtcx is always built as a static library. It supports both dynamic and static linkage to nvrtc and nvjitlink. We need to statically link both in Spark builds, but otherwise we typically want to dynamically link them both. We must also ensure that in all non-Spark distributions we also have the necessary dependencies set up so that these libraries are available at run time. Currently I think the libcudf wheel is missing the nvrtc dependency, and I am also unsure that we are properly dynamically linking in conda/wheels.
librtcx is always built as a static library. It supports both dynamic and static linkage to nvrtc and nvjitlink. We need to statically link both in Spark builds, but otherwise we typically want to dynamically link them both. We must also ensure that in all non-Spark distributions we also have the necessary dependencies set up so that these libraries are available at run time. Currently I think the libcudf wheel is missing the nvrtc dependency, and I am also unsure that we are properly dynamically linking in conda/wheels.