Would it be possible to document the runtime library requirements of TensorRT? In other words, exactly what other libraries are needed by TensorRT at runtime? I can think of CuDNN, CuBLAS, NVRTC. Are there others? What versions are needed?
I cannot find this information via readelf -d libnvinfer.so | grep NEEDED, since it’s not exposed.
Motivation: some advanced build systems (e.g. Bazel) need to model the dependencies explicitly to make them available at runtime. In this case, it’s not possible to install these libraries under /usr/local/cuda nor is it possible to put them all in a folder and set the LD_LIBRARY_PATH to that folder.