I’m running the Nvidia cuda python library inside a GitHub Action on a GPU hosted runner. The library works and detects the Nvidia T4 device just fine, demonstrating that the library is installed and working appropriately.
However, any attempt to use the NVRTC CUDA library to compile CUDA code results in an error that libnvrtc.so.12 is missing. This file supposed to be installed as part of Cuda Toolkit, which leads me to believe that the file wasn’t installed or is installed in the wrong place (e.g. not in $CUDA_HOME or not in $LD_LIBRARY_PATH).
I know that NVRTC is a lesser-used feature of CUDA and the GitHub GPU runners are still newish, so there’s a pretty good chance this is a bug.
I’m reaching out here because this is where GitHub support pointed me for a “partner image”.