[DOC]: instructions to compile PTX compatible with system version of cuda-toolkit #3379

Is this a duplicate?

• I confirmed there appear to be no duplicate issues for this bug and that I agree to the Code of Conduct

Is this for new documentation, or an update to existing docs?

New

Describe the incorrect/future/missing documentation

Hi, is there guidance available on how to install cuda-core with the cuda-toolkit included, without letting the cuda-toolkit version exceed the system's installed version?

I'm regularly running into trouble because the version of cuda-toolkit installed alongside cuda-core[cu13] exceeds the system version. Practically, what this means is that when I compile PTX code via cuda-core's Python API, I cannot run that PTX code unless the system drivers are up-to-date. I see two solutions

1. Update system CUDA version.
2. Restrict the installed cuda-toolkit version to a sufficiently low version number.

It's not clear to me what the recommended strategy is to resolve this. When compiling PTX with cuda-core, I already target the sm_... architecture of my card, but that is not enough. It seems I also need to restrict the CTK version as well, somehow.

Advice appreciated. I think this ought to be documented somehere, or a more intuitive solution be present.

If this is a correction, please provide a link to the incorrect documentation. If this is a new documentation request, please link to where you have looked.

• https://github.com/NVIDIA/cuda-python/discussions?discussions_q=is%3Aopen+toolkit+version
• https://nvidia.github.io/cuda-python/cuda-core/latest/search.html?q=toolkit+version#
• https://docs.nvidia.com/cuda/cuda-programming-guide/02-basics/nvcc.html#nvcc-ptx-and-cubin-generation