Mo Zhou wrote:
The bug is largely triggered by g++-12 being too old. However, the latest version of g++ supported by CUDA 12.3 is g++-12.2. We need to wait for the nvidia team to upload CUDA 12.4 (they have planned so) to unstable, and bump the g++ symlink in nvidia-cuda-toolkit-gcc to g++-13. At that time the issue should be automatically resolved without code change in src:pytorch-cuda.
I guess what you call CUDA version is this:
nvidia-cuda-dev | 11.8.89~11.8.0-5~deb12u1 | stable/non-free | amd64, arm64, ppc64el
nvidia-cuda-dev | 12.2.140~12.2.2-2 | testing/non-free | amd64, arm64, ppc64el
nvidia-cuda-dev | 12.3.101~12.3.2-1 | unstable/non-free | amd64, arm64, ppc64el
nvidia-cuda-dev | 12.4.99~12.4.0-2 | experimental/non-free | amd64, arm64, ppc64el
so I'm going to file a serious bug on nvidia-cuda-dev so that it
does not propagate to testing, not even accidentally.
(If you have a better idea, please say so)