I didn’t try yet. But I think it could work because the patch was provided 2 months ago.
Is there an issue with the sm_120 support supplied by the tensorflow package in the repositories?...
For some reason when I use nvcc I got error stdlib.h not found....
To be clear, I have no real use-case for the binary, I just spotted the missing .so and reported that.
$ sudo pacman -Fy libhsa-runtime64.so...Is that not an option for you to use nvcc to compile CUDA code and clang++ or g++ to compile the host code? Are you asking for sm_120 support to be...
The problem won’t happen because export TF_CUDA_CLANG=0 exists....
Does it work if you build the current tensorflow PKGBUILD in a clean chroot?
For me it doesn’t work if I use sm_>90
The current tensorflow sets HERMETIC_CUDA_COMPUTE_CAPABILITIES up to sm_120 https://gitlab.archlinux.org/archlinux/packaging/packages/tensorflow/-/...
Yes I’ve seen it....
sm_120 was only added in [CUDA] Add support for sm101 and sm120 target architectures (#127187). Is there an issue with the sm_120 support supplied ...
This happen when I try to build tf-nightly with cuda 12.8.1
Worked around upstream. I'm subscribed to the LLVM issue but I don't see it getting fixed anytime soon. I don't want to add the Debian patch unless...