Hi Jakob,
On 11/10/2023 12:02, Strube, Alexandre wrote:
Hi Jakob,
I saw that there is a TensorFlow config for foss/2023a, but not with CUDA
support. Looking at earlier versions, it looked like the only difference
between a non-CUDA and a CUDA version was adding the relevant dependencies, so
I tried that. However, if it were that easy somebody else would already have
done it :-)
Tensorflow fails to compile. When I google it, others have seen the same
error, and it's because it only supports CUDA 11.8, not CUDA 12.1.1. So I
tried to change to that version, and make the few dependencies with that
version too. No luck: CUDA 11.8 does not support GCC 12.X. Now I am clearly
in too deep water... And I assume there is no CUDA-supporting PyTorch for
similar reasons.
Is it CUDA 11.8 that doesn't support GCC 12.x, or the other way around?
We merged a PR today that could maybe be relevant here (not entirely
sure though):
https://github.com/easybuilders/easybuild-easyconfigs/pull/18854
regards,
Kenneth
How do you handle stuff like that in EasyBuild? Is it just a matter of waiting
till a new version of TensorFlow and PyTorch support CUDA 12.X? Or are there
some experts that hack stuff until it works?
So, you are right - and sometimes we patch things, and in other times, we wait.
Numba is a case where they always have stuff way behind and we try to patch it
here.
If tensorflow doesn’t support cuda 12 and cuda 12 doesn’t support gcc 12,
you’re in bad luck - I would not recommend patching this monstrosity.
In fact, considering that Tensorflow is almost abandonware at this point, I
would just use the previous toolchain for it.
Alex
