I am using an older Python 3.7 code, in which TensorFlow, Pytorch and Spacy all run in the same environment.
With Python 3.7, TensorFlow Cuda version is upper-limited by 11.2.
I am aware that Cuda 11.2 officially supports compute capabilities up to Sm_86 (Ampere).
However, I have Nvidia hardware with compute capability Sm_89 (Ada).
It seems that it’s possible to buld pytorch from source with Cuda 11.2 supporting Ada as follows:
TORCH_CUDA_ARCH_LIST="8.0 8.6+PTX" python3 setup.py install
How can I force Cuda 11.2+Cudnn 8.1to work with Ada? Perhaps, by masking Ada as if it were Ampere? Ideally, if system-wide.
Ubuntu 22.04, Miniconda with conda environment Python 3.7.17