OpenDriveLab
Using python 3.10 and cuda 11.8
Could you let me know if it is possible to set up the UniAD project with Python 3.10 and CUDA toolkit 11.8, even though the default setup is Python 3.8 and CUDA toolkit 11.1?
Based on my experiments, it appears that it is at least not feasible to run UniAD on newer GPUs with compute capability SM_89 or higher. I have attempted to use UniAD on both an H800 (SM_90) and a 4060Ti (SM_89) with CUDA 11.8 (the minimum version required for these GPUs) and CUDA 12.2, as well as PyTorch 2.x and Python 3.10. All attempts have failed.
The primary issue lies in the compatibility between UniAD and mmcv-full. The highest supported version of mmcv-full 1.x (required by UniAD) is 1.7.2, which is only compiled for PyTorch 1.x. However, the highest version of PyTorch 1.x is 1.13.2, which is compatible with CUDA 11.7 at most.
Thus, if you aim to use UniAD with these newer GPUs, significant modifications will be necessary. You will need to adapt UniAD to work with mmcv-full 2.x, and this may require additional substantial adjustments to the codebase to ensure compatibility.
Is it possible to run UniAD on Nvidia A6000 Ada gen GPUs? The minimum cuda version specified is 11.6
