I'm trying to start working with Nvidia's Isaac Gym software in a new Conda environment and am running into some issues.
Firstly, the software requires a Conda environment of Python 3.8.
Secondly, the software requires Pytorch be installed and functional with my GPU. I have a 3080 Ti, which means Pytorch needs to be installed with CUDA 11.3 (or, at least, newer than 10.2). As Isaac Gym's setup installs it with CUDA 10.2, I need to install Pytorch elsewhere.
Whenever I run the CUDA 11.3 Pytorch install in the Conda environment as described by the menu on https://pytorch.org/, it bumps the Python version in the Conda environment.
My workflow is described below:
1. Create the Conda environment:
conda create --name isaacEnv python=3.8
2. Activate the Conda environment:
conda activate isaacEnv
3. Check Python version, which returns Python 3.8.13
python --version
This returns Python 3.8.13
4. Install the rest of the Nvidia Isaac Gym stuff.
This step doesn't impact any later steps but is detailed in their installation documentation (downloadable from their website). It does not end up mattering because the script they use installs torch with Cuda 10.2, which--as stated--is incompatible with my 3080 Ti.
This step is also likely independent of the problem, as the next step causes issues regardless:
5. Install Pytorch with CUDA 11.3
As detailed on pytorch.org:
conda install pytorch torchvision torchaudio cudatoolkit=11.3 -c pytorch
At this stage, even if I hit "no" to cancel the installation, when I run python --version again, it will show Python 3.9.12, which is the default version on this PC.
tl;dr: Is there a way to install Pytorch with CUDA 11.3 without upgrading from Python 3.8 in the Conda environment in the process?
I have faced the same question today.
It seems like you just need to open another terminal, and the python version will revert to normal... Maybe it's something about CUDA.
Update: I have considered this question again. Now I tend to that this is a bug with conda updating. After I created the new env, I run an updating just as what conda advised me to do. And since then the python attaching to my terminal was always the one in base env.