After researching a lot on how to use PyTorch with a RTX 3060 card, specially with older versions or torch (0.4.0) and torchvision (0.2.1), I noticed that it was either impossible or very hard to do. RTX 3060 and these packages apparently doesn't have compatibility with the same versions of CUDA and cuDNN. I tried to do this by using different combinations with compiled versions available in conda, but didn't work, maybe it could work if you recompile from source these versions.
After all this, actually I was able to use RTX 3060 effectively with latest versions of all these dependencies with two methods:
- Using a conda env, and latest versions published in pytorch site (https://pytorch.org/get-started/locally):
conda create -n rtx_3060 python=3.6.5
conda activate rtx_3060
conda install pytorch torchvision torchaudio cudatoolkit=11.1 -c pytorch -c nvidia
- Using
lambda-stack(https://lambdalabs.com/blog/install-tensorflow-and-pytorch-on-rtx-30-series/):
LAMBDA_REPO=$(mktemp) && \
wget -O${LAMBDA_REPO} https://lambdalabs.com/static/misc/lambda-stack-repo.deb && \
sudo dpkg -i ${LAMBDA_REPO} && rm -f ${LAMBDA_REPO} && \
sudo apt-get update && sudo apt-get install -y lambda-stack-cuda
Reboot your machine.
LAMBDA_REPO=$(mktemp) && \
wget -O${LAMBDA_REPO} https://lambdalabs.com/static/misc/lambda-stack-repo.deb && \
sudo dpkg -i ${LAMBDA_REPO} && rm -f ${LAMBDA_REPO} && \
sudo apt-get update && \
sudo apt-get --yes upgrade && \
sudo apt-get install --yes --no-install-recommends lambda-server && \
sudo apt-get install --yes --no-install-recommends nvidia-headless-455 && \
sudo apt-get install --yes --no-install-recommends lambda-stack-cuda
Reboot your machine.
*** The good thing about this method is that, if you have your current global environment messed up, lambda-stack is going to actually fix it.
Related to:
Hello! I'm currently trying to get this to work with Pytorch 1.7.0 and CUDA 10.2, but am not able to verify CUDA availability with torch. Any suggestions?