PyTorch support CUDNN question ( testing Cuda version vs GPU )

Hi all, I am new to PyTorch.

Just recently installed it on my laptop with GeForce GT 550M card.
pytorch-cpu install and runs with no problems. Can’t install the latest version of Cuda, because it says, your GPU is two old ( I installed cuda 9 separately and it’s OK ). One of the participants here advised to install pytorch with lesser version of cuda using

**conda install pytorch torchvision cuda80 -c pytorch**

This time installation was OK and my example actually started in PyCharm , but pretty soon it failed with Error message that my card supports only Cuda 2.1

Few questions:

  1. Is there anyway to reach to PyTorch developers , customer support etc, or the ONLY way is to publish questions on this forum ?

  2. How does CudNN determine which cuda version is supported by my GPU and is it possible to make this check to be ignored by setting some option/variable ? My GPU is indeed old, but as I mentioned cuda 9 installed separately works just fine.

  3. If I install PyTorch from source and link it with Cuda 9 I already have, would it be staill failing inside CudNN on version vs GPU test ?

Thanks a lot

This GPU is too old to run PyTorch. To give you a sense for how old it is, last year ( we removed support for 3.0, the earliest Kepler; your GPU is the microarch version before that.

Hi Edward ,

Thanks for your response. I understand the motivations for removal it : you can’t support every old version, My only surprise is , that if the chip is capable to run Cuda 9 separately, even if “officially” only older version of Cuda is supported, I assume you (CudNN) don’t need to do anything special at all. But, of course, I don’t know constraint/limitations of your product. As I wrote , I realize now that even if I install pytorch from source and somehow “link” my existing cuda install, it will fail in CudNN