Building from GIT error: PyTorch needs CUDNN-8.5 or above but found 8.2.4

torch.compile
1

Hi,
I am trying to compile PyTorch on older PC with NVidia GT 730. This card requires CUDA-11.4 as newer CUDAs no more support this card. CUDNN-8.2.4 matches to CUDA-11.4 therefore I am not able to use newer version of CUDNN.

How can I “downgrade” the PyTorch to be able to build with CUDA-11.4/CUDNN-8.2.4 setup ?

Thank you for help.

2

Try to git checkout an old release where Kepler GPUs (sm_35) were still supported.

3

Hi, thank you for help. I somehow went through (actually switched off the CUDNN for a moment) but got the cmake error. Tried to google it but no success:

CMake Error at cmake/public/cuda.cmake:262 (set_property):
The link interface of target “torch::cufile” contains:

CUDA::cuFile

but the target was not found. Possible reasons include:

* There is a typo in the target name.
* A find_package call is missing for an IMPORTED target.
* An ALIAS target is missing.

Call Stack (most recent call first):
cmake/Dependencies.cmake:44 (include)
CMakeLists.txt:865 (include)

I was not able to find the CUDA::cuFile target anywhere. What am I missing? Thank you.

4

The PyTorch version you are trying to build might be too new for the CUDA toolkit you have installed (I don’t know which versions you are trying to build now). You could try to disable GDS via USE_CUFILE=0 or could try my previous suggestion to downgrade PyTorch to an old version.

5

Hi ptrblck,
to be honest I gave up and bought newer NVidia card (Turing) and torch can work with CUDA now even without compilation. I was not able to find which older release to checkout and compile. I am sorry that I wasted your time. Just imagine how your father starts to install the Linux on his computer after 15 years of sitting behind Windows as a user. Funny picture, isn’t it ?