Hi all,
What is the last PyTorch version to support CUDA10.1 when compiling from source?
The way I understand it is that the latest few versions of PyTorch do require CUDA10.2, even when compiling from source. I’m building a software plugin that needs to work with the host software which is compiled against CUDA10.1 and cuDNN 8.0.5. So to play together nicely I need to match these CUDA/cuDNN versions, but do at the same time want to use a PyTorch version as modern as possible.
Cheers,
David
The CUDA version was bumped to 10.2 as a requirement in PyTorch 1.10.0 if I’m not mistaken.
You could thus use pre-1.10.0 or check if the build process would still work with 10.1.
Thanks @ptrblck !
I’ve tried building PyTorch 1.11 under Linux using CUDA10.1 and it did actually work (to my happy surprise). But it doesn’t work under Windows unfortunately. I’m getting some compile error for “CUDASparseDescriptors.h” there. I guess there is nothing to do in that situation, right? Other than using a pre-1.10.0 as you suggest?
Cheers,
David
I found the commit that changed the requirements:
Does that mean that up to that commit should work with CUDA10.1?
Yes, I guess so as this is the first commit I can find removing this CUDA requirement.
