Compiling earlier PyTorch versions on Blackwell

1

It seems that if you have a 5000 series GPU (Blackwell) with Compute Capability 12.0, you are forced to use CUDA 12.8. Is it possible to compile older PyTorch versions (specifically 1.13 or 2.0) using CUDA 12.8 ? Unfortunately, I have some software that strictly requires 1.13 and Python 3.8.

I published this as an issue Compile earlier PyTorch versions on Blackwell · Issue #150034 · pytorch/pytorch · GitHub, but it is unfortunately not planned. Therefore, I am looking for some guidance in the community. Has anyone succeeded ?

2

Already released binaries won’t be rebuilt again as was already mentioned in the GitHub issue. I would probably not bother trying to backport all needed changes into ancient PyTorch branches but would instead check why your repositories require these old versions and update them.

3

Thanks for the quick answer, unfortunately it’s Vitis-AI which is stuck at 1.13 or 2.0.0 and Python 3.8 and I dont believe I can do much, but I will try, perhaps I can meet somewhere in the middle