Which versions should I install to get support for my 5070 rtx (12 gb vram) with CUDA v13? So far I am getting errors, even if I try to compile the sources myself…
Thanks in advance!
1 Like
You can install any PyTorch binary built with CUDA 12.8+ for Blackwell support.
I think, always use the latest version of CUDA is very dangerous, because it may cause some issue very very hard to solve, just become some small API changes… Solve these bug is really…
Maybe you can use the “nightly builds”, pip3 install --pre torch torchvision --index-url https://download.pytorch.org/whl/nightly/cu130, because MAYBE some of nightly builds perormances better in fitting new versions of CUDA than stable releases: pip3 install torch torchvision --index-url https://download.pytorch.org/whl/cu130
But nightly builds will cause some bad results ABSOLUTELY, like they may calc slowly, or they need some special DLLs. I think, the best plan, is ALWAYS WAIT FOR THE (NEXT) STABLE BUILD THAT MAY FIT YOU.
I disagree with this statement as generally the latest CUDA toolkit will perform the best for the latest hardware generation. Especially for Blackwell GPUs I would try our latest stack.
If you are seeing any issues, please let us know on GitHub.