Is there a table which shows the supported cuda version for every pytorch version

Is there a table which shows the supported cuda version for every pytorch version?
Thanks.

1 Like

You can build PyTorch from source with any CUDA version >=9.2 and the binaries ship with the mentioned CUDA versions from the install selection.

1 Like

If I only have cuda9.0, what’s the newest pytroch I can use?

I think 1.4 would be the last PyTorch version supporting CUDA9.0.
Note that you don’t need a local CUDA toolkit, if you install the conda binaries or pip wheels, as they will ship with the CUDA runtime. You would only have to make sure the NVIDIA driver is updated to the needed version corresponding to the CUDA runtime version.

1 Like

Thank you for your reply.

If I have to build some additional libraries, such as detectron2, do I need a local CUDA toolkit?

If these 3rd party packages are shipping with CUDA kernels, then you would need to install a local CUDA toolkit, yes.

1 Like

Would having the latest Nvidia driver do this, or are you saying that I should downgrade my Nvidia driver to match whatever CUDA version Anaconda is currently using?

2 Likes

hi, i hope you are well,
actually i have installed cuda 11.8 but i cant find the suitable version of pytorch which is compatible with cuda 11.8 and customized for pytorch. what should i do instead?

You can just select CUDA 11.8 in the install matrix and run the created command: