NVIDIA RTX A4000 with cuda capability sm_86 is not compatible with current Pytorch installation

1

I am having cuda version of 12.4 in my GPU’s.
It says current Pytorch installation supports cuda capabilities sm_37, sm_50, sm_60, sm_70, sm_75.
What should I do now?

2

You have most likely installed an old PyTorch binary shipping with CUDA 10.2 (you can check it via torch.version.cuda) and would need to install any binary shipping with CUDA>=11. Your locally CUDA toolkit won’t be used unless you build PyTorch from source or a custom CUDA extension.

3

import torch
print(torch.version.cuda)
11.8

It is showing 11.8 and not 10.2.

4

All PyTorch binaries compiled with CUDA 11.8 support up to sm_90, so you might need to double check if multiple PyTorch builds are installed or something else is conflicting. E.g. this old 2.0.0+cu118 installation shows:

python -c "import torch; print(torch.__version__); print(torch.version.cuda); print(torch.cuda.get_arch_list())"
2.0.0+cu118
11.8
['sm_37', 'sm_50', 'sm_60', 'sm_70', 'sm_75', 'sm_80', 'sm_86', 'sm_90']

which supports the RTX A4000.