Is it possible to use rtx 3080 gpu with pytorch 1.6

1

I have an rtx 3080 gpu, using Windows 10. Is it possible to use this gpu with
pytorch version 1.6 and any version (as necessary) of CUDA?

The reason I am asking is using the command
conda install pytorch torchvision cudatoolkit=11 -c pytorch-nightly
(recommended from the github issue https://github.com/pytorch/pytorch/issues/45028)
resulted in an install of a 1.8 dev nightly build that was having errors I was not encountering previously with the torchvision.models.detection repository where they
seemed to have added a new file anchor_utils.py . The authors seem to have refactored the anchor generator to its own file I guess, where previously the anchor_generator class was in rpn.py.

Thanks

2

You could use the 1.6 binaries with e.g. CUDA10.2, which will then however JIT compile the CUDA kernels in the first CUDA operation and will thus add a massive overhead to your runs, so I would recommend to adapt the source code and use the nightly binaries with CUDA11.

1 Like
3

What is the latest stable build of pytorch that supports RTX 3080/ CUDA 11?

4

1.7.0 is the latest stable release, which ships with CUDA11.0.

2 Likes
5

Yeah. So it can be “interpreted” as RTX 30 Series are still not well compatible with relative “old” PyTorch versions (< 1.7.0). Am I right?

6

Yes, and since the framework is not forwards compatible the CUDA11 support will not be backported to already released versions.

1 Like