This Dockerfile serves the latest possible pytorch (v1.10.2) and torchvision (v0.11.3) that could be compiled for the sm_30 NVIDIA architecture (compute capability 3.0).

CUDA 10.2 cuDNN 7.

Pre-built images are served on the dockerhub.

docker run -it --gpus all dizcza/pytorch-sm30 python

>>> import torch
>>> torch.__version__
'1.10.2'
>>> torch.cuda.get_device_capability()
(3, 0)
>>> torch.randn(5).cuda()
tensor([ 0.8824, -0.0490,  2.0234, -1.7939,  0.6414], device='cuda:0')

Note: it'll take 3+ hours to build the image.

docker build -t pytorch-sm30 .

Then

docker run -it --gpus all pytorch-sm30 python