dizcza / pytorch-sm30-docker

PyTorch 1.10.2 Dockerfile for CUDA capability 3.0

Pytorch Dockerfile for graphics cards capability 3.0

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.

Usage

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')

Local build

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

docker build -t pytorch-sm30 .

Then

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