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

pytorch-sm30-docker's Issues

Build PyTorch 1.12.0 or 1.13.0 on CC 3.0 GPU?

Is PyTorch 1.10.0 the highest version to be used with CUDA 10.2 on an Nvidia Compute Capability 3.0 GPU?

Or one can still build from source of PyTorch 1.12.0 or even 1.13.0 to be used on a CC 3.0 card?

Official PyTorch site seems to indicate, PyTorch can be built from source for sm_30 (CC 3.0) graphic cards until version 1.13.0. Version 2.0.0 drops support for sm_30 and CUDA 10.2 and below.

Is this right?

Python version

thanks for your work
What is the python version in this Docker?

Thanks

Just a big thank you from me, that's all. I've been trying to get a working GPU accelerated Pytorch install for my sm_30 card for days and have had a total nightmare.

Just getting a working CUDA 10.2 installation on my Ubuntu 22.04 O/S has been a nightmare enough with dependency issues from both the Ubuntu and NVIDIA packages, which I could only resolve by using the 18.04 repo and a custom built dummy package to trick it into being happy with my driver and X server versions.

Then, having been reassured from multiple sources that building Pytorch from source for CUDA 10.2 with an sm_30 GPU was possible and finding out it's not as straight forward as I was lead to believe and a lot of misinformation on exactly what version supports it (if any), I came across your project, tested the image and it worked, proving to me that it WAS possible just at the point I was about to give up.

I'm not using your Docker image, but the Dockerfile has given me the information that I need - basically that I can build NCCL for an sm_30 architecture and then build Pytorch with USE_SYSTEM_NCCL to avoid the compilation error I was getting due to the lack of the __funnelshift_r() intrinsic in 3.0 GPUS.

I don't know how much work you put into this, but there's at least one person here who appreciates your effort because I have been pulling my hair out trying to put this jigsaw puzzle together.

I might be jinxing myself as the build is still working and I have had so many instances where I thought I'd managed to get somewhere only to be presented with another compilation error, and I still don't understand how CuDNN can work because NVIDIAs support matrix says "no" for sm_30, but it has at least allowed me to move forward on to my next challenge.

So - thank you!