Pytorch wheel package built for Nvidia Kepler GPUs. Currently supports newer Kepler GPUs (cuda compute capability 3.5 and 3.7) on Linux.
Since the Maxwell architecture was also deprecated by the official builds, this unofficial builds starting with version 2.5 added cuda compute capability 5.0 to the support as well.
to use this pre-build package, simply download whl file from the release, and install it using pip:
python -m pip install torch-2.1.0a0+git7bcf7da-cp39-cp39-linux_x86_64.whl # change whl package name to real file name.You may also need to install the cuda and cudnn runtime packages via pip or conda.
This pre-build package only supports Linux (only tested on Ubuntu). The supported architecture is amd64. If you want to use Kepler GPU on arm64 or other architecture, please follow this guide to build pytorch from source.
This pre-build package only supports CUDA cc3.5 and cc3.7. The Kepler architecture is already 13 years, and considering the practical performance, only the top-line chips are still usable today. Here is a brief description of these chips.
| Chip Name | CUDA CC | Product Name |
|---|---|---|
| GK 110 | 3.5 | GTX 780, GTX TITAN, Tesla K20 |
| GK 110B | 3.5 | GTX 780Ti, GTX TITAN BLACK, GTX TITAN Z, Tesla K40, Quadro 5200, Quadro 6000 |
| GK 180 | 3.5 | Tesla K40c |
| GK 210 | 3.7 | Tesla K80 |