Getting started

Setting up the project

Option 1. For developers, we recommend build from source. Here's our build guide

Option 2. We have nightly pip wheel packages published on pypi.nvidia.com that works with PyTorch's nightly pip wheel. Meanwhile, pip wheel built against stable PyTorch is released on pypi.org. Package is named as nvfuser-cuNNN-torchXY, replace cuNNN & torchXY with the proper version, e.g. nvfuser-cu121-torch24.

Note that we are currently only releasing wheel for python 3.10 on linux x86_64.

Ensure you have the proper python installed on your system. Install pytorch nightly with desired cuda version, e.g.

pip install --pre torch torchvision torchaudio --index-url https://download.pytorch.org/whl/nightly/cu121

After this, you will need to install nvfuser pip wheel and patch the binary coming from pytorch installation. (Note that we are installing cu121 version, which needs to match the cuda version of your pytorch installation).

pip install --pre nvfuser-cu121 --extra-index-url https://pypi.nvidia.com

That's it. You should be able to use nvfuser along with pytorch now!

root@5ed0d864dbfa:/# python
Python 3.10.9 (main, May 18 2023, 01:38:23) [GCC 9.4.0] on linux
Type "help", "copyright", "credits" or "license" for more information.
>>> import nvfuser
>>>