Overview

Simple CMAKE make project for extending PyTorch with custom cuda kernels for doing all the things that I wanna try.

Build

New scikit-build core way

pip install scikit_build_core
pip install pyproject-metadata

# Install in editable mode
pip install -v --no-build-isolation -e .

This will invoke the pyproject.toml file and build the project with the scikit-build package. Including, calling cmake and building the shared library.

How to create a distributable wheel?

• Lets make a test env:

conda create -n test_env python=3.10
conda activate test_env

# Install nightly torch
pip3 install -U --pre torch --index-url https://download.pytorch.org/whl/nightly/cu121

• Lets run the build command

pip install .

This will invoke cmake which in turn will build the shared library and installed into driss_torch/lib/libdriss_torch.so

What does this do?

So this will go and build the libdriss_torch.so shared library. In which there will be a number of ops that have been registered. In the python package driss_torch there will be small wrappers around the cpp ops with better type hints and documentation.

Could use .ini but this more fun.

Tip (Maybe unsafe?)

The total install command from scratch can take a second. If you want a faster dev flow cd into build and once you change your sources of the lib just run ninja to rebuild the shared library. This will not work if you are adding and deleting files.

Intended Usage

You should be able to grab the ops by doing something like

import torch
import driss_torch

a = torch.randn(10, 10, device='cuda')
b = driss_torch.add_one(b)
print(b)

Install pre-commit hooks

pre-commit install

Running Tests

pytest test/

Old build way

This calls cmake directly in the build dir.

# In the root directory
mkdir build && cd build

# Configure the build, to build with compute arch
cmake -DCMAKE_BUILD_TYPE=RelWithDebInfo -GNinja  -DTORCH_CUDA_ARCH_LIST=9.0 ..