Accelerated multiplications with kernel matrices.
native/: CUDA/C++ code, separated into acommon/folder and several operators.matmul/contains the main KernelMatmul implementation.kernel_matmul/: Python library. This code is responsible for the on-demand compilation and autotuning of the CUDA/C++ operators (native_function.py,configurations.py), the sparsity pattern calculation (ranges.py), and contains the implementation of the GPyTorch functionality (gpytorch/andlinear_operator.py).examples/: Contains two simple examples demonstrating the use of the library.tests/: Contains the extensive unit test suite for this library. The tests innative/ensure correct functionality of the CUDA/C++ implementations, while thetest_linear_operator.pyensures correctness with the LinearOperator test suite from GPyTorch.
Note that the GPyTorch integration is available only if the corresponding extra (gpytorch or just linear_operator) is activated.
This project has the following dependencies:
- NVIDIA CUDA 12.1
- Python 3.11
- GPyTorch 1.11
- PyTorch 2.2.2
- And many other Python packages (see below)
The development environment is powered by devcontainers.
With VSCode, Docker and the devcontainer extension, you can just open this folder with VSCode and will be prompted to download, build and open the container.
The environment is defined by the files in the .devcontainer directory.
This config assumes a working NVIDIA Container Toolkit installation.
Note: Developing inside a container provides us with a reproducible environment which is similar on all platforms (e.g., with the correct CUDA and Python version). This reduces compatibility issues once the initial devcontainer installation is completed.
The project's dependencies are managed with Poetry. Poetry is already installed in the container. You can install the Python dependencies using the following command:
poetry installThis project uses pre-commit (configured in .pre-commit-config.yml) for running formatters and linters.
To install it as a Git commit hook, run
pre-commit install
inside the container. You can also invoke it manually with the following command:
pre-commit run
We provide a default VSCode configuration in .vscode.
This refers to a folder of include paths named include in the project root.
You can create this folder using the provided script find_include_paths.py:
python find_include_paths.pyThis script creates the include directory, queries Torch and Python for the include paths and symlinks them to the include directory.
You can then use the VSCode configuration to work with the C++/CUDA code.
We use pytest for unit tests. Just run it:
pytestThis will likely take a long time to complete, because of the compilation times for all the operator variants.