Dask-CUDA can be installed using
pip, or from source.
To use Dask-CUDA on your system, you will need:
NVIDIA drivers for your GPU; see NVIDIA Driver Installation Quickstart Guide for installation instructions
A version of NVIDIA CUDA Toolkit compatible with the installed driver version; see Table 1 of CUDA Compatibility – Binary Compatibility for an overview of CUDA Toolkit driver requirements
Once the proper CUDA Toolkit version has been determined, it can be installed using along with Dask-CUDA using
To install the latest version of Dask-CUDA along with CUDA Toolkit 11.0:
conda install -c rapidsai -c nvidia -c conda-forge dask-cuda cudatoolkit=11.0
When working outside of a Conda environment, CUDA Toolkit can be downloaded and installed from NVIDIA’s website; this package also contains the required NVIDIA drivers. To install the latest version of Dask-CUDA:
python -m pip install dask-cuda
To install Dask-CUDA from source, the source code repository must be cloned from GitHub:
git clone https://github.com/rapidsai/dask-cuda.git cd dask-cuda python -m pip install .
Other RAPIDS libraries¶
Dask-CUDA is a part of the RAPIDS suite of open-source software libraries for GPU-accelerated data science, and works well in conjunction with them. See RAPIDS – Getting Started for instructions on how to install these libraries. Keep in mind that these libraries will require:
At least one CUDA-compliant GPU
A system installation of CUDA