Some preliminary Conda packages can be installed as so. Replace <CUDA version> with either 9.2 or 10.0.
conda create -n ucx -c conda-forge -c jakirkham/label/ucx cudatoolkit=<CUDA version> ucx-proc=*=gpu ucx ucx-py python=3.7
The ucx recipe can be found here: https://github.com/conda-forge/ucx-split-feedstock/tree/f13e882cc0566e795ff12f2a039f490ce1653698/recipe
The following instructions assume you'll be using ucx-py on a CUDA enabled system. The instructions assume you're using CUDA 9.2 for unspecific reasons. Change the CUDA_HOME environment variable, and the environment created and used by conda to cudf_dev_10.0.yml in order to support CUDA 10.
These three libraries provide a powerful combination of HPC message passing tools. Using them involves using the correct dependencies, in the correct order:
git clone git@github.com:rapidsai/cudf.git
cd cudf
export CUDA_HOME=/usr/local/cuda-9.2
export CUDACXX=$CUDA_HOME/bin/nvcc
conda env create --name cudf_dev_92 --file conda/environments/cudf_dev_cuda9.2.yml
conda activate cudf_dev_92
./build.sh
cd ..
git clone git@github.com:rapidsai/dask.git
cd dask
pip install -e .
cd ..
git clone git@github.com:dask/distributed.git
cd distributed
pip install -e .
cd ..
conda install -c rapidsai dask-cuda
conda install -c conda-forge automake make cmake libtool pkg-config pytest-asyncio cupy distributed
git clone https://github.com/openucx/ucx
cd ucx
git remote add Akshay-Venkatesh git@github.com:Akshay-Venkatesh/ucx.git
git remote update Akshay-Venkatesh
git checkout ucx-cuda
./autogen.sh
mkdir build
cd build
../configure --prefix=$CONDA_PREFIX --enable-debug --with-cuda=$CUDA_HOME --enable-mt --disable-cma CPPFLAGS="-I//$CUDA_HOME/include"
make -j install
cd ../..
git clone git@github.com:rapidsai/ucx-py.git
cd ucx-py
export UCX_PATH=$CONDA_PREFIX
make install
You should be done! Test the result of your build with
pytest -v