Skip to content
Python bindings for UCX
Python C Shell Makefile
Branch: devel
Clone or download
Akshay-Venkatesh Merge pull request #157 from rapidsai/minor-changes-in-benchmarks
add options to dask-dot-product; listener config in send-recv-py-obj
Latest commit b900016 Aug 21, 2019
Permalink
Type Name Latest commit message Commit time
Failed to load latest commit information.
benchmarks Merge pull request #157 from rapidsai/minor-changes-in-benchmarks Aug 21, 2019
conda-recipes
doc test note Feb 18, 2019
scripts pass IB interface name to starter script May 23, 2019
tests Update tests/utils.py Aug 20, 2019
ucp buffer.buf = NULL when length is zero Aug 18, 2019
.gitignore rename library to ucp and make installable Apr 22, 2019
Makefile fix Makefile Apr 22, 2019
README.md Update the commit used to build the recipes Aug 16, 2019
setup.py Properly pulling endpoint pointer from internal struct Jun 19, 2019
ucx-setup.sh start-all script changes to use user suggested CVD list May 22, 2019

README.md

Python Bindings for UCX

Installing preliminary Conda packages.

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

All of the recipes used can be found here: https://github.com/jakirkham/staged-recipes/tree/ad6f8c51e9b08f34b800b19e9e15dd80cee6f7ea/recipes

Building ucx-py from source.

Ensure you have the build dependencies installed

  1. UCX
  2. Cython
  3. CUDA (optional)

Then run

make install

to build with CUDA support. To build without CUDA, run

This assumes that UCX is available in the "standard" place (next to this directory) and that CUDA is avaiable in /usr/local/cuda. If not, specify them like

UCX_PATH=/path/to/ucx/ CUDA_PATH=/path/to/cuda make install
You can’t perform that action at this time.