Release 0.9 is providing our first MNMG algorithms. It has taken a few iterations to get a reasonable abstraction together to enable the building of different MNMG algorithms, but now that we have something, we should put it in a developer guide.
Some key concepts-
General overview of the cumlCommunicator, cuml.dask packages and MG paradigms
Differences between algorithms architected for MPI and for Dask/Spark.
Data distributions, guarantees, and formats
General algorithm design guidelines & considerations (from prims all the way to Python)
Release 0.9 is providing our first MNMG algorithms. It has taken a few iterations to get a reasonable abstraction together to enable the building of different MNMG algorithms, but now that we have something, we should put it in a developer guide.
Some key concepts-
cumlCommunicator
,cuml.dask
packages and MG paradigms