Programming Model

Programming Model#

The programming model for oneCCL describes how to:

  • Set up distributed computations.

  • Perform collective communication operations (for example, ALLREDUCE, BROADCAST, ALLGATHER).

See also

See oneCCL specification that oneCCL is based on.

oneCCL supports a single rank/process per GPU device. The current implementation does not yet support a single process opening multiple devices.

Review the oneCCL generic workflow in the specification before getting started with the communication operations.

You can quickly get started with: