Changes and New Features

TL/UCP Special Service Worker

Added support for having a separate UCX UCP worker use UCC service collectives.

For further information, please see TL/UCP Special Service Worker section.

Data Type Support in CUDA Executor Component (EC)

Added out-of-box support for all datatypes and reduction operations for UCC collectives for GPUs.

For further information, please see Data Type Support in CUDA Executor Component section.

EC/CUDA One-shot Kernel with Cooperative Launch

Added support for using a single CUDA kernel for CUDA operations in UCC GPU collectives.

For further information, please see EC/CUDA One-shot Kernel with Cooperative Launch section.

Out-Of-Box Native GPU Allreduce

Added support for the UCC library to detect the NVIDIA NVLink topology and select the best GPU-based algorithms for supported collectives (Allgather/v, Reducescatter/v).

For further information, please see Out-Of-Box Native GPU Allreduce section.

Bug Fixes

See 

.