Out-Of-Box Native GPU Allreduce

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

To view the NVLink topology, run nvidia-smi topo -m

To activate this feature, make sure to enable the hierarchical component in UCC using the UCC_CLS environment variable as follows:

            

            
UCC_CLS=basic,hier.
        

To view all available UCC items and options, run ucc_info -f