raft_dask#
2026-08-27
1 min read time
Applies to Linux
Warning
raft_dask is experimental. Testing is limited and it may not be fully validated.
raft_dask extends hipRAFT to multi-node, multi-GPU (MNMG) environments using Dask
for distributed task scheduling. It provides communicator abstractions built on RCCL
(AMD’s ROCm implementation of the NCCL collective communication API) and UCX/UCXX that
enable scalable collective operations (all-reduce, broadcast, gather, etc.) across GPU
clusters. raft_dask is intended for distributed workflows where computation must be
coordinated across multiple GPUs on one or more nodes.