Hello, I am running an alltoall_perf test with a stacked Spark configuration on a two-node cluster where the nodes are connected via two links. (NIC Fusion) The NCCL output shows three issues : Despite specifying only two Ethernet interfaces (enp1s0f0np0, enp1s0f1np1) via the environment variable…

Have you tried Connecting Two DGX Spark Systems via 200Gb/s RoCE Network for Multi-Node GPU Training | by Doran Gao | Oct, 2025 | Medium ?

NCCL/RoCEv2 Issues with Duplicates, GPU Direct RDMA, and Fusion

AndrewMyers November 14, 2025, 11:45pm 2

GPU Direct is addressed in FAQ ( DGX Spark / GB10 FAQ )