UCCL on our cluster

06 Nov, 2025

Yang Zhou’s UCCL transport finally runs cleanly on gnerv2 ↔ gnerv7.

Recent checkpoints:

• ✅ nvidia_peermem verified on both nodes (200 Gb/s ConnectX-6 over InfiniBand).
• ✅ Smoke tests: all-reduce/broadcast passed with UCCL plugin active.
• 📊 Benchmarks (5× iterations, 1 MB & 16 MB payloads):
  • UCCL all-reduce hit 63.8 Gb/s at 1 MB (≈6% over stock NCCL).
  • NCCL still edges out UCCL on 16 MB packets (~125 Gb/s vs. 122 Gb/s).