- Set NCCL IB env vars before TF import (IB_DISABLE=0, GPU Direct RDMA)
- Use NCCL_DEBUG_FILE approach for transport logging (same as PyTorch DAG)
- Add IB pre-flight diagnostics (dev/infiniband, sys/class/infiniband, ibv_devinfo)
- Add per-epoch logging with timing for all workers
- Add NCCL transport summary after training (NET/IB, GPU Direct, Channel info)
Co-Authored-By: Claude <noreply@anthropic.com>
