diff --git a/dags/test-train-pytorch.py b/dags/test-train-pytorch.py index 8afc74c..3f2f252 100644 --- a/dags/test-train-pytorch.py +++ b/dags/test-train-pytorch.py @@ -113,14 +113,18 @@ def run_training_node_func(rank, world_size): os.environ['WORLD_SIZE'] = str(world_size) os.environ['RANK'] = str(rank) - # NCCL Configuration - Start with basic ethernet config (IB is causing issues) + # NCCL Configuration for InfiniBand os.environ['NCCL_DEBUG'] = 'INFO' os.environ['NCCL_TIMEOUT'] = str(NCCL_TIMEOUT) os.environ['NCCL_BLOCKING_WAIT'] = '1' - os.environ['NCCL_IB_DISABLE'] = '1' # Disable IB for now - use ethernet - os.environ['NCCL_SOCKET_IFNAME'] = 'eth0' # Use regular ethernet + os.environ['NCCL_IB_DISABLE'] = '0' # Enable IB + os.environ['NCCL_IB_HCA'] = '^mlx5_[0-9]+$' # Regex for Mellanox HCAs + os.environ['NCCL_SOCKET_IFNAME'] = '^ibp[0-9]+s[0-9]+$' # Regex for IB interfaces os.environ['NCCL_P2P_DISABLE'] = '0' # Enable P2P os.environ['NCCL_IGNORE_CPU_AFFINITY'] = '1' # Better compatibility + os.environ['NCCL_NET_GDR_LEVEL'] = '5' # Enable GPUDirect RDMA + os.environ['NCCL_IB_GID_INDEX'] = '3' # Use RoCE v2 (common for modern IB) + os.environ['NCCL_IB_TC'] = '160' # Traffic class for low latency print(f"[{rank}] Environment configured:") print(f" MASTER_ADDR: {MASTER_ADDR}")