diff --git a/dags/test-train-pytorch.py b/dags/test-train-pytorch.py index 61b7043..788af8a 100644 --- a/dags/test-train-pytorch.py +++ b/dags/test-train-pytorch.py @@ -40,11 +40,11 @@ def run_training_node_func(rank, world_size): import time # CRITICAL: Set NCCL environment variables BEFORE importing torch - os.environ['NCCL_DEBUG'] = 'TRACE' - os.environ['NCCL_DEBUG_SUBSYS'] = 'NET,COLL' + os.environ['NCCL_DEBUG'] = 'INFO' os.environ['NCCL_TIMEOUT'] = str(NCCL_TIMEOUT) os.environ['TORCH_NCCL_BLOCKING_WAIT'] = '1' os.environ['NCCL_IB_DISABLE'] = '0' # Enable InfiniBand + os.environ['NCCL_IB_GPU_DIRECT_RDMA_MAX_DISTANCE'] = '6' # GPU at bus 0, IB at bus 6 os.environ['NCCL_P2P_DISABLE'] = '0' # Enable P2P os.environ['NCCL_IGNORE_CPU_AFFINITY'] = '1' # Better compatibility