diff --git a/dags/test-train-pytorch.py b/dags/test-train-pytorch.py index e388d1c..7429a20 100644 --- a/dags/test-train-pytorch.py +++ b/dags/test-train-pytorch.py @@ -44,7 +44,7 @@ def run_training_node_func(rank, world_size): os.dup2(sys.stdout.fileno(), sys.stderr.fileno()) # CRITICAL: Set NCCL environment variables BEFORE importing torch - os.environ['NCCL_DEBUG'] = 'INFO' + os.environ['NCCL_DEBUG'] = 'TRACE' os.environ['NCCL_TIMEOUT'] = str(NCCL_TIMEOUT) os.environ['TORCH_NCCL_BLOCKING_WAIT'] = '1' os.environ['NCCL_IB_DISABLE'] = '0' # Enable InfiniBand