config/ds.json

{
  "train_micro_batch_size_per_gpu": "auto",
  "zero_optimization": {
    "stage": 1
  }
}