ds_config.json

{
    "zero_optimization": {
      "stage": 2,
      "offload_param": {
        "device": "cpu",
        "pin_memory": true
      },
      "offload_optimizer": {
        "device": "cpu",
        "pin_memory": true
      },
      "overlap_comm": true,
      "contiguous_gradients": true
    },
    "gradient_accumulation_steps": 1,
    "train_batch_size": 1,
    "fp16": {
      "enabled": true
    }
}