ds_batch_config.json

{
    "train_batch_size": 2,
    "gradient_accumulation_steps": 1,
    "optimizer": {
      "type": "Adam",
      "params": {
        "lr": 0.00015,
        "max_grad_norm": 1.0
      }
    },

    "fp16": {
      "enabled": true,
      "loss_scale": 0
    }
  }